experiment-log.md

Experiments

What was tried on the micro tasks, and what came of it. Run names (abil6-*, lamfocus-2, ...) are runs in the author's runs/ directory, shown by the dashboard; they are not part of the repository. Win rates are against the scripted opponent; scripted baselines use scripts/baselines.py, fitted scripts python -m warcraftsim.puffer.bc eval.

Summary

• On footmen2, PPO reaches 95% wins in about 1.5 minutes with tuned settings.
• On random mirror matches, PPO from scratch ends near "let the units fight on their own". Team tactics (focus fire plus pulling hurt units back) don't emerge from random exploration: each half alone doesn't pay.
• Starting from a behavior-cloned script fixes that: PPO keeps and sharpens the script's tactics.
• Longer credit (horizon 64, λ 0.95 instead of horizon 16) lets PPO discover pull-backs from a focus-fire clone, which horizon 16 never did.

footmen2

• Scripted baselines win 0% (random), 30% (noop), 65–70% (focus fire), 90% (focus fire, and pulling a footman back while it is low and being hit).
• Tuned by sweeps (f2-sweep1..5, f2-step* in the dashboard): 0.5 s steps, lr 0.01, minibatch 192, replay ratio 4, the learning rate annealed over 400k steps → 95% wins after ~0.2M steps (~1.5 min), 98-100% soon after; 1.0 s steps: 100% for both seeds.
• Most of the speed came from more updates per sample (1 → 32 per epoch), then from longer steps and a shorter annealing schedule. Horizon 32 learns fastest at first but ends lower.
• Early policies learned focus fire plus pulling a hurt footman back. The 100% policy instead holds position until the enemies arrive: the scripted opponent then splits its damage over both footmen, while ours focus one enemy.

Mirror matches: how many hit points

Hit points decide whether micro matters (scripted baselines, 120-150 episodes each; the time limit scales with hit points):

hit points	episode	noop	focus + pull back (pull35)
25%	21 s	55%	44%
35%	28 s	53%	64%
40%	31 s	54%	80%
50%	36-40 s	53%	79%

At 25% units die within a few hits and every order costs more than it gains (attacking the weakest enemy in range 48-50%, pulling back without focus 19%); RL at 25% converged to noop (lr 0.003: 52%). mirror_mix_hp400 is the training setting.

Hero abilities (mirror_mix_abil_hp400)

Results on mirror_mix_abil_hp400 (0.5 s steps; runs abil40-*):

policy	win rate
noop (no casting)	31%
pull35 (no casting)	29%
castnoop	50%
castpull35	50%
PPO from scratch, lr 0.001, 2.5M steps	42%
PPO from the fitted castpull35, lr 0.001, 2.5M steps	52%

Not casting against a caster loses badly. PPO from scratch found a shortcut: 60-66% of its actions are casts and 94% of those aren't possible. An impossible cast does nothing, so this means "do nothing, and cast each ability the moment it's ready".

Sweeps for learning speed

Sweeps for learning speed from scratch (with action masks; 600k steps, ~5 min each; runs abil6-*, abil6b-*). At this budget the win rate is still rising, so faster runs show as higher numbers:

round 1 (base: lr 0.001, minibatch 192, replay ratio 4, horizon 64, γ 0.99, λ 0.9, entropy 0.001)	win @0.3M	win @0.6M
lr 0.003	21%	29% (peak 34%)
replay ratio 8	9%	27%
base	4%	25%
lr 0.002	7%	23%
minibatch 384	8%	22%
hidden 256	4%	21%
λ 0.95 / entropy 0.0003	1%	15%
γ 0.97	0%	0% (never learned)

round 2 (base: lr 0.003)	win @0.3M	@0.45M	@0.6M
horizon 32	27%	35%	33% (peak 40%)
lr 0.005	26%	28%	33%
λ 0.8	24%	31%	33% (peak 38%)
γ 0.995	24%	31%	33%
clip 0.3	20%	33%	34%
base (lr 0.003; round 1: 21% / 29%)	19%	28%	33%
replay ratio 8	10%	26%	30%
entropy 0.003	7%	16%	24%

round 3 (base: lr 0.003, horizon 32)	30% reached at	win @0.3M	@0.45M	@0.6M
horizon 16	0.19M (1.8 min)	33%	39%	39%
clip 0.3	0.25M	34%	33%	40%
λ 0.8	0.27M	31%	37%	36% (peak 42%)
lr 0.005 + λ 0.8 + γ 0.995 + clip 0.3	0.30M	29%	38%	38%
lr 0.005	0.31M	29%	27%	35%
γ 0.995	0.33M	27%	34%	32%
base (horizon 32; round 2: 27% / 35% / 33%)	0.40M	21%	31%	36%
minibatch 96	0.45M	27%	30%	35%

round 4 (base: lr 0.003, horizon 16)	30% reached at	35% reached at	win @0.3M	@0.6M
λ 0.8 + clip 0.3	0.10M (1.0 min)	0.15M (1.5 min)	34%	45% (peak 47%)
horizon 8	0.13M	0.15M	37%	40%
horizon 8 + λ 0.8	0.12M	0.15M	33%	41%
clip 0.3	0.13M	0.24M	42%	41%
λ 0.8	0.15M	0.25M	35%	40%
base (horizon 16; round 3: 0.19M)	0.23M	0.29M	35%	38%
lr 0.005	0.24M	0.37M	32%	39%

Shorter horizons (64 → 32 → 16) give more, smaller updates per sample and learned fastest; horizon 8 costs throughput (the trainer updates twice as often). The winner, horizon 16 with λ 0.8 and clip 0.3, is the task's default now (Task.train_defaults, used where the command line doesn't set an option). It reaches 30% wins after 0.1M steps; the settings this started from needed 0.66M.

Seed noise is about ±4 points at 0.6M, where the learning rate has annealed to zero and most configs end alike; the earlier columns separate them better. The minibatch must be a multiple of the horizon (PufferLib asserts it).

The plateau

The plateau: from scratch the tuned settings reach 40-45% after 0.6M steps, and 42-44% after 2.5M (runs abilbest-*). What was tried at 600k steps (runs plat-*; base 40%, other seeds 45% and 38-45%):

change	win @0.6M
hidden 256 / 3 layers	39% / 39%
entropy 0.003 annealed to 0	43%
γ 0.995	40%
replay ratio 2 / 8	39% / 40%
relational features (_rel: nearest opponent, in range, threatened, weakest, time to die)	37-39%
tactical masks (_tac: retreat only while losing hit points, no plain moves)	38-40% (faster early; retreat still unused)
0.25 s steps (horizon 32)	33-35%
start from the fitted castpull35 (masked BC), lr 0.001	52% (50% after 0.27M steps)
start from the fitted castpull35, lr 0.003	48%
value-loss weight 0.5 / 4 (default 2)	42% / 38%
value clipping off / gradient norm 0.5 / lr floor 20%	40% / 40% / 41%
V-trace / momentum 0.9 / momentum 0.98	38% / 37% / 36%
one agent per unit, one shared policy (_units, 3M agent steps = 0.6M game steps)	43% / 45% (30% after 54k game steps, 0.6 min)

Scripts that cast more carefully don't do better either (smartcast: area spells only with two enemies inside, targeted spells on the biggest threat, heals below half; 200 episodes each): castnoop 53%, smartcastnoop 48%, smartcastpull35 55%. Against a casting opponent, reasonable strategies all end near a coin flip: the headroom here may be small.

The trained policy plays as well greedily as sampled (43% / 42%), so that is its level. With the scripted casting rule in place of its own casts it wins 45%; castnoop (no attack orders at all, the same casting) wins 50%. So the gap to the scripts is partly casting, and partly attack orders that do worse than letting units auto-acquire. Only a better starting point moved the plateau.

Longer from the fitted script (bclong, 2.5M steps, lr 0.001): 47% at 0.3M, 50% at 2.5M, flat. It keeps the script's style: 91% attack orders, 6% casts.

Self-play

Self-play (mirror_mix_abil_self_hp400, MirrorSelfPlayEnv: the policy plays both sides of the same game with the full action set; the observations and actions are the single-agent task's, so a checkpoint also plays the scripted opponent with bc eval). Starting from bclong, 3M agent steps (run selfplay1), against the scripted opponent over training: 50% → 51% (0.6M) → 44% (1.2M) → 42% (1.8M) → 41% (2.4M) → 49% (3.0M, lr at 0). Plain self-play against the latest self drifts away from what beats the script and doesn't improve it. PufferLib's self-play pool (older checkpoints as opponents in part of the games: --selfplay.enabled=1 --vec.num_policies=2 --vec.hist_policy_percent=0.5) works with this env; the second agent carries policy tag 1. With it (run selfpool1: half the games against a past checkpoint, resampled every 100k steps, from bclong, 3M agent steps), the results against the scripted opponent stay at 45-52% (47%, 48%, 52%, 48%, 45%, 49% over training): no drift, but no gain either.

Without abilities: pulling hurt units back (mirror_mix_sem*_hp400)

Without abilities the plateau is an exploration problem. pull35 (68%) is focus fire plus pulling hurt units back, and neither half works alone: focus fire 53% (noop 51-54%), pulling back without focus 19% (at 25% hit points). PPO from scratch with every improvement above ends at noop's level (runs sem6-*: 47-49%; with tactical masks semtac6-*: 42-48%, 97% noop and no retreats): trying either half alone is punished, so it never finds the pair.

From a fitted script (behavior cloning)

Results on mirror_mix_sem_hp400 (noop 54%, the pull35 script 67-68% at 0.5 s steps; runs mix40-*):

start	lr, entropy	win rate
random (PPO from scratch)	0.003, 0.001	44% after 3M steps; never learned to retreat
random (PPO from scratch)	0.001, 0.001	47-48% after 2.5M (92% noop, no retreats)
pull35 fitted exactly (62% sampled, 64% greedy)	0.001, 0.001	~70% throughout (stopped at 0.8M)
pull35 label-smoothed (33% sampled)	0.001, 0.001	67-69% after 1M, 70% (best 73%) after 2M
pull35 label-smoothed	0.001, 0.003	62% after 2M (more randomness, no new behavior)
noop label-smoothed (~18% sampled)	0.003, 0.001	42-47% at 1-1.5M (stopped): no better than from scratch, still 25% harmful attacks
noop label-smoothed	0.001, 0.001	49-50% after 2.5M (90% noop, no retreats)

Without a good script to start from, PPO converges to letting the units fight on their own (≈ noop); it never discovers pulling hurt units back. From the script, it converges onto the script's behavior (99% of attacks on "weakest", 3-4% retreats) rather than beyond it. Random actions are very costly here: 7.5% of them turn noop's 54% into ~18%, since a random retreat or retarget takes a unit out of the fight for a second or two. Exploration is therefore punished hard.

Tactical mode

Tactical mode (_tac) tries to make the pull-back discoverable:

• no plain moves;
• a retreat only for a unit below half its hit points that is losing them;
• a chosen retreat goes on, re-ordered every step, while the unit keeps losing hit points (at most 3 s), so one decision is the whole pull-back.

With the same mechanics pull35 still wins 68% (a first version with a fixed 1.5 s commitment kept units out too long: 51%). PPO still doesn't learn it. From a fitted focus-fire script (focus, 49%, no retreats; runs focusrl*) it holds ~50% and drops its retreats from 1-2% to 0%, even with the commitment. A reward for kills and losses (_kill, ±0.2 per unit) doesn't change that (48% from the focus clone; from scratch 49% with tactical mode, 48% without; runs kill-*, killnotac). From the fitted pull35 (pullrl2), PPO keeps its retreats (3% → 2% of unit decisions) and goes from 58% to 65-66% in 3M steps. PPO can value retreats when they come with the rest of the strategy: forbidden at evaluation (bc eval --forbid retreat), the final pullrl2 policy drops from 66% to 46% (240 episodes each). Pulling back pays because the attack-moving opponent keeps switching targets and chasing, and that happens only when the whole team pulls hurt units consistently. A lone retreat while the rest fight just loses that unit's damage. Isolated exploratory retreats therefore look useless, and the strategy has to come from somewhere else (a script, replays).

Pulling back only now and then

Yet pulling back only now and then already pays. With focus fire, a script that pulls a unit back only with probability P each step it could (pull<L>p<P>, 200 episodes each) wins:

focus	pull35	pull35p30	pull35p10	pull50	pull50p10
56%	70%	70%	65%	47%	66%

So near the focus policy, occasional retreats have a clear gradient (+10 points at a 10% rate), and yet PPO from the focus clone removes them.

The horizon

The horizon was the problem. All the runs above used horizon 16 (Task.train_defaults of the ability tasks, tuned there for speed over the first 0.6M steps). GAE then sums at most 16 steps (8 s) and trusts the value estimate after that, while a pull-back pays off later: the unit survives to fight on, and the enemy that chased it gets focused. From the focus clone (lr 0.001; runs lamfocus*):

horizon, λ	win @0.6M	@1.5M
16, 0.8 (focusrl4-*)	48-52%	48-50%
16, 0.95	49%	50%
64, 0.95 (2 seeds)	54% / 52%	60% / 59%
64, 0.99	54%	59%
128, 0.95 (minibatch 384)	52%	53%

With retreats forbidden at evaluation, the 60% policy drops to 48%. It learned to pull back, which no run from the focus clone had done. From scratch, horizon 64 doesn't help (41-44% after 1.5M; runs lamscratch-*). Near noop, pulling back alone doesn't pay, so there is nothing to follow yet: PPO first needs focus fire, which it doesn't find on its own either.

Rejoin

Pulling back without focus fire failed for a mechanical reason: a unit that pulled back stood where the move left it, out of reach, until an attack order came. Without focus fire, 1-2 steps later, it never came (nooppull35: 4% at 40% hit points, 1% at 70%, many draws). _rejoin (implies _tac) makes the pull-back one whole maneuver: when it ends and the unit is told nothing (noop), it attack-moves back into the fight. With it (200 episodes each):

noop	nooppull35	nooppull35p10	focus	pull35	pull35p10
52%	40%	57%	48%	70%	65%

Now occasional pull-backs from noop pay a little (+5), but PPO from scratch still doesn't find them (horizon 64, λ 0.95: 47% for both seeds, retreats 1%; runs rejoin-*; 41-44% without rejoin).

The fitted pull35 on the rejoin task (bc/mirror_mix_sem_rejoin_hp400-pull35; the script wins 67% over the 2000 recorded episodes) plays better than the script: 68% sampled and 79% greedy after 60 epochs (56% / 75% after 30, when its retreat recall was still rising: 0.80, then 0.86).

General orders and the entity network (mirror_mix_gen*, --trainer torch)

General orders have no built-in tactics (see the README). The same scripts, written with them (120 episodes each): noop 52%, focus 61%, pull35 71%, pull35p10 58%; with abilities, castnoop and castpull35 52%. The action space can express what the built-in retreat and target rules did.

• From scratch, PPO with the entity network collapses to noop within ~30k steps (entropy 7.6 → 0.01; run torchsmoke): random stops, holds and moves are costly, so the first thing it learns is to stop giving orders, and it stays there (~50%).

• Behavior cloning: the entity network fits pull35's general orders almost exactly (40 epochs, 3 min: order accuracy 99.4%, recall of the pull-back moves 0.98, of attacks 1.00; bc/mirror_mix_gen_hp400-pull35). PufferLib's network, fitted to the same script with the built-in retreat, recalled 0.80-0.86 of the retreats. Played with sampled actions and no updates, the clone wins 65.6% (2242 episodes; run genbc-1); the script 68-71%.

• PPO from the clone (lr 3e-4, horizon 64, λ 0.95; runs genbc-*):

  	0.2M	0.5M	1M	2M
  plain PPO (genbc-2)	57% → 21%	stopped at 0.4M (18-47%)
  with a KL of 0.1 to the clone (genbc-3)	69%	71%	74%	77%

  Plain PPO moved the policy too far per update (KL 0.03-0.06 per epoch, a fifth of the samples clipped): it walked away from what the clone knew before its value estimates were any good, and collapsed. A KL penalty toward the clone, as AlphaStar keeps its policy near the supervised one, kept it stable, and it went on past the script it was cloned from (77% against 71%).

• Gentle fine-tuning works better than the anchor (runs genft-*: lr 1e-4, 10 epochs that train only the value first, each epoch's updates stopped at a KL of 0.02):

  	0.5M	1M	1.5M	2M
  gentle (genft-1)	75%	81%	87%	91%
  gentle, with the KL to the clone (genft-2)	71%	73%	74%	75%
  lr 3e-4 with the KL to the clone (genbc-3)	70%	72%	75%	76%

  The collapse came from the first updates, made before the value head could judge anything. With the value trained first and small steps, the anchor only holds the policy back.

  What the 91% policy does differently (scripts/target_choice.py, 40 episodes): it keeps the script's pull-backs (7% of orders are moves) and its focus fire (99% of the attacks in a step go to one enemy), but not its target: 72% of its attacks go to the enemy with the fewest hit points, where the script's go there always. When it picks another, that is mostly the enemy with the lowest share of its hit points left (71% of those), often the hero (55%). It finishes the most damaged unit, often the hero, which carries more hit points and does more damage. Its side loses 1.25 units per game instead of 2.3. Only a pointer at units could learn this: the rule-based targets of the older tasks had no rule for it.

Hero abilities with general orders (mirror_mix_gen_abil_hp400)

The same recipe on the ability task, against the casting scripted opponent: castpull35 with general orders (52% over 2000 episodes) fitted in 40 epochs (order accuracy 98.8%, recall of casts 0.96, moves 0.96, attacks 0.99), then fine-tuned gently (runs genabil-*):

	win rate
the clone, no updates	51%
fine-tuned, 0.5M / 1M / 1.5M / 2M steps	61% / 68% / 69% / 68%
PufferLib's network: from scratch / from the fitted script (earlier)	42-45% / 52%

The plateau of the older tasks ("reasonable strategies all end near a coin flip") was the network's, not the task's.

League self-play (mirror_mix_gen_self_hp400)

The first league run (genleague: from the fitted pull35, gentle fine-tuning, 3M steps; half the games against itself, a quarter against past snapshots by PFSP, a quarter against the scripts noop, focus and pull35) looked good on its own terms: it beat its past snapshots 63-70% and the focus and pull35 scripts 70-80% (40% at the start). Head-to-head it beat the 91% policy genft-1 68% (scripts/match.py: 46 wins, 11 losses, 43 draws).

Against the game's scripted opponent, which attack-moves at the nearest enemy, it won 4%.

It had learned to run: 38-61% of its orders were moves away from the enemy. In a mirror fight the side that walks into range is hit first, so against copies of itself waiting and backing off pays, and the games ran into the time limit (draws went from 18% of its self-play games to 46%). None of its opponents chased relentlessly, so nothing punished running; the scripted opponent does exactly that, and ran it down. Self-play found an equilibrium of its own games, not strength.

Two changes for the next run: a draw counts as a loss for both sides (_nodraw), and the league's scripts include amove, which attack-moves at the nearest enemy each step like the game's scripted opponent: an anchor that punishes running away.

The second run (genleague2, with both) beat its amove 92-100%, and the game's scripted opponent still only 18%. The script was broken: it re-ordered every unit every step, and a new order cancels an attack in progress, so its units hardly landed a hit. The game's scripted opponent orders only idle units (at the nearest enemy's position) and leaves fighting units alone. With that behavior (amove now), the script matches it: 84% and 22% for two policies that score 91% and 18% against the game's opponent.

Meanwhile the three policies form a cycle, as non-transitive games do (scripts/match.py, 120 episodes each):

	against
genft-1 (trained against the scripted opponent)	wins 91% against the scripted opponent
genleague2	wins 70% against genft-1 (65-16, 39 draws), 57% against genleague
the scripted opponent	wins 82% against genleague2

The fine-tuned policy waits for the attack-moving opponent and punishes it; the league policies out-wait the fine-tuned one, which walks into them; the scripted opponent keeps coming and runs the league policies down. A league needs every one of these styles among its opponents; this is what AlphaStar's league exploiters are for.

The third run (genleague3: no draws, the corrected amove, scripts chosen by PFSP too, so the ones it still loses to come up more) broke the cycle:

	genleague3	genleague2	genft-1
against the game's scripted opponent	82%	18%	91%
head-to-head against genft-1	79% (87-18, 15 draws)	70%
head-to-head against genleague2	60% (44-21, 55 draws)

It gives up nine points against the scripted opponent to the policy trained only against it, and beats that policy four games in five. The anchors decided it: with an opponent in the league that chases like the scripted one, running away stopped paying.

League self-play with hero abilities (mirror_mix_gen_abil_nodraw_self_hp400)

genleague5 continued genleague3 on the task with abilities: the unit features grow by the ability features (their input weights start at zero, so it starts as genleague3 exactly), and the cast order's logit gets +6 (--cast-bias): genleague3 never had a cast possible and would have tried one 0.5% of the times it could. The league's scripted anchors cast like the game's scripted opponent, and a main exploiter plays a quarter of the games against the main learner. 3M steps in four pieces (genleague5, 5b, 5c, 5d: resumed after trainer speedups, the league and the learning-rate schedule carried over).

	against the casting scripted opponent	head-to-head
genleague5d (3.0M)	56%	beats genabil-2 56% (64-49, 7 draws), genleague3 71% (81-31, 8 draws)
genabil-2 (fine-tuned against the scripted opponent only)	66%
genleague3 (its start: never cast)	36%

The same pattern as without abilities: the league policy gives up ten points against the one opponent the specialist trained against, and beats the specialist head-to-head. Casts stayed at about 4% of its orders throughout. The exploiter's win rate against the main learner sank from 47% to 10-20% in the first million steps and never reached the 70% that restarted it; now it also restarts below 20%, from the main learner's current policy.

Whole games: cloning the built-in AI on duelrush

Demonstrations: built-in AI (normal) against built-in AI on duelrush (50% handicap, 0.5 s steps, all races; fullgame/collect.py). Games last 1-4 minutes of game time (a tie at 4). The fits are fullgame-167, -253, -263 (the number of games); the policy plays the normal AI with fullgame/play.py.

What mattered, in the order it was found:

• Not every recorded order is a decision. The AI's order events include the engine's own: internal orders it gives to most units (851974 alone is 29% of the events), resumeharvesting (7.7%: a worker going back after a drop-off), returnresources and autocasts. A policy that learned resumeharvesting interrupted working harvesters, and the game refused two thirds of its orders. These are now dropped from the labels.
• Re-issued harvest orders. The AI re-orders harvesting workers to harvest every few seconds, with no effect in its games. Learned as decisions, the policy kept sending its workers to other trees and reset their work: 240 gold from 27 workers in 80 s. Harvest orders to workers that already harvest are dropped from the labels and skipped at play time: 4860 gold in two minutes with the same policy.
• Production was invisible. A building's current order is 0 while it trains, so the policy could not see a full queue and kept ordering more. Each building now has queued (its accepted train and research orders not yet done) and busy (from the production events) features.
• Points: x and y bins sampled independently paired the x of one place with the y of another; y is now chosen given x.
• Overfitting: at 250 games the validation loss rose after 9 of 20 epochs. Dropout 0.1 and keeping the best epoch fix that. A global temperature below 1 made the policy issue fewer orders; --order-temperature sharpens only which order a unit gets.
• fullgame-263 (263 games, before the production features): 2 wins in 16 against the normal AI. Its economy works; it under-spends, often builds no barracks, and now and then gives rare orders (battle stations, board).

Cloning under the rebalanced rules (fullgame-rush2)

3043 games under the v12 rules (walking workers carry more, 7× mines; all races, normal and insane AI; demos-rush2-1), 8 epochs: validation loss 3.05 (still falling at the end), the AI's order given a unit gets one: 69% (target unit 86%, point within 5.4 bins), and the rate at which units get orders matches the AI's (9.8% vs 9.9% of unit-steps). Against the normal AI in mirror matchups: 2 wins in 32 (both undead), most games lost within 1–2 minutes. About a third of its ~150 orders a game are refused: workers it cannot afford or house yet, buildings where they don't fit.

Fewer wasted orders: an availability mask and snapping

A third of the clone's orders were refused, most of them for being unaffordable. The AI only orders what it can pay for, so the clone never saw an unaffordable order and never learned what "can't afford" looks like.

The fixes:

• An availability mask (fullgame/costs.py, --avail-mask): each order class's gold, lumber and food cost comes from the game's tables, scaled by the map's cost rule. The classes the player can't pay for at the step are masked.
• Harvest snapping: a harvest order aimed at anything other than a gold mine goes to the nearest mine in view.
• Build snapping: in the harness, a build that doesn't fit tries rings of spots around its point.

The share of refused orders fell as follows:

orders	refused before	refused after
all	38%	25%
builds	53%	19%
train and research	58%	38%
harvest	77%	63%

The clone plays no better for it: in 32 mirror games against the normal AI it won 1, tied 3 and lost 28.

Self-play from the clone (fgself-2)

The setup was PPO from fullgame-rush2, mirror matchups, a league (itself, past snapshots, the easy and normal AI), material shaping and the tie-break.

After 2.1M agent steps it beat its past snapshots about 60% of the time, most of the rest ties. Against the built-in AI it had one tie in its last 75 games (easy and normal) and no wins. Self-play improves the policy against its own kind, but that doesn't carry over to the AI. The clone is the weak link: it starts too far behind for the anchor games to give a learning signal. The next steps are a better clone (more games, memory, a value head trained on the demonstrations), then self-play again.

The clone hoards: stale queues in the demonstrations

The clone's economy is not the problem. It gathers as much as the AI (6791 gold per game vs 6028), but it doesn't spend it.

A frame at 62 s of a night elf mirror game:

side	gold held	lumber held	food	material (units and buildings)
the clone	2552	2272	17/30	3114
the AI	1512	982	53/60	7452

The cause is in the demonstrations. The AI retries its train orders until it can afford them, and the recorded order events include every attempt:

• 61% of the AI's train and research orders were unaffordable when issued, and 80% of those started nothing.
• Each attempt raised the building's queued feature. Nothing ever finished to bring it down.

The effect on the recorded state, over the AI's finished production buildings:

• They showed "5+ queued" while not busy for 36% of their steps.
• The clone learned to train mostly in that state: 16% per step, against 2.3% at an idle building.
• In its own games only accepted orders count, so its buildings are almost never in that state (1% of building steps). It sat mostly at idle buildings, where the learned rate is low.

The fix (Encoder.encode with costs):

• A cost-bearing order counts, both as a label and in the queue, only if the player could pay for it at the step. Costs are deducted in order within the step.
• A train or research order counts only if production started at that building within 2 steps, or the building was already busy (the order queues behind).
• BC trains with the same availability mask the policy plays with.

52% of the AI's train labels go. The share of building steps with 5+ queued falls from 23% to 2%, and with a queue but not busy from 22% to 3%.

The clone cancels its own builds, so it is supply-blocked

The label fix alone didn't stop the hoarding. The ablation below compares two fits on the same 1000 games (3 epochs each), each playing 16 mirror games against the normal AI:

labels	wins, ties, losses	gold + lumber held (clone vs AI)	food at 1 minute (clone vs AI)
old (refused retries kept)	0, 1, 15	not recorded	not recorded
new (Encoder with costs)	0, 0, 16	4074 vs 1746	19 vs 46

Composition at 60 s showed why: the clone is supply-blocked.

• Undead: 10/10 food with no ziggurat, against the AI's 50/50 with 3–4.
• Human: 1 farm, against the AI's 7.
• Night elf: 1–2 moon wells, against the AI's 5 (+3 building).

With no free food the availability mask blocks all training, and the gold piles up.

Over 8 games, the clone had 80 accepted build orders and 50 constructions started (62%). The AI had 128 and 122 (95%).

Following each build showed the cause. Of 102 accepted builds, 48 never started, and in every one of those the worker got another order before reaching the site. All 45 builds that started were left alone. The policy picks every unit's order every half second, so a worker on its way to build keeps getting a chance of another order, which cancels the build. The AI leaves its builders alone until the building stands.

The fix: BCAgent keeps a worker whose current order is a building (on its way, or constructing) from getting other orders. This is features.building, with a 30 s timeout, and it applies in play and self-play.

In the same 8 games (seed 7), the fix changed:

	builds started	held (clone vs AI)	food at 1 minute (clone vs AI)
before	44%	4760 vs 1942	15.6 vs 47.9
after	68%	3847 vs 1760	18.6 vs 37.0

The builds that still fail lost their order in the game, probably at placement, not to another order from the clone.

Evaluations now record each side's gold and lumber on hand (mean over the game) and its food at one minute. The dashboard charts both.

The opening, and demonstrations from the clone's own states

Timelines of 8 mirror games (fullgame-rush2 with builder commitment) against the normal AI:

time	clone food	AI food	clone gold on hand	AI gold on hand	clone army	AI army
20 s	15	23	2154	1248	0.4	3.2
30 s	16	29	3280	1349	0.5	5.0
60 s	16	52	6043	1681	1.4	11.8

Workers are on par, so the economy works. The clone is behind within 20 s, before any gold piles up.

Checks on the cause:

• Calibration on the AI's states: fine. In the first 10 s the model expects 3.6 build orders per side against the AI's 3.1, and 8.6 train orders against 8.4.
• Clipping gold and lumber at play time: a little better (army at 60 s: 2.9 instead of 1.4).
• Several train orders to one building in one step: 93% of the kept train orders are the only one to their building in that step, so one label per step loses little.
• Timing: at the steps where the AI builds something, the model expects only 0.28 workers to build it (median 0.07). The rest of its build probability sits on steps where the AI builds nothing. The totals match, but the AI's script runs on timers and conditions the model can't see.

One human game step by step:

step	the AI	the clone
1	trains a peasant, builds an altar	trains a peasant, builds an altar
2–4	builds its first farm at 7/12 food	builds a barracks
11–12	builds a barracks	builds a second barracks
20	trains a Paladin, builds another farm	builds its first farm, already at 12/12
21–28	trains footmen	stuck at 12/12 food

After an early mistake the clone is in a state the AI's games never show, and cloning has nothing to learn from there. The answer is DAgger with the built-in AI as the expert:

• Takeover games (collect.py --policy): the clone plays one side for 5–90 s, then the built-in AI takes it over (protocol.StartAI) and its orders are recorded.
• BC skips the taken-over side's steps before the takeover.

demos-takeover-1: 4000 such games with fullgame-rush2 as the clone. They also measure how fast the clone loses a game. Evenly matched AIs would win about 50%.

the clone plays until	its side wins, the AI playing on	games
5–20 s	35%	725
20–40 s	23%	925
40–60 s	9%	999
60–90 s	6%	1351

Even the first 20 seconds cost about 15 points, and after a minute the game is mostly lost.

More data and the new labels alone don't fix the opening. fullgame-rush3 is fitted on the 8000 demos-rush2-1 games with the new labels. After epoch 4 (validation loss 2.93, order accuracy 70%) it played 32 mirror games against the normal AI:

	fullgame-rush2	fullgame-rush3, epoch 4	the AI
wins, ties, losses	1, 2, 29	1, 1, 30
gold + lumber held (mean)	4563	4178	~1900
food at 1 minute	20.6	21.8	46–49
refused orders	~25%	16%

Refused orders fell, but the spending gap stayed. The model is calibrated on the AI's states and still fails on its own, which is why the takeover games are the next step.

Where the clone's army goes missing (fullgame-rush3-tk)

fullgame-rush3-tk is fullgame-rush3 (epoch 5) fine-tuned on its 8000 games plus the 4000 takeover games. After epoch 2 it played 32 mirror games against the normal AI:

	result	gold + lumber held	food at 1 minute	refused orders
fullgame-rush3-tk, epoch 2	1 win, 31 losses	3951	24.1	13%
the AI		1965	43.3

The first minute, per game, over 8 of those mirror games:

	clone	AI
train orders	24.8 accepted (+6.8 refused)	251 (mostly retries)
training started	22.9	27.5
units trained: workers	~14	~11
units trained: army	~4	~14.5
units trained: heroes	~0.4	~1.2
constructions started	9.3	11.4

The types are much the same: moon wells 2.9 vs 3.4, burrows 2.1 vs 2.6, altars and barracks about equal. The Ancient of War is 0.2 vs 0.5.

Two suspects are ruled out:

• Placement. The clone builds slightly closer to its hall than the AI (median 703 vs 816 units).
• Training decisions. At idle, finished production buildings that can afford a unit, the clone trains at the AI's rate or above: Ancient of War 0.22 vs 0.06 per step, orc barracks 0.25 vs 0.10, Tree of Life 0.22 vs 0.21, Great Hall 0.07 vs 0.07.

Every local decision is roughly calibrated. The army gap comes from compounding small deficits: army buildings come later, occasional supply blocks, and a unit mix that leans toward workers. What's left is timing, which cloning can't pin down: the AI's script runs on timers the model doesn't see. Self-play's material shaping rewards every unit and building made, which is the signal this needs.

Right after a takeover, even the AI orders modestly: 0.13 train orders and 0.11 build orders per step in the first 10 s. The model on those states: 0.19 and 0.09.

Self-play from the new clone (fgself-3)

fgself-3 is PPO self-play from fullgame-rush3-tk, with the value head trained in BC (a 2-update warm-up), builder commitment and the fixed labels. It runs at 850–980 agent steps/s. Against the built-in AI it still loses almost everything: over its last 100 such games at 1.49M steps, 1 win, 2 ties, 72 losses against normal and 25 losses against easy.

The same spending evaluation as for the clones shows what changed. Mirror games against the normal AI:

	BC v3 (the start)	fgself-3 at 1.48M steps	the AI
wins, ties, losses	1, 0, 31	0, 1, 15
gold gathered	6577	10230	~7500
food at 1 minute	24.1	31.9	~47
gold + lumber held	3951	5292	~2150

It grew its economy, most likely with workers: cheap, safe material that the shaping rewards. It still hoards. Army takes production buildings first, which RL hasn't found yet.

A curriculum against the built-in AI (fgself-4)

fgself-3 won 2 of its last 100 games against the AI after 3.3M steps (1 tie, 33 losses against normal; 11 ties, 53 losses against easy). From games it always loses, it learns little about beating the AI. Self-play already sits at 50% against itself, but that doesn't carry over.

Neither knob alone brought the clone near 50% against the easy AI:

setting	wins, ties, losses
the clone's units at twice the AI's hit points	1, 1, 6
the AI starting 30–48 s late	0, 3, 10, while the clone gathered 2–3× the gold
both: twice the hit points, the AI 55–60 s late	2, 5, 3

The curriculum (League, --curriculum) is a level per AI difficulty that a loss raises by 0.02 and a win lowers:

• From 0 to 0.5 the learner's hit points rise to twice the AI's.
• From 0.5 to 1 the AI also starts late, up to 120 s.
• At 0 it's the real game.

fgself-4 starts at 0.75 and sends half its launches against the AI. After its first 45 such games it had 9 wins, 5 ties and 13 losses against normal, and 6 wins, 8 ties and 4 losses against easy. The levels settled near 0.71 (easy) and 0.83 (normal). The progress to watch is the levels falling toward 0.

The mean level per 0.5M steps. A single level moves by 0.02 a game, a random walk around the 50% point (±0.2 over ~100 games), so only the averages show the trend:

steps	easy	normal
0–1M	0.92–0.95	0.95–0.97
1.0–1.5M	0.89	0.89
1.5–2.0M	0.74	0.72
2.0–3.5M	0.65–0.67	0.67–0.76
3.5–4.0M	0.77	0.54

In its first million steps the learner needed nearly the most help: twice the hit points and the AI 110–115 s late. From 2M steps it held 50% at about 0.7 (twice the hit points, the AI ~50 s late). It improved against the AI, which fgself-3 never did, and then held level.

The decline went on slowly, about 0.03 a million steps: 0.57 (easy) and 0.60 (normal) by 8–9M steps. The real game (even hit points, the AI on time) doesn't show it yet. The checkpoint at 8.4M steps in 32 mirror games against the normal AI:

	clone (BC v3)	fgself-4 at 8.4M	the AI
wins, ties, losses	1, 0, 31	0, 2, 30
gold gathered	6577	11353	~7400
gold + lumber held	3951	5969	~2170
food at 1 minute	24.1	27.8	~44

It learned to gather, not to spend. With twice the hit points a small army wins fights, so nothing in the curriculum pushed it to build a bigger one, and that is the skill it lacks. fgself-5 goes on from that checkpoint with the late start alone (--curriculum-hp 0, up to 180 s, starting at 90 s) and even hit points, so the fights stay as they are.

With even hit points the levels fell fast from 90 s. fgself-4's learner beat the AI starting 30–45 s late, where the clone had lost 10 of 13 such games. Over its first 2.8M steps the late start needed for 50% was 25–80 s (normal) and 30–75 s (easy), per half million steps. A 0.02 level step (3.6 s a game) makes it a random walk of about ±36 s over 100 games, so from there on the step is 0.01. 10% of the AI launches now play the real game (--real-share), and those results are charted as the yardstick.

The late start is exploitable too: a tax on the AI's income (fgself-6)

fgself-5 held 50% with the AI starting 30–50 s late, but the real game stayed at about 4% (5 wins, 12 ties, 168 losses). Its videos showed why: it attacked the idle AI. 65% of its curriculum wins came before the AI had started.

Two knobs have now been gamed:

• Twice the hit points let a small army win, so the learner never had to spend.
• A late start left an opponent that doesn't defend itself.

The third knob is a tax (--curriculum-mode tax). The built-in AI plays from the start, but each step play_one takes a share of what it gathered (and of its starting gold and lumber) with SetResources. That makes a poorer opponent that still defends, builds and attacks.

Smoke test against the easy AI at a 45–61% tax: 6 wins, 1 tie, 15 losses in 22 games. In several losses the learner had gathered 2–3 times the AI's gold (22,295 vs 7,845), so this curriculum presses on the skill it lacks: turning gold into an army.

fgself-6 continues from fgself-5 at 6.9M steps. It starts at level 0.6 (a 54% tax, up to 90%) with a controller step of 0.01, and 10% of its AI launches play the real game.

After 1.5M steps the learner's race decided most curriculum games. Mirror matchups, the learner's score:

race	vs the taxed easy AI	vs the taxed normal AI	real game
night elf	0.94	0.96	1 win, 2 ties, 3 losses (normal)
undead	0.69	0.74	0 wins in 32
human	0.14	0.07	0 wins in 18
orc	0.08	0.13	0 wins in 12

One level per difficulty settled where night elf wins balanced human and orc losses. Every race's games were then nearly decided and taught little: the easy AI's level sat at the 90% maximum while human and orc still lost. The curriculum now keeps a level per difficulty and learner's race.

Human and orc had no lumber: harvest switches were dropped as "redundant"

Per race, human and orc stayed pinned at the 90% maximum tax. Against an AI that kept a tenth of its income, the learner scored:

• As human: 1 win, 19 ties, 20 losses, gathering 29,183 gold.
• As orc: 3 wins, 24 ties, 13 losses, gathering 44,561 gold.

The videos show the cause: in one human game it held 8225 gold and 187 lumber at 72 s, and 18,731 gold, 445 lumber and 24/24 food at 143 s. Human and orc farms and barracks cost lumber, so it could only hoard gold.

The bug: the rule that drops the AI's re-issued harvest orders (they reset the workers' work) dropped every harvest order to a harvesting worker. A worker's current order doesn't say gold or lumber. So a miner sent to the trees was dropped too, from the labels and at play time, and the clone learned that harvesting workers never switch. In the AI's games, 50% of human harvesting workers are on lumber, 36% of orc and 35% of night elf.

The fix:

• features.harvest_resource says what an order harvests.
• The encoder and BCAgent track what each worker was last sent to.
• redundant() only drops orders to the resource a worker already harvests.
• A new entity feature marks workers on lumber (F 28 → 29; older checkpoints get a zero column).

In 60 games this recovers 148 (human) and 238 (orc) switch labels. fullgame-rush4 fine-tunes fullgame-rush3-tk on the fixed labels: 8000 AI games plus both takeover collections, 2 epochs.

The first version still missed most switches. The starting workers mine by the map's melee setup and new ones by the town hall's rally point, so no recorded order sent them and their assignment was "unknown", which counted as the same resource. fullgame-rush4 gathered 514 lumber as human (the AI 1880) and 398 as orc (1408). Unknown now counts as gold. That recovers 358 (human) and 365 (orc) switch labels in the same 60 games, and fullgame-rush5 fine-tunes rush4 on them for 2 epochs.

fullgame-rush5 in 32 mirror games against the normal AI, the real game:

race	wins	lumber (clone vs AI)	food at 1 minute (clone vs AI)
night elf	4 of 9	2853 vs 3215	35.8 vs 44.4
undead	2 of 10	3862 vs 1387	34.9 vs 39.3
orc	0 of 5	1195 vs 2269	25.0 vs 48.2
human	0 of 8	555 vs 1290	11.8 vs 36.2

That's 6 wins in 32 (19%), where every clone before won 0 or 1. Orc lumber tripled. Human is still broken: little gold (3551) and 11.8 food at one minute. fgself-7 starts self-play from fullgame-rush5 with the per-race tax curriculum.

In a human game's video the problem is the army, not supply: 21/36 food and 2860 gold at 39 s, food barely rising, no footmen from the barracks. The human AI's early army kills it at about 70 s (human games last 0.9–1.3 minutes).

fgself-7 got worse against the AI from its first updates:

• After 1.4M steps all the taxes had risen to 65–90%, night elf included.
• The real game went 0 wins in 60.
• In its first 477 games the curriculum score was 0.14 at a mean 56% tax. The KL to the clone jumped to 0.096 by 250k steps.

The harness plays the same policy as play.py: with a learning rate of 0 it won 6 of 64 real games against normal. By race: night elf 3 of 8, undead 3 of 14, human 0 of 20, orc 0 of 22. That matches play.py race by race, so RL did the damage.

The likely causes: BC's value head learned the AI's returns, much higher than the clone's, so the first advantages were large and wrong; and half the games were mirror self-play. fgself-8 starts from rush5 with:

• 20 value-only updates first
• the KL to the clone at 0.2 (was 0.05)
• a learning rate of 3e-5 (was 5e-5)
• 75% of launches against the AI
• the curriculum starting at level 0.3 (a 27% tax)

fgself-8 improves in the real game, the first run that does. Its 10% real-game games by race, in fifths of its first 17,400 games (7.2M steps):

race	1st	2nd	3rd	4th	5th
undead	13%	26%	33%	53%	45%
night elf	–	10%	26%	28%	50%
orc	3%	8%	11%	0%	11%
human	2%	2%	4%	0%	4%

That's the share of wins; the rest are ties and losses. The last fifth, for example: undead 27W 1T 32L, night elf 24W 8T 16L. The clone won about 21% (undead) and 38% (night elf).

The taxes fell to match. By 7.2M steps undead needs no tax against normal: the curriculum game is the real game. Night elf needs 25% (normal) and 45% (easy), orc 43–49%, human 47–74%.

From 7M to 11M steps undead and night elf held at about 40–50% real-game wins, and human and orc at about 5%. The Production tab (per game, the last 800 games) shows why human loses in the real game:

	learner	the AI
footmen trained	40.6	19.2
peasants trained	33.8	12.0
heroes trained	0.3	1.3
lumber gathered	696	2513
kills	12.9	84.8
losses	87.6	21.1

It spends now, but on workers (cheap, safe material that the shaping rewards) and footmen that trade 1:7. The AI fields footmen, riflemen, knights, mortar teams, gryphon riders and heroes, with its upgrades. The learner builds an Altar of Kings in only 0.6 games out of 1. Orc is similar (grunts 30 vs 19, peons 21 vs 12, kills 31.8 vs losses 50.5, in curriculum games). A likely cause is piecemeal attacks: each unit picks its own attack order, so the reinforcements trickle into the enemy army.

An orc loss against the easy AI (at 12M steps) shows it:

• 53 s: the learner is ahead in material (5722 vs 5172) at 42/50 food, but holds 4093 unspent gold, and its units are spread around its base.
• 76 s: the AI attacks as one group, with a Far Seer's chain lightning, and wins the fight in the learner's base.
• 100 s: the learner is down to 8/10 food (2458 vs 6561 material), still with 4529 gold.

Two coordination problems: spending while fighting, and gathering an army when an attack comes. Both are hard for a policy that decides every unit's order separately. At 13M steps night elf reached 0% tax against normal, as undead had at 7M. Both then held about 35–45% real-game wins.

fgself-8 stopped at 16M steps. Its real games in total:

race	wins	ties	losses
night elf	173	64	205
undead	166	89	251
orc	25	27	310
human	19	9	395

Its last few million steps held about 44% (night elf) and 20–40% (undead), and undead's tax crept back up to 25%.

Next is a second DAgger round from its states. demos-takeover-3 holds 4000 takeover games with fgself-8's last checkpoint as the policy, handing over after 10–60 s: the AI's orders from the states the learner reaches (unspent gold, a scattered army, no altar). With the RL policy playing the first 10–60 s, its side then won about 31% under the AI, whenever the AI took over (10–25 s: 32%, 25–40 s: 31%, 40–60 s: 31%). The clone's side had dropped from 34% to 13% over the same window. So the RL policy no longer loses ground in the opening.

Fine-tuning the checkpoint on these games did not work. The fine-tune ran 1 epoch at learning rate 3e-5 with the value head left alone (fullgame-rl8-dagger). 64 mirror games against the normal AI, the real game, same seed:

	human	night elf	orc	undead	total
fine-tuned	0/16	4/21	0/17	2/10	6 wins (9%)
fgself-8, 16M steps	3/16	10/21 (1 tie)	4/17	2/10 (3 ties)	19 wins, 4 ties (30%)

Cloning the AI's orders overwrote what RL had learned. Self-play goes on from the checkpoint; the takeover games would have to come in as an auxiliary loss during RL, as AlphaStar's KL to its supervised policy, rather than as a fine-tune. The 30% is also the cleanest measure of fgself-8, up from the clone's 19% (fullgame-rush5).

DAgger as an auxiliary loss (fgself-9)

fgself-9 goes on from fgself-8 at 17.3M steps. It adds the cloning loss on the 4000 takeover games from its own states to each PPO minibatch, times 0.005 (--bc-data, --bc-coef). The cloning loss fell from 6.7 to about 3.8 in the first 70 updates. After a jump in the first few updates (0.02), the per-update KL settled at fgself-8's level (0.004–0.005). The real game after 1.2M steps:

race	fgself-9	fgself-8 (in its last millions of steps)
human	18W 31L (37%)	~5%
night elf	17W 5T 9L (55%)	~44%
orc	6W 1T 15L (27%)	~7%
undead	8W 9T 26L (19%)	~35%

Human and orc, which RL alone never moved, win a quarter to a third of their real games. The takeover games show the AI's decisions from the learner's own states, such as building an altar, training heroes and gathering lumber. The cost is throughput (357 steps/s): the loader workers encode the demonstrations next to the games.

Night elf had almost no real games at first (6 against 46–96 for the other races). The real-or-curriculum draw came right after the race choice, and the actors' seeded random streams correlated them. A replay of those streams gave night elf 12 of 91 real launches. The real game is now a launch kind of its own, and the replay gives 17–24 per race.

A second skew, found on a day of 15 restarts (2026-09-30): a game thread's random stream was seeded by the thread alone, so every restart began with the same launches. Of the real launches against the normal AI, 43 were human's and 6 orc's. The seed now includes the update the run starts from; over 200 simulated restarts every race gets its share (101 to 132 of 923 real launches per race and difficulty).

Games between agents had no heroes: the scripted reset (fgself-1 to fgself-9)

Found while profiling (next section): the games between agents restarted by script (GameSetup.melee_reset: remove every unit, respawn the start; 0.1 s) instead of reloading the map, and that was not a new game.

per game, fgself-9	against the built-in AI (map reloaded)	between agents, scripted reset	between agents, map reloaded
heroes trained, orc	1.30	0.07	1.33
heroes trained, undead	1.69	0.13	1.91
heroes trained, night elf	0.80	0.08	0.91
research started, orc	5.7	0.35	5.5
research started, undead	4.7	0.55	5.7
food used after a minute, median (lowest)	41 (3)	18 (-345)	39 (0)

• Heroes. The engine keeps counting a removed hero. After one game with a Blademaster the altar refused another and a Far Seer needed a Stronghold (the second hero's requirement); a game later the Far Seer was refused even with a Fortress. GetPlayerTechCount, the type and hero limits and the hero tokens all read as in a new game. Handing the hero to the neutral player before removing it changed nothing.
• Food and research. A third of those games had less than 5 food used after a minute, and research all but stopped. A simple reset test (units in training, a research under way, a dead hero) reproduced neither: the food count was right after it and the research was accepted again. They were not chased further.
• These were more than half of all games (the learner against itself and against past snapshots), in every run so far.

Every game now reloads the map (--scripted-reset 0, the default), in fgself-9 from 9.4M steps on (the table's last column). The lesson for environment changes: compare per-game statistics by kind of game (heroes, research, food), not only win rates.

The real game (the built-in AI without the curriculum's tax, 10% of the launches against it) before and after:

wins in the real game, fgself-9	all	human	night elf	orc	undead	easy AI	normal AI
the 4 hours before the fix (to 9.4M steps)	252/642 (39%)	30%	80%	21%	25%	34%	45%
the 2 hours after it	89/223 (40%)	30%	83%	12%	25%	36%	45%
2 to 4 hours after it (11M to 15M steps)	309/559 (55%)	43%	89%	41%	40%	48%	62%
5 hours after it (17M to 19M steps, five games per load)	214/375 (57%)	50%	88%	39%	49%	50%	64%
21M to 24M steps	242/407 (59%)	50%	81%	62%	43%	50%	69%
30M to 33M steps	276/426 (65%)	51%	88%	74%	54%	56%	73%
39M to 42.6M steps (the end, stopped for a shutdown)	390/522 (75%)	64%	91%	73%	70%	72%	78%

The races' columns are the measure: the restarts of that day skewed the mix of races in "all" and in the difficulties' columns (see the note on seeds above). Nothing else about the training changed in between (the speed work below changed how fast the same games are played), so the gain is the games between agents being whole games, or training that would have come anyway; there is no control run.

Over the run's last 20M steps human still trained almost no heroes (0.03 a game), and undead fewer (1.7 → 1.2) while it won more. fgself-9 played mirror matchups only (--mirror 1, as every run so far). The next run (fgself-10) goes on from its last checkpoint with mixed matchups: a curriculum level per matchup against the built-in AI and a tax between races in games between agents (--mirror 0, League.balance).

Mixed matchups (fgself-10)

fgself-9's last checkpoint (42.6M steps, mirror matchups only) in 320 real games against the normal AI over all 16 matchups (play.py --race all --ai-race all), the learner's race by row:

learner \ AI	human	orc	undead	night elf	all
human	59% (22)	29% (17)	50% (18)	0% (26)	33%
orc	65% (26)	85% (13)	71% (17)	7% (14)	59%
undead	48% (25)	11% (27)	50% (20)	0% (16)	28%
night elf	88% (25)	82% (17)	87% (15)	91% (22)	87%

70% in the mirror matchups it trained on, 44% in the others. Against the night elf AI the other races won 1 game in 56: night elf is the strongest race on duelrush (78% of the built-in AI's games). fgself-10 goes on from this checkpoint with fgself-9's recipe and mixed matchups (--mirror 0):

• the curriculum against the built-in AI keeps a level per difficulty and matchup (32), starting at 0.3 (a tax of 27% on the AI);
• games between agents of two races tax the stronger race's income (League.balance): a level per pair of races that the learner's games against itself move towards a 50% score (0.02 a game, a tax of at most 90%); games against past snapshots pay the same tax. In the first 150 games night elf's tax against human rose to 11%.

The tax was a subsidy at first. By 8.6M steps every pair's level had run to the end of its range, and the taxed side still won: night elf, taxed 90% against every race, won 76% of those games. A taxed night elf trained as many units as its opponent (36 against 38 a game) from a tenth of its income. The tax was the last of a step's commands: it set the taxed player's gold and lumber to what the observation showed minus the tax, which undid what that player's orders earlier in the same commands had just spent. A taxed agent trained and researched for free. The tax on the built-in AI was not affected, because the AI spends between steps. Now the tax is a step's first command, and the taxed agent sees what it has left. fgself-10 went on from 8.6M steps with the levels back at 0. In those 8.6M steps about 40% of the games (the mixed games between agents) gave one side free production.

fgself-10 ran to 38.7M steps (stopped for a shutdown, 2026-10-02 00:00). The real game, by steps:

steps	all	mirror	mixed	human	orc	undead	night elf	vs human AI	vs orc AI	vs undead AI	vs night elf AI	easy AI	normal AI
0–8.6M (the tax a subsidy)	73%	79%	70%	55%	80%	57%	98%	79%	63%	84%	66%	66%	78%
8.6–20M	75%	86%	72%	52%	86%	66%	97%	84%	71%	83%	62%	72%	78%
20–30M	80%	86%	78%	60%	89%	71%	98%	86%	80%	83%	70%	81%	79%
30–38.7M	79%	87%	76%	57%	89%	77%	98%	85%	77%	87%	67%	78%	80%

By matchup from 30M steps on (real games against both difficulties, about 40–100 a cell):

learner \ AI	human	orc	undead	night elf
human	69%	57%	70%	36%
orc	90%	97%	90%	81%
undead	89%	67%	90%	61%
night elf	96%	95%	100%	100%

Mixed matchups went from 44% (the start's 320 games against the normal AI) to 76%; the cells that were near zero rose most (undead against the orc AI 11% → 67%, orc against the night elf AI 7% → 81%, undead against it 0% → 61%, human against it 0% → 36%; the start's numbers are from the normal AI alone). Human stayed the weak race at 52–60% and trained almost no heroes (0.02–0.05 a game) all run. At the end the taxes between races stood at 25–58% on the stronger race (the most: orc against human 54%, night elf against human 58%), and the curriculum still eased only human's games (against the night elf AI 0.3–0.36, the easy orc AI 0.12).

Resumed on 2026-10-02, it got the insane AI as a third difficulty from 40.4M steps (--ai easy,normal,insane, a curriculum level per matchup starting at 0.3). In the real game against the insane AI it won 16 of its first 30 games. It stopped at 41.4M steps to move on to longer games.

Human's footmen and riflemen without heroes are a duelrush strategy: human built 0.02 altars a game against the human AI's 1.0, and its games with a hero were won less often (43% against 54%, 10k games). Two-minute games do not pay for a hero.

Longer games: duelfast (fgself-11)

duelfast is the duel map with the game's own speed, a third of the build, train and research times, and half the costs and hit points (no extra harvest). 48 built-in AI games (runs/fullgame/probe-duelfast, normal and insane, all races) lasted 5.9 minutes on average (median 5.4, longest 12.7, no ties within 15). The insane AI beat the normal AI in 20 of 23 games. Between different races orc won 17 of 22 and night elf 11 of 14, human 6 of 25 and undead 4 of 15 (few games).

fgself-11 goes on from fgself-10's last checkpoint on duelfast, with:

• a memory core (a minGRU added as a no-op, --memory 1): the games are three times longer and the enemy is out of sight most of the time;
• a longer horizon: gamma 0.999 instead of 0.997 (about 8 minutes instead of 3), and 10 updates of the value alone first;
• the auxiliary cloning loss on new takeover games on duelfast from fgself-10's policy (demos-fast-takeover-1*: the policy plays one side for 10 seconds to 5 minutes, then the built-in AI takes over);
• the rest as fgself-10: mixed matchups, the curriculum per matchup and difficulty (easy, normal, insane), the tax between races, the clone's KL term (now towards fgself-10's policy).

The takeover collection was bound by its own Python (24 games' policy calls in one process; the machine 65% idle), so three processes collected side by side.

fgself-11 ran 14.8M steps (5 hours) and stalled:

	the first hour	hours 3 to 5
curriculum games won (the AI taxed ~0.65)	1–3%	31–34%
tied at the time limit	17%	33–41%
the real game	0 of 80	4 of 346 in all (1%)

Fixes on the way:

• The clone's KL term spiked (96 and 4.4, gradient norms of 726 and 39 in two updates): the cloning loss on the new demonstrations allowed new (unit type, order) pairs to the learner but not to the clone, whose log-probability for them stayed at -1e9. The clone now gets the learner's allowed orders.
• Games tied at 15 minutes (59% of the curriculum games at one point, the learner ahead in material in most), and a tie leaves the curriculum's level alone: 20 minutes now.
• Six games per launch kept a launch's races and opponent for 10 minutes of duelfast games: the real games came in streaks of one matchup. Two now.

What it did in the real game, against the AI, per game (the last 150): 32.6 workers trained against 10.1, an army of 23 against 28, 38 food at most against 60, 0.1 heroes against 1.6, almost no research (the AI researched in a quarter of its games), 8 kills against 42. These are duelrush habits: there, walking workers carried 6.5 times the gold, so more workers paid, and a game was over before heroes and research paid. At the AI's own hero decisions in the duelfast demonstrations the policy gave the hero order 0.6% (fgself-10: 2.5%, orc 8%), and a stronger cloning loss (×0.02 for an hour) did not move it.

The clone's KL term held the policy to fgself-10's, a duelrush policy: its distance stayed at 0.002 a unit for the whole run. AlphaStar's KL term is towards a supervised policy of the same game. So the next run starts from a clone of the built-in AI on duelfast (fgself-10's network fine-tuned on 3000 built-in AI games on duelfast and the 1017 takeover games), with that clone as the anchor.

The clone, fullgame-fast-1 (3 epochs, 46 minutes): validation loss 2.48 → 2.38, the AI's order 77% of the time, its target 94%, the value explaining 90% of the returns' variance. It lost all 45 real games it played against the normal AI (6 minutes on average, 15 food at one minute against 21).

fgself-12 (from the clone, the clone as the anchor; memory core, gamma 0.999, 20-minute games, 2 games per launch, the cloning loss on all the duelfast demonstrations) ran 12.9M steps (4 hours). Its habits stayed the AI's: 1.2–1.4 heroes a game against the AI's 1.4 (fgself-11: 0.1), about one altar a race, 14–15 workers against 10 (fgself-11: 32). But it did not learn to win:

	first hour	last hour
curriculum games won	4–9%	1–4%
tied at the time limit	14–29%	78–86%
the AI's tax	0.32–0.43	0.67–0.76
the real game	1 of 234 in all

• Ties. In the tied games it was ahead in material (lead 1.6, ahead in 79%), with armies even and kills even, and the AI kept ~11 buildings. A tie with that lead was worth ~+0.5 against +1 for a win: --tie-break 0.1 (from 5.7M steps). Ties went on rising.
• The curriculum stood still: a tie left a level where it was, so with 78% ties the AI's tax stayed at 0.67 for an hour. --curriculum-tie 0.5 (a tie raises a level by half a loss's step, from 10.3M): the levels rose from 0.72 to 0.88 on average (26 of 48 at the most) in an hour; the curriculum games won only 4% at a tax of 0.76.
• It does not spend. It ended games holding ~3,600–3,800 gold and lumber (the AI ~800), gathered 40% less gold than the AI with half again as many workers, and trained fewer basic units than the AI (21.7 footmen against 33.7, 22.7 ghouls against 35.4) but more supply buildings (1.7 times) and some high-tech units. Its material shaping counts units and buildings but not what is held, so spending was already rewarded.

Why it holds its resources (scripts/production_probe.py: games on the CPU, the policy's production buildings looked at every step). Its barracks (and ancients of war, crypts, the great hall) were busy 23–42% of the time and idle with something affordable 50–75%; there the policy trained in 4–9% of the steps. The clone fullgame-fast-1 was the same: the built-in AI keeps its production queued, so its games hardly show an idle barracks with money in the bank, and the clone did not learn what to do there. The KL term (×0.2) then held the policy to the clone's rate: from 5% to 50% at such a building is ~0.8 nats a step, against a few hundredths of shaping reward for the unit it buys.

From 13.0M steps fgself-12 goes on with the held gold and lumber counted against the shaping's potential (--float-penalty 0.5: spending is credited at the order, not ~14 steps later when the unit appears, half the credit after GAE's lambda) and the KL term ×0.05.

The KL weight ×0.05 let RL lower the production orders: the chance the policy gave the AI's own hero orders at the AI's decisions (scripts/order_probability.py) fell from 7.8% (the clone) to 2.7% at 14.1M steps, its basic units from 39% to 20%. From 15.07M steps the KL term is ×0.2 again and the cloning loss ×0.02: 5.9% and 32% at 16.9M.

fgself-12, 30-minute windows	60–90 min before 16.9M	30–60 min	the last 30 min
curriculum games won	3%	5%	9%
tied	95%	90%	83%
lost	2%	5%	7%

At 17.8M steps (5.5 hours) it had won 2 of 334 real games. Games lasted ~19 of 20 minutes at a tax of 0.84. It held ~4,000 gold and lumber against the AI's ~330 and had the larger army (28 food against 17), but killed 7 units a game and lost 15: it out-builds the AI and does not attack.

The real game is not a tie. The untaxed AI wins in 5–7 minutes. The last 150 real games against the curriculum games of the same hours:

per game (learner vs AI)	real, normal AI	real, insane AI	curriculum (tax 0.84)
result	49 of 50 lost	62 of 62 lost	83% ties
minutes	6.7	5.3	19
food at 1 minute	15 vs 21	16 vs 20	15 vs 7
most food	34 vs 53	31 vs 62	72 vs 34

The tax cuts the AI to 7 food at one minute, so the curriculum hides the policy's slow opening. The clone was already behind (15 against 21). Many matchups were at level 1.0 (the AI without income) and the policy still tied them.

Learn from short wins (AlphaStar fine-tuned its supervised policy on winning replays). Of the 3,000 built-in AI games on duelfast, 13 were ties and 55% were decided within 6 minutes. fullgame-fast-win6: the clone fine-tuned on the winning sides of the games decided within 6 minutes (2,180 of 8,034 game sides, takeover games included; bc --winners-only --max-minutes 6). fullgame-fast-win6 (4 epochs, the best the third): validation loss 2.314 on held-out short wins, the AI's order 81% of the time.

fgself-12 goes on from 17.77M steps with it as the KL anchor and its data as the cloning loss (--bc-winners 1 --bc-max-minutes 6). The cloning data also gets takeover games from fgself-12's own states, with the AI taking over at 0.5–5 minutes, before the real games are lost (demos-fast-takeover-12).

After the switch (17.77M to 21.0M steps, 1.5 hours): curriculum games won 9% → 24%, tied 83% → 69%. The real game: still no win in 46 games, but games lasted 9.7 minutes instead of 7.6 and the policy killed 12 units a game instead of 7.5. The chance of the AI's hero orders: 6.3% (the old clone 7.8%, fullgame-fast-win6 8.6%), basic units 30% (39%, 39%). The takeover collection ended with 600 games: 512 taken over (the rest lost before the takeover), 102 won by the taken-over side within 6 minutes of it. From 21.0M steps the cloning loss reads all 600.

After 50 minutes of distillation (24.13M → 25.65M steps): the real game (never advised) 13 of 50 won (26%: easy 5 of 14, normal 7 of 18, insane 1 of 18; before: 1 of 98), 6% ties, 8.5 minutes a game, food at one minute 19.0 against 21.6. Unadvised curriculum games 82% won with the AI's tax falling 0.84 → 0.60; unadvised self-play 13% ties (82% before). The policy agrees with the advisor's order 59% of the time; its KL per update 0.003, its distance to the anchor 0.006.

After 2 h 50 min of distillation (31.7M steps, 17:49): the real game 55 of 227 won (24%: easy 33%, normal 35%, insane 9%, insane 18% in the last 30 minutes). The chance of the AI's own orders at its decisions: basic units 40% (30% at 20.9M; the clones 39%), heroes 7.2% (6.3%; the clones 7.8-8.6%): distillation brought production back to the clones' level. The curriculum's tax fell to 0.31.

A plateau after 3 hours (33.4M steps): the real game held at ~24% in every half hour since the first (normal 30-40%, insane ~8%), and the agreement with the advisor at ~58% (the distillation loss 2.6, flat). The real games are lost on tempo:

real games against normal and insane (learner vs AI)	won (50)	lost (171)
food at 1 minute (normal)	21.4 vs 20.5	16.9 vs 22.5
tier 2 reached	100% vs 50%	63% vs 91%
heroes	2.0 vs 1.3	1.2 vs 2.0
minutes	8.3	6.2

In games of 6 minutes or more both sides reach tier 2 (94% and 86%): the learner techs, but late. It trains 1.3-2.5x the AI's workers a minute (human 4.4 vs 2.3, undead 1.3 vs 0.5) and holds twice the resources (2,800 vs 1,300). Most losses to insane come in minutes 3-5. Human wins 2 of 46 such games, orc 14%, undead 23%, night elf 39%. From 33.4M steps the distillation loss weighs x0.15 (was x0.05).

3 h 20 min at x0.15 (33.4M → 46.4M steps, 21:56): the real game 95 of 366 won (26%: easy 34%, normal 40%, insane 4%), so no change. The agreement with the advisor rose 58% → 64%; the KL per update stayed at 0.0033, the distance to the anchor rose 0.006 → 0.0095. Against normal the tempo is now even, against insane it is not:

real games, learner vs AI	normal (118)	insane (120)
food at 1 minute	19.1 vs 19.2	18.5 vs 20.9
tier 2 reached	86% vs 73%	72% vs 94%
heroes	1.8 vs 1.6	1.4 vs 2.2
workers per game minute	1.55 vs 1.11	2.21 vs 1.35
kills / units lost	0.70	0.30

The losses are decided in the fights: in real losses against normal and insane the learner trained about the AI's army (27 vs 25 units) but killed 8.8 units for 40 lost (in wins 39 for 17). The chance of the AI's own orders at its decisions (46.35M): heroes 5.1% (4.9% at 40.8M, 7.2% at 31.6M), basic units 33% (40% at 31.6M). By race the heroes are human 3.1% (the clones 11.5-12.6%), undead 1.8% (4.9-5.9%), night elf 4.6% (5.1-5.3%), orc 9.8% (8.4-9.6%); human basic units 31% (45-46%). Those are the races that lose: against normal and insane human won 2 of 48, undead 8 of 62, orc 13 of 66, night elf 29 of 62.

Without the cloning loss (from 58.0M steps, 2026-10-10 19:11; --bc-coef 0, was 0.02). The distillation teaches the built-in AI's orders at the policy's own states; the cloning loss taught them at the AI's states in its winning games, at a third of an update's time. The KL anchor to fullgame-fast-win6 stays (0.2). Before it, since the distillation weighs x0.15: the real game 150 of 589 won (25%: easy 31%, normal 38%, insane 6%; human 2 of 75 against normal and insane). What to watch: the real game, the chance of the AI's production orders at its decisions, the distance to the anchor. At 58.03M, just before: heroes 6.0% (human 7.5%, undead 3.3%; at 46.35M 5.1%, human 3.1%), basic units 33% (human 39%, orc 23%; the clones 39%).

Without it (58.0M → 72.8M, 3 hours): the real game 100 of 443 won (23%: easy 28%, normal 30%, insane 9%), against 25% before (not a clear difference). But the chance of the AI's hero orders at its decisions fell steadily, 6.0% → 5.1% (62.1M) → 4.0% (72.2M; human 7.5% → 3.2%, undead 3.3% → 1.9%, no order at all 76% → 85%), while basic units held (33-35%). In its own real games it still trained as many heroes (1.4-1.5 against insane's 2.2). So only the cloning loss had kept the AI's habits at the AI's own states. From 72.8M (22:10) it is back at x0.02 on half the batch (32 steps a minibatch: an update 2.24 s alone, against 2.64 s with 64 and 1.59 s without).

With it back (72.8M → 81.8M, 2 hours): the real game 60 of 250 (24%: easy 24%, normal 44%, insane 6%), the AI's hero orders 4.5% at 81.4M (from 4.0%; human 4.6%).

The anchor halved (from 82.8M, 2026-10-11 00:17; --ref-kl 0.1, was 0.2). For 40M steps the policy sat ~0.01 nats from fullgame-fast-win6 and the real game at ~24-25%; a weaker pull lets the reinforcement learning move it further. Watch: ties and held resources (this run hoarded and tied without an anchor), the distance to the anchor, the AI's production orders, the real game over 300+ games.

Stopped for a shutdown at 85.9M (00:55) after 30 minutes: the real game 21 of 93 (23%), ties as before (self 18%, past 26%), resources held at a real game's end 3,261 (3,506 before; the AI ~1,390), the distance to the anchor 0.010 → 0.014. Not judged yet: it resumes from 0000000085906150.pt with the same flags.

The anchor at 0.1, judged (302 real games, 82.8M → 93.1M): against normal and insane 42 won where the built-in AI would win ~96 in the same matchups (44% of its record, as before: 43-45%), but ties rose in every kind of game: self-play 18-22% → 27%, past snapshots 20-25% → 31%, unadvised curriculum ~15% → 27% (resources held at a real game's end stayed lower: ~3,200 against ~3,500). No gain for the run's old drift towards ties: back at 0.2.

A quarter of the AI games real (from 93.1M, 2026-10-11 12:47; --real-share 0.25, was 0.1, with the anchor at 0.2): more of the training against the untaxed built-in AI, the yardstick's own opponent. Judged on 300+ real games against the AI's record in the same matchups (43-45% before), ties and the curriculum's tax.

Judged after an hour (93.0M → 98.1M, 328 real games): 35% of the AI's record against normal and insane, 28% after the first half hour (13 won where the AI would win 46), against 41-45% in every earlier window; back at 0.1 from 98.1M (13:48). Ties had been creeping up since the evening before, independent of these changes: self-play 17-19% (18-21h on the 10th) → 20% (22h-0h) → 26-30% (11-13h on the 11th) → 33%, past snapshots 19-20% → 35%, unadvised curriculum 13-15% → 25-28%; real games stayed at 7-16%. The run is watched for whether ties and the yardstick recover with the recipe as before both experiments (anchor 0.2, real share 0.1).

Rolled back (2026-10-11 14:16): the run reached its 100M target at 14:09, the last hour at 28% of the AI's record and ties at 30-37%. It goes on to 200M from the policy before both experiments (0000000082830609.pt, 44-45% of the AI's record, ties ~20-24%; selfplay_restart.py --weights), with the recipe as then (anchor 0.2, real share 0.1, the cloning loss on 32-step batches, 48 games). If ties rise again from there, the suspects are what changed the evening of the 10th (48 games and 1.8 instead of 1.3 updates of staleness from 21:17, the cloning batch 32 from 22:10).

They did: from the roll back, ties rose again within 1.5 hours (self-play 22% → 30%, past snapshots 23% → 31%, curriculum 16% → 27%), so the drift comes from the training as it is now, not from the policy. Between 19:10 and 21:17 on the 10th (40 games, no cloning loss) they held at 17-19%. Test 1 (2026-10-11 16:23, from 0000000082830609.pt again): the cloning loss on its full batch (--bc-batch 64, was 32); everything else as before (48 games; the inference server now the C++ loop, which answers as the Python server did). If ties stay near 20% for two hours, the half batch was the cause; if not, test 2 is 40 games.

Test 1 after two hours (16:23 → 18:28, 110.8M → 121.1M): ties flat instead of rising (self-play 26, 27, 24, 27%; past snapshots 30, 27, 31, 26%; curriculum 20, 17, 17, 24% by half hour; after the first roll back they had risen 8-11 points in the same time), though above the 17-20% of the evening of the 10th (the league kept the past snapshots of the drift). The real game after the first half hour: 40 of 198 won against normal and insane where the AI would win ~92 (44% of its record, its level before the experiments). The AI's hero orders at its decisions 5.2%, basic units 34% (121.0M). The cloning loss stays on its full batch (64).

Moved to native Linux (2026-10-11): stopped at 19:10 at 124.56M steps on WSL and resumed at 21:19 on the same machine under native Ubuntu 26.04 (Python 3.14, torch 2.14 with CUDA 13), with the same recipe and checkpoint. The first launch lost its learner at its first update (Python 3.14's new default for starting processes broke the cloning loss's data workers; fixed). The real-game yardstick restarts from 21:19, and its first half hour is discounted as after any restart. --lr 6e-5 stays queued until the native baseline has 300+ real games. Throughput work restarted the run about a dozen times between 21:19 and 00:49 (each restart front-loads short lost games; the recipe never changed); one restart (23:54) died after three updates of a device-side assert in the inference loop (stale padding, fixed). From 00:49: ~2,100 steps/s (the GPU shared better between learner and inference, --update-pause-ms 4), staleness ~2 updates instead of ~1.3: the real-game baseline for --lr 6e-5 starts there. Paused 01:00-09:47 on the 12th for a machine shutdown (143.20M); from 09:55 the games run with the equivalent speedups of their own code (--game-cuts 1529, the same simulation): ~2,280 steps/s. The baseline starts at 09:55.

Why the fights are lost: group orders scatter (2026-10-12). When several of the policy's units got the same point order (attack-move, move, smart) in one step, their targets spread a median 304 world units from their mean (groups of 5+: 713), and it gave such group orders ~6.5 times a game; the built-in AI's spread 0 (groups of 5+: 10) and it gave ~69 a game. Every unit samples its target bins with a noise draw of its own, so a group's units go to different points. (The AI gives unit-target attacks to groups only 0.4 times a game: it attack-moves and lets units acquire targets.) The policy does not over-order: 1.0-1.3 orders a step against the AI's 1.96.

First try, in fgself-12 (11:18-11:28): half the games drew one noise vector a step for all of a side's units, orders included. Those games built fewer buildings (15.6 against 18.9 a game), had less food at one minute (16.4 against 18.6) and held more (4,051 against 3,222): with one draw, similar units took a rare order (build a farm) all together or not at all.

fgself-13 (the first experiment as a run of its own, from fgself-12 at 155.16M, scripts/selfplay_fork.py; the anchor stays fullgame-fast-win6): the same A/B with the shared noise for the targets only (unit, x, y; the orders keep a draw each), half the games at random (--shared-noise 0.5; episodes record shared_noise). Judged per arm on real games against the AI's record (real_games.py --split shared_noise) and on the scatter of group targets. A new league and curriculum.

fgself-13 after an hour (11:29-12:28, 7.0M steps): no gain. Real games against normal and insane 13 of 56 with independent noise (48% of the AI's record) and 12 of 66 with shared target noise (38%); unadvised curriculum games 33% and 30% won (~420 each), kills per unit lost 0.77 and 0.80. The shared noise did not tighten groups of 5+ (their targets' spread 353 against 320 world units): a group's units stand in different places and want different points, so one draw aligns only units that already agree. Coherent groups need a representation for them (a group command, or features of where each unit is already sent). Ties were higher than in fgself-12 in both arms (self-play 38%, past snapshots 40-42%), with a new league whose snapshots are close to the policy.

fgself-14 (from fgself-13 at 7.0M, 12:31): two low-risk fixes together. The anchor's KL term as forward KL over the orders (--ref-kl-mode forward, AlphaStar's direction: reverse KL's pull on an order shrinks with the policy's own probability of it, so orders RL drove down stayed down). And two unit features with zero weights at first: the hit points' change since the last step and the position as sines and cosines at 5 frequencies (F 29 → 50). The shared-noise A/B goes on at 0.5.

fgself-14 after two hours (12:31-14:27, 13.6M steps; its first half hour discounted): against normal and insane 46 of 184 won where the AI would win 97 (47% of its record), falling window by window (60% → 55% → 33% in the last half hour); all real games 28% won (fgself-12: ~24%). Ties: self-play 32%, past snapshots 34%, curriculum 31%. No clear gain within two hours; features with zero weights at first may need longer, so they stay.

fgself-15 (from fgself-14 at 13.6M, 14:30): where each unit is already sent. Each own unit's last order the player gave it: its target relative to the unit (a point, or the target unit where it is now), point or unit, and how long ago (F 50 → 55; zero weights at first). Everything else as fgself-14.

A faster yardstick: evaluation games (2026-10-12, --eval-games-per-actor). The real games had come from 10% of the AI games (~200 an hour), and a verdict of ±7 points on the AI's record took two hours. Each actor now also plays games that only evaluate: the real game against normal or insane (half each), races at random, never advised, nothing trained on (episodes eval: true; real_games.py --split eval). A real game is ~6.5 minutes of game time but only ~16 s of wall time, so 8 such slots (one per actor) play ~1,200 an hour: ±4-5 points in ~25 minutes, at ~10% of the training's throughput (16 slots cost a quarter). fgself-15's policy (from fgself-14's, 14:37-14:49): 125 of 457 won, 54% of the AI's record, ±4.5.

Forked runs lose ground in their first hours. fgself-15's evaluation games fell from 54% of the AI's record (14:37-14:50, 457 games) to 49% (14:51-15:10, 422) and 44% (15:10-15:30, 349); fgself-14's real games fell 60% → 33% in two hours. The learner's order entropy is ~0.122 since fgself-14 against ~0.107 in fgself-12: forward KL to the anchor makes the policy cover everything the anchor gives, so it spreads its orders.

fgself-16 (from fgself-15 at 4.8M, 15:30): the anchor term back to reverse KL; everything else as fgself-15 (its features). Watched on the evaluation games' trend (4 in all, ~600 an hour: 8 cost ~20% of the throughput, --eval-games). A halt or a recovery points at forward KL.

fgself-16's evaluation games kept falling: 35% (15:30-15:38, 100 games), 30% (15:38-15:46, 109), 27% (15:46-15:58, 137). The decline runs on through the fork (54 → 49 → 44 → 43 in fgself-15, then 35 → 30 → 27): it is the training, not the fork's first minutes, and reverse KL did not stop it within half an hour. In the real games, heroes and research fell with it (fgself-14 1.85 heroes and 8.5-8.9 research a game, fgself-16 1.46 and 4.6).

fgself-17 (control, from fgself-12 at 155.16M, 16:00): fgself-12's recipe (reverse KL, the unit features as before with --features 29: the new channels zeros, no shared noise) plus 4 evaluation games. Holding near fgself-12's level would put the decline on today's changes; falling too would put it on forking (a new league and curriculum, a fresh optimizer: checkpoints do not keep Adam's state).

fgself-17 (16:00-16:30): 49 of 367 evaluation games won, 28% of the AI's record (22%, 29%, 29% by ten minutes): fgself-12's final policy is far below fgself-14's (54%) on the evaluation games. The training real games read higher than the evaluation games for the same policy in the same windows (fgself-15 59% against 50%, fgself-16 52% against 36%, fgself-17 45% against 25%; 44-79 training games a window), so runs are compared on the evaluation games only. Read so, fgself-13/fgself-14 (shared target noise, forward KL, hit-point-change and position features, a new league) raised the policy from ~25-29% to ~54%, and the decline began with fgself-15's last-order features.

fgself-18 (from fgself-14 at 13.6M, 16:30): fgself-14's recipe (forward KL, --features 50: the last-order channels zeros, shared target noise 0.5) plus 4 evaluation games. Holding near 54% puts the decline on the last-order features.

Head to head (--eval-checkpoints, 16:35-17:25 in fgself-18: its 8 evaluation games split at random between the current policy and three fixed checkpoints, the same conditions at the same time; ~275 games each against normal and insane):

policy	of the AI's record
fgself-14's last (13.6M)	59%
fgself-18 now (fgself-14's recipe without the last-order features)	58%
fgself-16's last (with the last-order features)	39%
fgself-12's last (155.16M, the plateau)	28%

fgself-13 and fgself-14 doubled the policy's record against the AI (28% → 59%); the last-order features undid much of it (39%); without them fgself-18 holds the gain. Next head to head: fgself-13's last against fgself-12's and fgself-14's, to see which run made the jump.

Second head to head (17:28-17:58, ~150 games each, ±8): fgself-18 now 57%, fgself-14's last 43%, fgself-12's last 38%, fgself-13's last 19%. Pooled over both rounds: fgself-18 58%, fgself-14 53%, fgself-12 32%, fgself-13 19% (the rounds differ by up to 16 points for one checkpoint: ~150-280 games each are still noisy). fgself-13 (shared target noise, a new league) was no better than fgself-12; the jump came with fgself-14's changes: forward KL over the orders, and/or the hit-point-change and position features.

fgself-19 (from fgself-18 at 7.56M, 18:10): the target heads at temperature 0.8 (--target-temperature, a buffer of the model now, so a past snapshot in an inference slot keeps its own). Its evaluation games play the current policy against fgself-18's last checkpoint (the same weights at 1.0) and fgself-14's last.

fgself-19 after 45 minutes (18:10-18:55, ~300 evaluation games each): the current policy at temperature 0.8 49% of the AI's record, fgself-18's last (the same weights at 1.0) 46%, fgself-14's last 51%. No measurable effect of the temperature (it stays at 0.8, as AlphaStar's). The line fgself-14 → fgself-19 holds ~50%.

The runs since fgself-15 had updates of ~4.6 s instead of ~3.5: the last-order features' per-unit loop ran in the cloning loss's data workers even when --features 50 zeroed them. It is skipped now (an update alone 3.2 → 2.7 s).

fgself-20 (from fgself-19 at 4.23M, 19:05): a value with the opponent's information. The global vector ends with a privileged block (23 numbers: the opponent's resources and supply; each side's workers, army, buildings, heroes, the army's hit points and centre, seen or not) that only the value head reads, through a zero-initialised projection. Its evaluation games play the current policy against fgself-19's last (the same policy at the start) and fgself-14's last.

fgself-20 after 1 h 23 min (19:05-20:28, 568 evaluation games each): the current policy 49% of the AI's record (42%, 49%, 53% by window), fgself-19's last 55%, fgself-14's last 51%. The value fit better (its loss 0.0063 against 0.0075) but the policy did not gain measurably. The line holds ~50-55% (fgself-12 32%).

fgself-21 (attribution, from fgself-13's last at 7.0M, 20:28): what made fgself-14's jump (19% → ~51-53%)? fgself-14 changed forward KL and two unit features; this run repeats it with forward KL only (--features 29), shared target noise 0.5 as before. Evaluation games: the current policy against fgself-13's last (its start) and fgself-14's last. (It also has the value with the opponent's information, now always on; fgself-20 found no measurable effect of it.)

fgself-21: forward KL alone made the jump. Head to head over its first hour (20:28-21:30, ~240-280 evaluation games each): the current policy (forward KL from fgself-13's last, no new features) 54% of the AI's record, fgself-13's last (its start) 28%, fgself-14's last (forward KL and the features) 49%. So forward KL to the anchor over the orders took the policy from ~28% to ~54% within an hour; the hit-point-change and position features added nothing measurable (the line goes on with --features 29). Reverse KL's pull on an order shrinks with the policy's own probability of it, so heroes and production that RL drove down stayed down; forward KL pulls them back towards the clone's.

fgself-22 (from fgself-21 at 4.87M, 21:29): the forward KL term stronger, ×0.5 (was 0.2). Evaluation games: the current policy against fgself-21's last and fgself-14's last.

fgself-22 (forward KL ×0.5, 21:29-22:26, ~310-350 evaluation games each): the current policy 50% of the AI's record, fgself-21's last (its start, ×0.2) 61%, fgself-14's last 54%. A stronger anchor did not help; ×0.2 stays. As in fgself-20, the run's policy evaluated a little below its starting checkpoint: from a good start the training holds ~50-60% rather than climbing.

fgself-23 (from fgself-21's last at 4.87M, 22:29): forward KL over every head (--ref-kl-mode forward-all: the orders, the target unit, the point's x and y; ×0.2), as AlphaStar anchors all action arguments.

fgself-23: forward KL over every head helps too. Head to head over its first hour (22:29-23:27, ~340-380 evaluation games each): the current policy 69% of the AI's record (70%, 67% by half hour), fgself-21's last (its start) 56%, fgself-14's last 53%. Ties fell from ~25-35% to ~10% (self-play 10%, past snapshots 11%, curriculum 7-9%); the curriculum's tax on the AI fell 0.29 → 0.24. Anchoring the targets (the unit, x and y) to the clone's makes the policy finish its games. From the morning's ~28-32% the policy is at ~69% of the built-in AI's own record on the evaluation games.

fgself-23 held rather than climbed: 70%, 65%, 67% of the AI's record by half hour (22:29-23:57); the curriculum's tax on the AI fell to 0.20 while the policy won 58% of those games.

fgself-24 (from fgself-23's last at 10.2M, 23:57): a moving anchor. The KL term's reference (--init) becomes fgself-23's last checkpoint instead of the clone fullgame-fast-win6 (DeepNash and ProRL move theirs to a recent policy): the clone gave the policy the production habits it has now, and the pull back to its weaker parts may cap it. Forward KL over every head ×0.2. Evaluation games: the current policy against fgself-23's last (its start and its anchor) and fgself-21's last.

fgself-24 after 27 minutes (23:57-00:24, stopped for a machine shutdown at 3.14M): the current policy 76% of the AI's record (135 evaluation games), fgself-23's last (its start and anchor) 60% (175), fgself-21's last 53% (173). Promising but early (±9): to be judged after it resumes.

fgself-24: the moving anchor helps. Head to head (23:57-00:25 on the 12th and 08:30-09:00 on the 13th after the shutdown, ~290-340 evaluation games each): the current policy 71% of the AI's record (76%, 67%), fgself-23's last (its start and anchor) 60%, fgself-21's last 52%. Ties 14-15%. Where it still loses (fgself-23's own 540 evaluation games): against normal it matches the AI's record (101%), against insane 41%; orc and undead are weakest against insane (34%, 25%), human and night elf reach 49-56%.

fgself-25 (from fgself-24's last at 6.22M, 08:59): the anchor moved again, to fgself-24's last; nothing else changed. Evaluation games: the current policy against fgself-24's last and fgself-23's last. Next in line: training on normal and insane only (--ai normal,insane: a third of the AI games are still against easy).

fgself-25 (the anchor moved again, 08:59-10:00, ~340-370 evaluation games each): the current policy 84% of the AI's record (89%, 79% by half hour), fgself-24's last (its start and anchor) 77%, fgself-23's last 59%. The second move added less than the first (+7 against +11, and level in the second half hour).

fgself-26 (from fgself-25's last at 6.66M, 09:58): the AI games against normal and insane only (--ai normal,insane; a third had been against easy), the anchor where fgself-25 had it (fgself-24's last). Evaluation games: the current policy against fgself-25's last and fgself-24's last.

fgself-26: training against normal and insane only helps. Head to head (09:58-11:00, ~330-370 evaluation games each): the current policy 87% of the AI's record (87%, 87% by half hour), fgself-25's last (its start) 74%, fgself-24's last 77%; against insane alone 48% against 38% and 38%.

fgself-27 (from fgself-26's last at 6.69M, 10:59): the AI games against insane only (--ai insane), the anchor unchanged. Evaluation games: the current policy against fgself-26's last and fgself-25's last.

fgself-27 (insane only, 10:59-12:00, ~330-370 evaluation games each): the current policy 78% of the AI's record (normal 118%, insane 45%), fgself-26's last (its start) 72% (115%, 34%), fgself-25's last 66% (100%, 35%). A modest gain, all of it against insane; normal kept. (The same checkpoint reads differently from run to run, fgself-26's last 87% in its own and 72% here: compare within one run's evaluation games.)

fgself-28 (from fgself-27's last at 6.67M, 11:59): the anchor moved again, to fgself-27's last (it had been fgself-24's since fgself-25); insane only as before. Evaluation games: the current policy against fgself-27's last and fgself-26's last.

fgself-28 (the anchor moved a third time, 11:59-13:00, ~340-380 evaluation games each): the current policy 82% of the AI's record (insane 55%), fgself-27's last (its start) 88% (insane 51%), fgself-26's last 69%. The anchor's moves gave +11, +7 and now nothing: fgself-27's last stays the best.

fgself-29 (from fgself-27's last, with fgself-27's flags, 12:59): the distillation from the built-in AI twice as strong (--opd-coef 0.3, was 0.15): the insane AI advises half the games at the policy's own states, and insane is where it loses. Evaluation games: the current policy against fgself-27's last and fgself-26's last.

fgself-29 (distillation ×0.3, 12:59-14:00, ~340-360 evaluation games each): the current policy 79% of the AI's record (insane 45%), fgself-27's last (its start) 87% (49%), fgself-26's last 65%. No gain.

How the best policy loses to insane (fgself-27's last in fgself-29's evaluation games: 49 won, 139 lost): in the losses it is behind early (food at one minute 17.9 against the AI's 22.1; the wins 20.9 against 18.0), builds twice the AI's workers (18.6 against 8.8) while the AI builds an army, reaches tier 2 less often (78% against 99%) and loses the fights (13 kills for 48 units lost; the wins 40 for 16). Most losses are against orc (53 of 139). It does not answer early aggression, and its fights are inefficient: next, structural changes (credit per unit for the fights, a memory of enemy units seen).

fgself-27 resumes (14:00) with its recipe to see whether longer training improves it; its evaluation games against its own last checkpoint before (6.67M) and fgself-26's last.

fgself-30 (from fgself-27's 6.67M, 14:05; fgself-27 had resumed at 14:00 for longer training and was stopped for this): credit per unit, OpenAI Five's team spirit (--unit-credit 0.3 --unit-gamma 0.8). Each own unit's advantage also gets a share of its own: the material it lost (hit points, death) and killed (DEATH events' killer) over the next few steps, relative to the step's other units (the team's credit stays the step's advantage) and normalised over the batch, ×0.3. Evaluation games: the current policy against fgself-27's 6.67M and fgself-26's last.

fgself-30 (credit per unit, 14:05-14:57, ~290-330 evaluation games each): the current policy 82% of the AI's record (normal 118%, insane 54%), fgself-27's 6.67M (its start) 90% (131%, 54%), fgself-26's last 70%. No gain; against insane on par.

Forks dip in their first 20 minutes. The current policy's share by 20-minute windows: fgself-28 83 / 77 / 87%, fgself-29 71 / 86 / 77%, fgself-30 70 / 88 / 86% (~85-125 games a window, ±9). After its first 20 minutes fgself-30 read 87% against fgself-27's 90%. A fork started Adam afresh: its first steps move each weight by ~lr whatever the gradient, and the value warm-up's updates pull only toward the anchor and the clones. Checkpoints now keep Adam's state (<steps>.adam, the newest three), and forks and resumes from them load it. Verdicts also look at the games after the first 20 minutes.

fgself-31 (from fgself-27's 6.67M, 15:01; no Adam state for it yet): UPGO, AlphaStar's upgoing policy update (--upgo 0.5; its ablation: 73% → 82% average win rate). Each step's advantage also gets ×0.5 of its UPGO advantage, normalised over the batch: the return follows the trajectory while the actions after did at least as well as the value expected (Q ≥ V), else bootstraps from the value, so a good action's credit is not cancelled by worse ones later. Per-unit credit off. Evaluation games: the current policy against fgself-27's 6.67M and fgself-30's last.

fgself-31 (UPGO ×0.5, 15:02-15:57, ~300-330 evaluation games each): the current policy 102% of the AI's record (normal 145%, insane 67%), fgself-27's 6.67M (its start) 79% (122%, 41%), fgself-30's last 74%. Ties 4%. A gain of ~23 points, the largest since the forward KL; the first policy above the built-in AI's own record over normal and insane. It showed from the first 20 minutes (102%, then 103% and 99%): no fork dip, though this fork started Adam afresh too.

fgself-32 (from fgself-31's 5.92M, 15:58, with its Adam state: the first fork that carries it): UPGO ×1.0. Evaluation games: the current policy against fgself-31's 5.92M (its start) and fgself-27's 6.67M.

fgself-32 (UPGO ×1.0, 15:58-16:57, ~320-340 evaluation games each): the current policy 111% of the AI's record (normal 156%, insane 73%), fgself-31's 5.92M (its start) 94% (134%, 58%), fgself-27's 6.67M 82%. Ties 5%. A gain of ~17 points, rising over the hour (98%, 110%, 122% by 20-minute windows). No fork dip with Adam's state carried.

fgself-33 (from fgself-32's 6.41M, 16:58, with its Adam state): the anchor moved to the UPGO policy (--init = fgself-32's 6.41M; it was fgself-24's 6.22M, ~60% of the AI's record). UPGO ×1.0. Evaluation games: the current policy against fgself-32's 6.41M (its start and anchor) and fgself-27's 6.67M.

fgself-33 (anchor moved, 16:58-17:57, ~300-340 evaluation games each): the current policy 113% of the AI's record (normal 153%, insane 82%), fgself-32's 6.41M (its start) 108% (156%, 66%), fgself-27's 6.67M 87%. Ties 8% against 6%. A small gain, mostly against insane (±7 each there); by 20-minute windows 111%, 128%, 97%. By race against normal and insane: human 44/86, night elf 66/80, orc 33/57, undead 29/78.

fgself-34 (from fgself-33's 6.25M, 17:59, with its Adam state): UPGO ×2.0; the anchor stays fgself-32's 6.41M. Evaluation games now two thirds against insane (--eval-ai normal,insane,insane): normal is at ~150% of the AI's record, insane is the gap. Against the current policy: fgself-33's 6.25M (its start) and fgself-27's 6.67M.

fgself-34 (UPGO ×2.0, 17:59-18:57; evaluation 2/3 against insane, so its shares read lower than the runs before): the current policy 89% of the AI's record (normal 145%, insane 67%; 256 games), fgself-33's 6.25M (its start) 85% (131%, 62%; 349), fgself-27's 6.67M 66% (117%, 40%; 305). Ties 9% against 8%. No clear gain over ×1.0; by 20-minute windows 93%, 97%, 76%. Back to ×1.0.

fgself-35 (from fgself-33's 6.25M, 18:58, with its Adam state): twice the learning rate (--lr 6e-5, was 3e-5, AlphaStar's; queued since 2026-10-11 and never tested). UPGO ×1.0, the anchor fgself-32's 6.41M. Evaluation games (2/3 against insane): the current policy against fgself-33's 6.25M (its start) and fgself-27's 6.67M.

fgself-35 (--lr 6e-5, 18:58-19:58): the current policy 103% of the AI's record (normal 170%, insane 76%; 263 games), fgself-33's 6.25M (its start) 93% (142%, 70%; 353), fgself-27's 6.67M 73%. But it fell over the hour: 136%, 97%, 74% by 20-minute windows; in its second half 82% against the start's 101% (insane 55% against 76%). Its training losses stayed calm (KL to the last policy 0.0039 against 0.0029, clip fraction 1.7% against 1.5%). Twice the learning rate: a fast rise, then worse than the start. Back to 3e-5.

fgself-36 (from fgself-33's 6.25M, 19:59, with its Adam state): the best line on, the anchor moved again (--init = fgself-33's 6.25M). UPGO ×1.0, lr 3e-5. Evaluation games (2/3 against insane): the current policy against fgself-33's 6.25M (its start and anchor) and fgself-35's last (6.25M).

fgself-36 (the anchor moved again, 19:59-21:02): the current policy 94% of the AI's record (normal 161%, insane 66%; 214 games), fgself-33's 6.25M (its start) 84% (301), fgself-35's last 85% (ties 11%; 279). On one matchup mix (below): 84%, 79%, 83%. No clear gain; by 20-minute windows 98%, 96%, 83%.

The evaluation games repeated in every fork. The games' random draws were seeded by the thread and the update the run started from, and every fork starts from update 0: all forks drew the same evaluation games in the same order (fgself-34, -35 and -36: the current policy 256-266 games a hour, the AI's mean record in its matchups 0.450-0.456, against 0.49-0.51 for the checkpoints; the current policy also played fewer launches, 107 against 141-154). The AI-record share depends on the matchup mix (the policy is far above the AI's record as human, near it as night elf). Re-scored on one matchup mix (each matchup that every policy played, weighted by its games over all of them; real_games.py --split now prints it), the verdicts since fgself-30 hold within a few points:

run	current	its start	the other
fgself-30 (credit per unit)	84%	91%	72%
fgself-31 (UPGO ×0.5)	102%	79%	72%
fgself-32 (UPGO ×1.0)	108%	92%	80%
fgself-33 (anchor moved)	112%	106%	81%
fgself-34 (UPGO ×2.0)	85%	81%	64%
fgself-35 (lr 6e-5)	101%	93%	71%
fgself-36 (anchor moved)	84%	78%	83%

(Bootstrapped 90% intervals are about ±9 points each.) Now each evaluation slot plays a matchup (races, sides, difficulty) with every policy in turn, and the run's name is part of the seed.

fgself-37 (from fgself-36's 6.25M, 21:02, with its Adam state): half the KL to the anchor (--ref-kl 0.1, forward over every head; was 0.2): the anchor is now a strong policy, not a clone. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-36's last (its start) and fgself-33's 6.25M (the anchor).

fgself-37 (--ref-kl 0.1, 21:02-21:58, ~265-270 paired evaluation games each): on one matchup mix the current policy 86% of the AI's record (normal 156%, insane 58%), fgself-36's last (its start) 82% (insane 51%), fgself-33's 6.25M (the anchor) 87% (insane 64%). No gain. The pairing works: the three policies played the same 23+ matchups, 263-270 games each.

How the best policy loses to insane now (fgself-33's 6.25M in the evaluation games of fgself-34 to -37, 867 games: 276 won, 565 lost, 26 tied): as before, a greedy economy. In its losses it builds 22 workers against the AI's 10, has 1.7 heroes against 2.5, is behind at one minute (food 19.4 against 22.0) and loses the fights (19 kills for 63 units lost); its wins: 14 workers against 17, 2.1 heroes against 1.4, 43 kills for 16. Losses are longest as undead and human (167 and 186 of 565). 90% of its games against the AI are curriculum games, in which the AI loses up to ~80% of its income to the tax: there greed pays.

fgself-38 (from fgself-33's 6.25M, 21:59, with its Adam state): half the games against the AI untaxed (--real-share 0.5, was 0.1). It wins ~30% of the real games against insane now, so they carry a signal (at 0.25 on 2026-10-11 it won almost none). --ref-kl back to 0.2. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-33's 6.25M (its start and anchor) and fgself-37's last.

fgself-38 (--real-share 0.5, 21:59-23:00, ~260 paired games each): on one matchup mix the current policy 75% of the AI's record (insane 52%), fgself-33's 6.25M (its start) 81% (59%), fgself-37's last 75% (62%). No gain; back to 0.1. Its training games against the real insane stayed at ~28% won. The evaluation games against insane also correct the diagnosis: the policy trains more army than the AI (~60 units to ~40) and still loses the fights (~25 kills for ~50 units lost), with 1.8 heroes to the AI's 2.3. The fights decide, not the economy.

fgself-39 (from fgself-33's 6.25M, 23:01, with its Adam state): no dropout in the updates (--rl-dropout 0; the network kept the cloning's 0.1). The actors act in eval mode, without dropout, while the updates trained with it: each update's ratios, KL and clipping counted the dropout's noise (3 layers, dropout in attention, residuals and feed-forward), and so did the gradients. --real-share back to 0.1. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-33's 6.25M (its start and anchor) and fgself-38's last.

Its first launch died at once: with the dropout changed in its config, its evaluation checkpoints' network (dropout 0.1) found no inference slot. model.arch_key now ignores dropout (nothing at inference). Relaunched at 23:01. Its first updates: the KL to the last policy 0.0009-0.0010 (with dropout 0.003-0.004) and to the anchor 0.0009 (0.004-0.009): most of what the updates measured as KL was the dropout's noise.

fgself-39 (no dropout in the updates, 23:01-23:56, ~290-300 paired games each): on one matchup mix the current policy 109% of the AI's record (normal 172%, insane 81%), fgself-33's 6.25M (its start) 93% (180%, 55%), fgself-38's last 100% (173%, 67%). Ties 4% against 7%. A gain of ~16 points, ~25 against insane; by 20-minute windows 112%, 94%, 125%. The updates' KL to the last policy stayed ~0.001 and their clip fraction 0.8% (with dropout 1.5-1.8%); the KL to the anchor rose from 0.003 to 0.005 over the hour. New best: fgself-39's last, runs/fgself-39/checkpoints/0000000006018863.pt (with its Adam state).

Paused 2026-10-13 23:56 for a machine shutdown. fgself-39 stopped cleanly (status "stopped"). Next: go on from fgself-39's last with its recipe (--rl-dropout 0, UPGO ×1.0, forward KL ×0.2 over every head, --ai insane, --real-share 0.1); first the anchor moved to it (--init = that checkpoint; it is still fgself-33's 6.25M), then the fights (the policy trains more army than insane and loses the fights ~2:1, with fewer heroes): hero levels in the shaping's material (a hero counts its cost only), focus fire, a memory of enemy units seen.

fgself-40 (from fgself-39's last, 2026-10-14 10:15, after the machine's restart, with its Adam state): the anchor moved to the new best (--init = fgself-39's last; it was fgself-33's 6.25M). The recipe otherwise fgself-39's. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-39's last (its start and anchor) and fgself-33's 6.25M.

fgself-40 (the anchor moved, 10:15-11:18, ~355 paired games each): on one matchup mix the current policy 129% of the AI's record (insane 139 of 246 won), fgself-39's last (its start) 129% (142 of 248), fgself-33's 6.25M 83% (72 of 244). Ties 2%, 2%, 9%. No gain from moving the anchor (the third time it did nothing after the first two moves helped). The dropout-free policy leads fgself-33's 6.25M by ~46 points here (16 in fgself-39's own games, which started from it): against insane it now wins more than half its games.

fgself-41 (from fgself-40's 6.58M, 11:19, with its Adam state): hero levels in the shaping (--hero-level-value 150): a hero's material also counts 150 for each level above the first (its cost, 525, did not grow as it levels; a dead hero counts nothing until revived). The anchor stays fgself-39's last. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-40's last (its start) and fgself-39's last (the anchor).

fgself-41 (hero levels in the shaping, 11:19-12:19, ~390 paired games each): on one matchup mix the current policy 133% of the AI's record (insane 144 of 254 won), fgself-40's last (its start) 130% (141 of 254), fgself-39's last 124% (127 of 253). Ties 1-2%. No clear gain; off again.

fgself-42 (from fgself-41's 6.41M, 12:19, with its Adam state): twice the learning rate again (--lr 6e-5), now without dropout in the updates. fgself-35 tried it with dropout on, when most of what an update moved was the dropout's noise; without it the KL per update is ~0.001 (it was 0.003-0.004), so the policy moves slowly. The anchor stays fgself-39's last. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-41's last (its start) and fgself-39's last (the anchor).

fgself-42 (--lr 6e-5 without dropout, 12:19-13:19, ~370 paired games each): on one matchup mix the current policy 120% of the AI's record (insane 120 of 247 won), fgself-41's last (its start) 132% (140 of 255), fgself-39's last 127%. Worse, falling over the hour (125%, 120%, 115%): twice the learning rate hurt with dropout and without. 3e-5 stays.

How the best policy loses to insane now (fgself-41's last, 511 evaluation games against insane: 285 won, 222 lost). The weakest matchups are undead against night elf and undead (52-53% of the AI's record) and the human mirror (85%). In its losses it trains masses of tier-1 units (as undead ~47 ghouls a game, 31 in its wins; as human ~40 footmen, 31 in its wins) against the AI's ~25-35 units in all, researches less, has fewer heroes (1.8 against 2.4), reaches tier 2 less often (91% against 100%) and loses those fights 17 kills to 65 units lost. It holds less than the AI (~1,150-1,400 gold and lumber on average against ~1,900): the shaping's penalty on unspent resources (--float-penalty 0.5, from when it hoarded ~3,700) now pushes it to spend at once.

fgself-43 (from fgself-41's 6.41M, 13:19, with its Adam state): a smaller penalty on unspent resources (--float-penalty 0.2, was 0.5), so saving for tech and heroes pays more. lr 3e-5; the anchor stays fgself-39's last. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-41's last (its start) and fgself-39's last (the anchor).

fgself-43 (--float-penalty 0.2, 13:19-14:19, ~384 paired games each): on one matchup mix the current policy 135% of the AI's record (insane 137 of 256 won), fgself-41's last (its start) 142% (150 of 253), fgself-39's last 136%. No gain. Against insane after its first half hour it held a little more (~1,420 against ~1,200) but researched no more and had no more heroes. Back to 0.5. Small knobs of the reward and the learning rate have not moved the line since fgself-39 (~130-140% on one mix).

fgself-44 (from fgself-41's 6.41M, 14:19, with its Adam state): twice the steps per update (--batch-steps 16384, was 8192; snapshots and checkpoints every 10 updates to keep their spacing in steps): each gradient from twice the data, half as many updates. Twice the learning rate hurt twice, so the updates look limited by their gradients' noise, not their size. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-41's last (its start) and fgself-39's last (the anchor).

fgself-44 (--batch-steps 16384, 14:19-15:19, ~365 paired games each): on one matchup mix the current policy 135% of the AI's record (insane 132 of 246 won), fgself-41's last (its start) 143% (138 of 249), fgself-39's last 131%. No gain (by 20-minute windows 118%, 141%, 146% against the start's 136%, 145%, 146%). Back to 8192. fgself-41's last has read 132-148% in four runs and every run from it lower: the recipe's knobs (shaping, learning rate, batch) do not move the line.

fgself-45 (from a grown fgself-41's 6.41M, 15:19): a deeper network, 5 transformer layers (was 3; 3.02M weights, was 2.13M). scripts/grow_model.py adds the layers after the last, each the identity at first (its attention's output projection and its feed-forward's second layer zero: a pre-norm layer then adds nothing), so the network plays exactly as fgself-41's last (the encoder's outputs equal; its first update's KL to the anchor 0.0050, as fgself-41's); Adam's state carried by parameter name for the old weights. A run can now start from another architecture than its anchor's (the anchor stays fgself-39's last, 3 layers), and the inference warms the evaluation checkpoints' architectures. Its updates take 4.3 s (3.7 s), ~1,680 steps/s (~1,880). Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-41's last (its start, 3 layers) and fgself-39's last (the anchor).

fgself-45 after its first hour (~340 paired games each): on one matchup mix the current policy 131% of the AI's record (insane 119 of 224 won), fgself-41's last (its start) 135% (123 of 224), fgself-39's last 131%. Level, as a network that started as its start and grows its new layers from zero may be for a while: it goes on for another hour or two before a verdict. ~1,575 steps/s.

fgself-45 stopped at 1.5 hours (15:19-16:49, ~500 paired games each): on one matchup mix the current policy 129% of the AI's record, fgself-41's last (its start) 138%, fgself-39's last 129%; its second hour alone 126% against 146%. No gain from the deeper network on this time scale (it may need far longer: the grown checkpoint stays at runs/fullgame/grown/fgself-41-6.41M-5layers.pt).

A pattern: fgself-41's last has read ~140% in five runs (~2,000 paired games: 142%, 143%, 142%, 143%, 138%), and every run trained from it settled near ~130% (fgself-42 to -45), the level of their anchor (fgself-39's last, 127-136% in the same games). The forward KL to the anchor (×0.2 over every head) may pull the policy back to it.

fgself-46 (from fgself-41's 6.41M, 16:49, with its Adam state): the anchor moved to fgself-41's last (--init = it), 3 layers. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-41's last (its start and anchor) and fgself-39's last (the old anchor).

fgself-46 (the anchor at its start, 16:49-17:51, ~377 paired games each): on one matchup mix the current policy 139% of the AI's record (insane 146 of 243 won), fgself-41's last (its start and anchor) 142% (148 of 242), fgself-39's last 133%. Level: with the anchor at its start the line held (the runs from fgself-41's last with the anchor at fgself-39's last fell 7-12 points below it). No gain either. (real_games.py failed on a game over in its first second, against an AI that never started: 2 of ~56,000 games since fgself-39; such a game has no food at one minute now.)

fgself-47 (from fgself-46's 6.58M, 17:52, with its Adam state; fgself-46's recipe goes on): an evaluation-only experiment, sampling at temperature 0.8 for the orders and the targets (AlphaStar's) against 1. The network has an order temperature now (a buffer, as the targets'; --order-temperature), and scripts/checkpoint_variant.py writes a checkpoint's copy at other temperatures. The evaluation games play fgself-41's last as it is and its copy at 0.8, head to head on the same matchups, beside the current policy.

fgself-47 (17:52-18:20, ~139 paired games each): fgself-41's last at temperature 0.8 (orders and targets) 71% of the AI's record (ties 10%), the same weights at 1 136% (ties 3%), the current policy 134%. Decisive at once: a lower order temperature concentrates each unit's choice on "no new order" (its likeliest class most steps), and the policy goes passive (cloning saw the same: a temperature below 1 issued fewer orders). Stopped early for the next experiment.

fgself-48 (from fgself-47's 2.80M, 18:20, with its Adam state): no advice from the built-in AI (--opd-share 0, was 0.5; --opd-coef 0, was 0.15). In half the games against itself, past snapshots and the curriculum the insane AI shadowed the policy's side (its orders labels, undone; some stood until the step ended) and a cloning loss pulled the policy toward them. The policy now wins ~60% of its games against insane: the teacher is weaker than the student, and its orders briefly play in those games. The cloning loss on the AI's winning demonstrations (--bc-coef 0.02) and the forward KL to the anchor stay. Its first updates' KL to the last policy: 0.0002-0.0003 (0.001 with the advice): the distillation was much of what an update moved. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-47's last (its start) and fgself-41's last (the anchor).

fgself-48 (no advice, 18:20-19:18, ~400 paired games each): on one matchup mix the current policy 152% of the AI's record (insane 172 of 273 won, 63%), fgself-47's last (its start) 137% (140 of 272), fgself-41's last (the anchor) 145% (161 of 270). Ties 0-1%. A gain, by 20-minute windows 135%, 166%, 158%, and ~2,030 steps/s against ~1,870 (no shadow AI in the games, no labels or distillation loss in the updates). The student beats its teacher now. New best: fgself-48's last, runs/fgself-48/checkpoints/0000000006907732.pt.

fgself-49 (from fgself-48's 6.91M, 19:18, with its Adam state): no cloning loss on the AI's winning demonstrations either (--bc-coef 0, was 0.02). On 2026-10-10 (fgself-12) that let the hero orders decay; now the forward KL to the anchor over every head keeps them. Its updates are faster without the cloning batches (its second update at ~2,690 steps/s). The anchor stays fgself-41's last. Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-48's last (its start) and fgself-41's last (the anchor).

fgself-49 (--bc-coef 0, 19:18-20:18, ~410 paired games each): on one matchup mix the current policy 147% of the AI's record (insane 192 of 280 won), fgself-48's last (its start) 152% (200 of 277), fgself-41's last (the anchor) 129%. Level (by 20-minute windows 131%, 150%, 162% against the start's 136%, 164%, 160%), the hero orders held (1.56 heroes a game against the start's 1.45), and ~2,220 steps/s against ~2,030: the updates take 2.4 s (3.4 s), and the learner now waits for the games (1.3 s an update). Kept.

fgself-50 (from fgself-49's 7.73M, 20:18, with its Adam state): the anchor moved to the best (--init = fgself-48's last, ~150%; it was fgself-41's last, 129% here): anchored below the line, runs fell back toward their anchor (fgself-42 to -45); anchored at their start, the line held (fgself-46). Evaluation games paired on matchups (2/3 against insane): the current policy against fgself-49's last (its start) and fgself-48's last (the anchor).

fgself-50 (the anchor moved to the best, 20:18-21:18, ~418 paired games each): on one matchup mix the current policy 179% of the AI's record, fgself-49's last (its start) 179%, fgself-48's last (the anchor) 176%. Level. (Higher than the runs before for the same policies: this run's matchups differ.) Against insane the three won ~77% of their games, against normal ~99% (157 of 158).

fgself-51 (from fgself-50's 7.73M grown to 5 layers, 21:20, with its Adam state for the old weights): the deeper network again, now that the updates take 2.3 s and the learner waits 1.5 s an update for the games: the two layers cost little throughput. fgself-45 tried it under the old recipe (no gain in 1.5 hours); this one runs ~2 hours. Evaluation games now against insane only (--eval-ai insane). The anchor stays fgself-48's last (3 layers). Evaluation games paired on matchups: the current policy against fgself-50's last (its start, 3 layers) and fgself-48's last (the anchor).

fgself-51 (the deeper network, 21:20-23:12, ~793 paired games each, all against insane): on one matchup mix the current policy 156% of the AI's record (80% of its games won), fgself-50's last (its start) 148% (76%), fgself-48's last (the anchor) 140% (72%). A modest gain; ahead in every 20-minute window but the last. By matchup (the current policy): orc 97%, night elf 98%, human 65%, undead 60% of its games won; the mirrors human 93%, orc 98%, night elf 100%, undead 62%; the hardest, undead against orc, 47% (the AI itself 23%).

The duelfast milestone. The policy beats the built-in insane AI in every matchup, untaxed, above the AI's own record. The user's next step (2026-10-15): the real game on Echo Isles (see the road).

fgself-52 (from fgself-51's last, 2026-10-15 10:19, after the machine's restart, with its Adam state): the 5-layer line on, the anchor moved to its last (--init = fgself-51's last). Evaluation games paired on matchups (insane): the current policy against fgself-51's last (its start and anchor) and fgself-50's last.

Toward Echo Isles (2026-10-15): positions and point targets in each map's own frame (features.set_geometry; self-play on another map sets it); two test games of the built-in AI against itself on Echo Isles (both at full strength, insane, 8 game minutes in 16-19 s); then runs/fullgame/demos-echo-1: 1,500 games of the built-in AI against itself (normal and insane, all races, --handicap 100, up to 40 minutes), ~600 an hour, for a clone of the real rules as the anchor.

fgself-52 (the 5-layer line, the anchor at its start, 10:19-11:13, ~270 paired games each against insane): on one matchup mix the current policy 172% of the AI's record (81% of its games won), fgself-51's last (its start) 172% (81%), fgself-50's last (3 layers) 154% (73%). Level. (~1,700 steps/s: the demonstrations' collection shares the CPU.)

fgself-53: the duelfast policy on Echo Isles, zero-shot (from fgself-52's last, 11:24; --map "(2)EchoIsles" --pairs 1 --handicap 100 --max-minutes 35, against insane with its curriculum). Its first launch failed at once: self-play's games defaulted to five pairs of players a map load (--pairs 5), which only the duel maps have; --pairs 1 for Echo Isles. Then 0 of 58 games won in its first minutes, all lost in 7-10 game minutes, the curriculum's taxed ones too. Against the AI: heroes 0.1 a game against 1.9, army units 5 against 14, most food 22 against 60, tier 2 in none of 58, 1 kill for 35 units lost; it trained workers (~32 a game, many replacing the killed) and held ~1,900 gold and lumber. Its duelfast habits (costs halved, production three times faster there) fail under the game's rules. Stopped after a few minutes: the GPU went to a clone of the real rules.

fullgame-echo-1 (behaviour cloning, 11:33): the duelfast 5-layer policy (fgself-52's last) fine-tuned on the built-in AI's winning sides on Echo Isles (demos-echo-1, 1,019 games at its start; 6 epochs, lr 1e-4). Its loader failed on a tie's game at first (no winning side with --winners-only: fixed). The clone becomes the next runs' start and anchor.

A curriculum of rules instead (the user, 2026-10-15: "start with halved costs and faster production, and anneal to the real rules later ... as long as we stay on echo isles"). The clone was stopped (its first launch had also read a game the collection was still writing: fixed). Echo Isles under other rules: --map echofast (duelfast's rules in the real map: its tables in the map, data/duelmap.py) and echoP (P% of the way to the game's rules; echo100 is the real game).

fgself-54 (echofast, --handicap 50, from fgself-52's last, 12:29): still 0 of 66 against insane, behind from the first minute (food 10.6 at one minute against 24). Its build orders were refused: in a test game 5 of 67 barracks orders taken (on duelfast 2 of 3). The map frame scaled to Echo Isles' size put every learned offset 2.3 times as far from the base, in the trees.

The frame, anchored. Each player now sees positions relative to its own start location at the duel maps' scale, its base where duelfast's left base was (on the duel maps exactly the frame of before). Point bins reach over the whole map (the network's bins; scripts/grow_bins.py widens a checkpoint's: the old bins in place, the new outer ones the nearest edge's row 10 lower; a constant low bias first let the outer bins win where every old logit was low, builds 10,000 units north and south).

fgself-55 (12:42, anchored frame, 416 bins): it played a real game (tier 2 in 86 of 89, 95 army units to the AI's 49, 2.0 heroes to 2.7) but won none of 89 against insane and tied 41: its attacks went to the enemy base's old place, 4,600 units out, mid-map on Echo Isles (9,856).

The frame, squeezed. Between the bases x is squeezed to the duel maps' 4,600 (as is within 1,500 of either base; ~200 units a bin mid-map): the enemy base is where duelfast's was. 352 bins.

fgself-56 (12:52): in its first minutes 33 of 73 games against insane won (15 tied), 4 of 19 against the untaxed AI; 100 army units to the AI's 38, 32 kills for 56 lost. Evaluation games against insane on echofast: the current policy against its start (the widened fgself-52's last).

fgself-56 (echofast, 12:52-14:14, 341 paired evaluation games each against the untaxed insane AI): the current policy won 44% (ties 21%), its start 31% (ties 28%); on one matchup mix 91% of the AI's record against 64%, rising (by 20-minute windows 58%, 95%, 84%, 118%, 152%). A gain. Its wins are quick (~6 minutes, 41 kills for 24 lost); its losses long, with huge armies of cheap units that lose the fights (132 trained against the AI's 46, 17 kills for 87 lost); its opening slow (food 14 at one minute against 23). By race: orc 34 of 52 won, night elf 24 of 42, human 14 of 60, undead 10 of 60 (at 13:43).

fgself-57 (from fgself-56's last, 14:15, with its Adam state): echofast on, the anchor moved to fgself-56's last, games up to 30 minutes (~25% tied at 20). Evaluation games against insane on echofast: the current policy against fgself-56's last (its start and anchor) and the widened fgself-52's last.

fgself-57 (14:15-15:14, ~165 paired games each against the untaxed insane AI on echofast): the current policy won 52% (ties 11%), fgself-56's last (its start) 57% (8%), the widened fgself-52's last 42%. No gain. With 30-minute games ties fell to ~8% (from ~25% at 20). On echofast the line holds ~55-60% of its games against insane.

fgself-58 (from fgself-56's last, 15:15): the first step toward the real rules, --map echo25 (hit points 0.625, build, train and research times 0.5, costs 0.625 of the game's) and --handicap 62 (was 50); games up to 30 minutes; the anchor its start. Evaluation games against insane on echo25: the current policy against fgself-56's last (what the step costs) and fgself-57's last. The schedule: echo25 / 62, echo50 / 75, echo75 / 88, echo100 / 100 (the real game), each once the policy is past its start there.

fgself-58 (echo25, 15:15-16:16, ~147 paired games each against the untaxed insane AI): the current policy won 37% (ties 9%), fgself-56's last (its start) 43% (10%), fgself-57's last 46%. The step cost ~15 points (fgself-56's last won ~57% on echofast), and an hour did not win them back.

fgself-59 (from fgself-56's last, 16:16): echo25 again with a quarter of the KL to the anchor (--ref-kl 0.05, was 0.2): the anchor learned duelfast's rules, and its pull keeps their habits (on duelfast the anchor's rules always matched). Evaluation games against insane on echo25: the current policy against fgself-56's last (its start and anchor) and fgself-58's last.

fgself-59 (echo25, --ref-kl 0.05, 16:16-17:15, ~150 paired games each): the current policy won 52%, fgself-56's last (its start) 52%, fgself-58's last 53%. Level: a weaker anchor did not let it adapt in an hour either.

fullgame-echo-1 again (17:20): after echofast an hour of RL did not win back a step of the rules, so a clone of the real rules first (the project's recipe: clone, then RL anchored to it). The best echofast policy (fgself-56's last, with the anchored and squeezed frame and 352 bins) fine-tuned on the built-in AI's winning sides of the 1,500 Echo Isles games under the game's own rules, 2 epochs at lr 1e-4 (~15 batches a second, ~15 minutes an epoch; bc.py now prints its progress). Then RL under the real rules (echo100, --handicap 100) from the clone, anchored to it.

fullgame-echo-1 (17:20-17:52, 2 epochs of ~13,300 batches, ~16 minutes each): validation loss 3.51 after the first, 3.29 after the second; the AI's orders (given one) 82%, its target units 68%, its points within 7.2 bins, the value's explained variance 0.99.

fgself-60: the real game on Echo Isles (18:13): the game's own rules (--map echo100), both sides at full strength (--handicap 100), games up to 40 minutes, against the insane AI with its curriculum; from the clone, anchored to it (forward KL ×0.2 over every head). Its first 54 games against insane: 3 won (2 of 12 against the untaxed AI), 10 tied, ~25 game minutes each (zero-shot the duelfast policy had won 0 of 58). Its economy now matches the AI's (food 8.5 at one minute against 7.9, 16 workers against 16, tier 2 in 48 of 54); its fights do not (1.5 heroes against 2.3, 17 kills for 50 units lost). Evaluation games against insane under the real rules: the current policy against the clone (its start and anchor) and fgself-56's last (the echofast policy as it is).

fgself-60's evaluation games after half an hour (~43 each against the untaxed insane AI under the real rules) turned it around: the current policy 0 of 42 won, the clone 0 of 40, fgself-56's last (the echofast policy, as it is) 22 of 45 (49%). The frame (anchored at the own start, squeezed between the bases) was what Echo Isles needed; the rules matter little (at echo25 it had won 43-52%, here 49%); and cloning the AI's games undid the policy's fighting (its training games' 3 of 54 were mostly against the taxed AI). Stopped.

fgself-61 (from fgself-56's last, 18:46): the real game on Echo Isles (echo100, --handicap 100, games up to 40 minutes) from the echofast policy, anchored to itself. Evaluation games against insane under the real rules: the current policy against fgself-56's last.

fgself-61 (the real game, 18:46-20:46, ~306 paired games each against the untaxed insane AI): the current policy won 47% (ties 10%), fgself-56's last (its start) 46% (12%). Level: in the real game on Echo Isles the policy is about even with insane. In both policies' evaluation games: the wins quick (11.6 game minutes, 35 kills for 17 lost, it attacks early); the losses long (25.8 minutes: 35 workers against the AI's 15 and 90 army units against 42, 21 kills for 85 lost, 1.2 heroes against 2.7, tier 2 in 78% against 100%; it holds ~4,000 gold and lumber, the AI ~7,600); ties at 40 minutes 11%. By race: orc 123 of 168, night elf 97 of 172, human 42 of 128, undead 25 of 145 (against orc 0 of 25). As undead it trains ~32 ghouls a lost game, as human ~17 footmen.

fgself-62 (from fgself-61's last, 20:46): a longer horizon for the real game (--gamma 0.9997, was 0.999: ~28 minutes of game time ahead instead of ~8; the duelfast games lasted ~6, the real game's ~25); it misses what pays off late. Evaluation games against insane under the real rules: the current policy against fgself-61's last.

fgself-62 (--gamma 0.9997, 20:46-22:46, ~325 paired games each against the untaxed insane AI): the current policy won 59% (ties 7%), fgself-61's last (its start) 55% (4%); on one matchup mix 112% of the AI's record against 104%, ahead in 5 of 6 twenty-minute windows. A small gain; kept. Its losses: 0.87 heroes a game against the AI's 2.69, ~47 workers. Heroes by race (the current policy against the AI): human 0.57 against 2.22 (its altar in 48 of 72 games, the AI's in all), undead 1.20 against 2.15, orc 0.96 against 1.59, night elf 1.35 against 1.72; games won: human 18 of 72, undead 17 of 60, orc 78 of 94, night elf 78 of 98.

fgself-63 (from fgself-62's last, 22:47): the cloning loss back, on the built-in AI's games in the real game (--bc-coef 0.02 --bc-data runs/fullgame/demos-echo-1 --bc-winners 1), for the heroes: in 2026-10-10's runs it was the one term trained at the AI's own states and kept its hero orders; here the anchor is the policy itself. The full clone undid the fighting; a ×0.02 term should not. Evaluation games against insane under the real rules: the current policy against fgself-62's last.

fgself-63 (the cloning loss, 22:47-23:46, ~275 paired games each against the untaxed insane AI): the current policy won 65%, fgself-62's last (its start) 65%; heroes a game 0.6-0.7 against the start's 0.9-1.0 (the AI's ~1.7), and ~1,220 steps a second against ~1,600 (the cloning batches of Echo Isles' games, many units each, made the updates 6.2 s). No gain; off again.

fgself-64 (from fgself-62's last, 23:47): hero levels in the shaping again (--hero-level-value 150), now in the real game, where the policy fields 0.6-1.0 heroes a game to the AI's ~1.7 and 25-minute games give levelled heroes their weight (on duelfast's 6-minute games, with cheap heroes, it was neutral). Evaluation games against insane under the real rules: the current policy against fgself-62's last and fgself-63's last.

fgself-64 (hero levels in the shaping, 23:47-00:45, ~133 paired games each against the untaxed insane AI): the current policy won 69% (0.92 heroes a game), fgself-62's last (its start) 67% (0.81), fgself-63's last 73% (0.72). Level; the heroes a little up. Stopped for a machine shutdown.

fgself-65 (from fgself-64's last, 2026-10-16 09:07, after the restart, with its Adam state): the real game on, a longer run of the recipe as it stands (--gamma 0.9997, hero levels in the shaping, no cloning loss, the anchor its start). Evaluation games against insane under the real rules: the current policy against fgself-64's last and fgself-62's last.

fgself-65 (the longer run, 09:07-11:14, ~255 paired games each against the untaxed insane AI): the current policy won 67% (0.72 heroes a game), fgself-64's last (its start) 61% (0.66), fgself-62's last 74% (0.84). It rose above its start but not to fgself-62's last, the best in every run since: the hero shaping (fgself-64, fgself-65) likely cost a little. Off again. The curriculum then taxed the AI only for the learner's weak races (undead 45% of its income, human 29%, orc and night elf none).

fgself-66 (from fgself-62's last, 11:15): more self-play (--ai-share 0.5, was 0.75: half the games against itself and past snapshots). The policy is above insane now. fgself-62's recipe otherwise (γ 0.9997, no hero shaping, the anchor its start). Evaluation games against insane under the real rules: the current policy against fgself-62's last and fgself-65's last.

fgself-66 (more self-play, 11:15-12:16, ~120 paired games each against the untaxed insane AI): the current policy won 58%, fgself-62's last (its start) 74%, fgself-65's last 68%. Worse: with half its games against itself it trained less against the AI it is measured on. Back to 0.75.

fgself-67 (from fgself-62's last, 12:16): the weak races trained more (--learner-races human:2,undead:2,orc:1,nightelf:1: the learner plays human or undead in two thirds of its training games, 63% in its first 139; evaluation games stay uniform). In the real game orc and night elf win ~80% of their games against insane, human and undead ~25%. Evaluation games against insane under the real rules: the current policy against fgself-62's last.

fgself-67 (the weak races trained more, 12:16-13:44, ~290 paired games each against the untaxed insane AI): the current policy won 64%, fgself-62's last (its start) 65%. Level, the weak races too: human 24 of 77 against 26 of 75, undead 32 of 68 against 34 of 71 (orc 62 of 67, night elf 66 of 74). More games as human and undead did not teach them more in 1.5 hours. Next: whether the game refuses their builds on Echo Isles' terrain (undead build on blight; as human the altar came in two games of three).

The builds, refused. One 8-minute game a race against insane under the real rules (fgself-62's last, build orders sent and taken): human 15 of 22, orc 13 of 21, undead 25 of 100 (its ziggurats, its supply, 8 of 67: off the blight), night elf 11 of 54 (its altar 1 of 25). The spots the policy learned on the flat duel map often miss Echo Isles' blight, trees and cliffs, and the harness tried other spots only up to 256 units away. Rings up to 768 units (16 spots a ring beyond 256): undead 26 of 33, night elf 14 of 19.

fgself-68 (from fgself-62's last, 14:09): the wider build search in every game (the evaluation games of the start too: the head to head measures what training adds; the fix itself shows in fgself-62's last against its ~65-74% in the runs before). Races uniform again. Its first 63 games against insane: 52 won.

The yardstick on Echo Isles. The AI-record share weighted each matchup by the built-in AI's win rate on duelfast. On Echo Isles under the game's rules (its 1,500 games, ~620 decided between equals) the races stand otherwise: in mixed insane games the AI as human wins 14%, as orc 43%, as night elf 58%, as undead 88%. So against the AI's own record the policy's human (~37%) is strong and its undead (~45%) the weak one, not as the duelfast table made it. real_games.py now takes the AI's games of the run's own map (demos-echo-1 for Echo Isles under the game's rules).

fgself-68 (14:09-15:45, ~283 paired games each against the untaxed insane AI): the current policy won 62% (122% of the AI's own record on Echo Isles), fgself-62's last (its start, with the wider build search too) 57% (113%). A small gain, ahead in 3 of 5 twenty-minute windows; the new best, runs/fgself-68/checkpoints/0000000008877917.pt. The start checkpoint did not gain from the wider search alone (57-59% here against its ~65-74% in other runs' matchups): the policy had to learn to use it (its undead 29 of 65 against the start's 24 of 66, its night elf 52 of 52 against 41 of 52 at 15:13).

fgself-69 (from fgself-68's last, 15:46): fewer games against itself (--ai-share 0.95, was 0.75): more self-play cost 16 points (fgself-66). Evaluation games against insane under the real rules: the current policy against fgself-68's last and fgself-62's last.

fgself-69 (15:46-17:46, ~294 paired games each against the untaxed insane AI): the current policy won 71% (146% of the AI's own record on Echo Isles), fgself-68's last (its start) 69% (143%), fgself-62's last 61% (125%). Level with its start, ties a little fewer (3% against 5%); kept. The real game stands at ~70% of its games against insane. By race (the current policy): orc 55 of 58, night elf 48 of 50, human 31 of 66, undead 20 of 45. Against the AI's own record undead is the weak race: as undead the policy floods cheap units (37 acolytes and 74 ghouls a game against the AI's 8 and 21), reaches tier 2 in 53% of its games (the AI in all), has 1.05 heroes against 1.82 and hardly any crypt fiends, gargoyles or abominations; 23 kills for 72 units lost.

fgself-70 (from fgself-69's last, 17:48): a cloning loss on the AI's undead alone (--bc-coef 0.02 --bc-data runs/fullgame/demos-echo-1-undead-wins --bc-winners 1 --bc-batch 32: the 439 Echo Isles games it won as undead). The loss on all its games (fgself-63) was neutral; this one only shows the undead's way. Evaluation games against insane under the real rules: the current policy against fgself-69's last.

fgself-70 (17:48-19:17, ~406 paired games each against the untaxed insane AI): the current policy won 63%, fgself-69's last (its start) 62%. Level. Its undead changed its way (tier 2 in 64% of its games against 37%, fewer acolytes) but won no more (31 of 103 against 35 of 103), at 16% fewer steps a second. Off.

Why undead loses: lumber. In the evaluation games of fgself-69 and fgself-70 (per game minute, the policy against the AI as the same race): undead gold 813 against 873 but lumber 83 against 213; human 319 against 395, orc 295 against 301, night elf 402 against 394. Undead's lumber workers are its ghouls, its first fighters too: the policy fights with them and hardly sends them to the trees, so no lumber, no tier 2, and its ghoul flood.

fgself-71 (from fgself-69's last, 19:17): no penalty on unspent resources (--float-penalty 0, was 0.5, from when the duelfast policy hoarded): in the real game the AI holds 3,000-6,500 gold and lumber at a time (saving for tech), the policy 700-1,800, and a gathered resource lowered the potential until spent. The cloning loss off. Evaluation games against insane under the real rules: the current policy against fgself-69's last.

fgself-71 (19:17-20:15, ~217 paired games each against the untaxed insane AI): the current policy won 64% (117% of the AI's own record), fgself-69's last (its start) 69% (128%). Worse: it held more (~1,540 against ~1,210) and its undead gathered a little more lumber (86 against 62 a game minute, 15 of 52 won against 12 of 52), but its human fell (11 of 43 against 22 of 45). Back to 0.5.

A harvest order to a ghoul works and the features mark it a lumber worker: the policy just stopped giving it (the clone had it from the AI; RL on duelfast's cheap rules wore it away).

fgself-72 (from fgself-69's last, 20:15): a stronger cloning loss on the AI's undead alone (--bc-coef 0.1, five times fgself-70's, the same 439 games; --bc-batch 32), for undead's lumber. Evaluation games against insane under the real rules: the current policy against fgself-69's last.

fgself-72 (20:15-21:15, ~290 paired games each against the untaxed insane AI): the current policy won 59%, fgself-69's last (its start) 68%. Worse. Its undead gathered more lumber (92 against 72 a game minute) but won fewer games (12 of 61 against 23 of 65), and its human (18 of 70 against 25 of 69) and night elf (66 of 77 against 76 of 78) fell too: at ×0.1 the cloning reached the other races through the shared network. Off; the undead cloning route is closed.

fgself-73 (from fgself-69's last, 21:16): half the shaping (--shaping 0.5, was 1.0), so the win itself weighs more: with γ 0.9997 the outcome reaches ~28 minutes back, and the shaping may steer toward trading material rather than winning. Evaluation games against insane under the real rules: the current policy against fgself-69's last.

fgself-73 (half the shaping, 21:16-22:46, ~345 paired games each against the untaxed insane AI): the current policy won 72% (139% of the AI's own record on Echo Isles), fgself-69's last (its start) 62% (121%), ahead in 4 of 5 twenty-minute windows. A gain, the weak races most: human 44 of 83 against 28 of 82, undead 48 of 98 against 31 of 96 (orc 84 of 88, night elf 73 of 78). With less weight on the material lead the policy plays for the win. New best: runs/fgself-73/checkpoints/0000000008386478.pt.

fgself-74 (from fgself-73's last, 22:46): less shaping again (--shaping 0.25). Evaluation games against insane under the real rules: the current policy against fgself-73's last and fgself-69's last.

fgself-74 after an hour (22:46-23:47, ~138 paired games each; not judged): the current policy won 74%, fgself-73's last (its start) 76%, fgself-69's last 68%: level so far. Paused 2026-10-16 23:47 for a machine shutdown (stopped cleanly, status "stopped"). The best: fgself-73's last, runs/fgself-73/checkpoints/0000000008386478.pt (~72-76% of its games against the untaxed insane AI in the real game on Echo Isles). Next: fgself-74's question (--shaping 0.25) from fgself-73's last, or less shaping still; undead's lumber (its ghouls).

fgself-75 (from fgself-73's last, 2026-10-17 09:58, after the restart, with its Adam state): fgself-74's question again, a quarter of the shaping (--shaping 0.25). Evaluation games against insane under the real rules: the current policy against fgself-73's last.

fgself-75 after half an hour (09:58-10:31, ~128 paired games each, not judged): the current policy won 73%, fgself-73's last (its start) 78%: level so far. The verdict at ~300 games each. 1,610 agent steps a second (8 actors of 6 games, 8 evaluation games), the learner's update 4.0 s. It started with a league of its own: forks start with their parent's league only since its launch (--league-from: the past snapshots that fit, their records, the curriculum's levels and balance).

fgself-75 (a quarter of the shaping, 09:58-11:25, ~327 paired games each against the untaxed insane AI): the current policy won 75% (143% of the AI's own record on one matchup mix), fgself-73's last (its start) 76% (146%). Level, as fgself-74 was after an hour: human 41 of 90 against 44 of 92, undead 48 of 72 against 48 of 74. The shaping line ends at half (fgself-73). The best stays runs/fgself-73/checkpoints/0000000008386478.pt.

fgself-76 (from fgself-73's last with its Adam state, 2026-10-17 11:27, --shaping 0.5 again; the league from fgself-75): AlphaStar's main exploiter (--exploiter-share 0.2). A second network from the same weights trains only against the main policy and looks for what beats its rush. Its snapshots join the league when they beat the main policy in 70% of their last 50 games, and then it starts over. 20% of the launches are main-exploiter games (each side trains its own network), 76% are against the built-in AI. To run with the native inference loop, the exploiter's versions are numbered files (runs/<run>/exploiter/<version>.pt, exploiter.json names the newest) that a game plays like a past snapshot; its update has its own CUDA graphs. The run also logs the gradient noise scale (grad/noise_scale, McCandlish et al. 2018: the graphed and eager updates agree, and it costs no measurable time). Evaluation games against insane under the real rules: the current policy against fgself-73's last.

fgself-76 (a main exploiter, 11:27-13:23 with three restarts for infrastructure; the first run with the laptop's evaluation games, ~965 paired games each against the untaxed insane AI): the current policy won 75% (152% of the AI's own record on one matchup mix), fgself-73's last 77% (156%); after 12:30, ~540 games each, 74% against 75%. Level. The exploiter never reached 70% against the main policy: it started over twice after 300 updates (scores 0.50 and ~0.49), so no exploiter snapshot joined the league. Two fixes on the way: the exploiter's versions for new games only every 8 of its updates (--exploiter-publish: one every update kept ~10 networks in play and halved the main policy's steps), and the noise scale at fixed weights. The main policy got ~1,250 steps a second, the exploiter ~440.

The gradient noise (fgself-76, 60 measurements at fixed weights, every 10 updates; McCandlish et al. 2018's simple noise scale, in agent steps): the policy term ~31,000, the KL to the anchor ~1,800, the value term ~950; the whole network ~2,350 (in Adam's metric ~20,000). By part: the target heads ~79,000, the transformer layers 14,000-28,000, the order head ~15,000, the value head ~700. A minibatch of 256 steps carries under 1% of the policy's signal, while the KL and the value pull with clean gradients.

fgself-77 (from fgself-73's last with its Adam state, 13:25, the league from fgself-76, no exploiter): a quarter of the KL to the anchor (--ref-kl 0.05, was 0.2). The KL's clean pull may hold each fork at its start (fgself-74/75/76 level); it was needed when the start was a clone. Evaluation games against insane under the real rules: the current policy against fgself-73's last.

fgself-77 (a quarter of the anchor's pull, 13:25-15:24, ~1,280 paired games each against the untaxed insane AI): level for its first ~1.5 hours, then ahead. After 14:50 (~380 games each) the current policy won 82% (166% of the AI's own record on one matchup mix), fgself-73's last 74% (152%): human 57 of 89 against 34 of 88, undead 47 of 75 against 39 of 71. A gain, and the new best: runs/fgself-77/checkpoints/0000000011183066.pt. The KL's clean pull held each fork at its start; with a quarter of it the policy moved (its KL to the anchor ~0.014, about three times the earlier forks') and improved after ~1.5 hours. Learning changes need that long before a verdict.

fgself-78 (from fgself-77's last grown with a strategy block, runs/fullgame/grown/fgself-77-z.pt, 15:25; the anchor moved to it, --ref-kl 0.05; a fresh league: the past snapshots have no strategy block): strategy conditioning, AlphaStar's statistic z (fullgame/strategy.py). Half the learner's training sides get a target strategy drawn for their race and matchup, winners first, from 8,072 (runs/fullgame/strategies-echo.json: the built-in AI's 1,500 Echo Isles games and 2,543 people's 1.29 1v1s; per item of the race's 64 most common, its first time within 15 minutes; people reach tier 2 in 93-100% of their games at ~3-5 minutes). The strategy is an input (a block of the global vector, zero weights at first) and a reward (a potential of minus 0.02 per item made by now by one of the side and the strategy, not both). Evaluation games get none. Evaluation games against insane under the real rules: the current policy against fgself-77's last.

fgself-78 (strategy conditioning, 15:25-17:58, ~1,250 paired games each against the untaxed insane AI on the main machine and the laptop): level for its first hour, then ahead. After 16:55 (~500 games each) the current policy won 89% (179% of the AI's own record on one matchup mix), fgself-77's last 84% (168%): undead 107 of 131 against 84 of 129, human 104 of 122 against 93 of 117, night elf 136 of 137 against 135 of 142, orc 104 of 118 against 105 of 111 (orc fell behind in the run's second hour, 69% against 92% after 16:25, and recovered). A gain, and the new best: runs/fgself-78/checkpoints/0000000014224474.pt. The strategies' mismatch at a game's end fell only a little (13.7 to ~13.3 items of the race's 64), and the training games with a strategy won 74% at first, ~82% later, as many as those without. How much of the gain is the strategies and how much more training at the lower anchor pull, this run does not tell.

fgself-79 (from fgself-78's last with its Adam state, 17:58; the anchor moved to it): more weight on following the strategy (--z-reward 0.05, was 0.02). Throughput: 2 evaluation games on the main machine (was 8; the laptop plays ~900 an hour) and a league of 6 past snapshots (was 12): the inference loop was saturated, half its calls for past snapshots and the evaluation checkpoint. Evaluation games against insane under the real rules: the current policy against fgself-78's last.

fgself-79 (the strategy reward x2.5, 17:58-19:52, ~730 paired games each after 18:55): the current policy won 89% (180% of the AI's own record on one matchup mix), fgself-78's last 87% (177%); after 19:20 87% and 87%. Level. It followed its strategies a little better (the mismatch 13.7 to 12.9 items; the games with one won 74% at first, 83% later), not enough to win more.

fgself-80 (from fgself-79's last, 19:52): the strategies taught by cloning, as AlphaStar's supervised policy learned what z means: an auxiliary cloning loss (--bc-coef 0.05) on the built-in AI's 1,500 Echo Isles games and people's 95 (winners and losers), each side with its own strategy as the input (--bc-z, strategy.hindsight). Worse: after 20:05 (~310 games each) 71% against fgself-78's last's 88% (139% against 173% of the AI's record), undead 53% against 81%, heroes 0.44 a game against 0.80. The cloning loss fell (9.1 to 4.3) but dominated the gradient (its norm 0.8 against ~0.13) and took the policy from its anchor (KL 0.038 and rising, other runs ~0.01): cloning undid the fighting, as the clone of the AI's games and the undead cloning loss did before. Stopped after 30 minutes.

fgself-81 (from fgself-79's last with its Adam state, 20:25; the anchor moved to it; no cloning loss): the material shaping annealed to 0 (--shaping-anneal 600: 0.5 to 0 over ~45 minutes, then none). Potential shaping is the same as a critic of the learned part plus the potential (Wiewiora 2003): the material lead is a prior, and it over-values what lasts (the learner's ~33 workers in long losses against the AI's ~19). The critic, trained for many hours, can take the term in gradually. Evaluation games against insane under the real rules: the current policy against fgself-78's last.

fgself-81 (the shaping annealed to 0, 20:25-21:53): level while the weight fell, then behind; after 21:30 (~266 games each, the weight 0 since ~21:10) the current policy won 84% (160% of the AI's own record), fgself-78's last 89% (168%); human 65% against 81%. Worse by ~5 points, partly recovered: the value's loss rose while the material term went (0.00056 to ~0.0007) and came back. In lost games the learner made fewer workers (23 to 17) and buildings (18 to 16). Its KL to the anchor (fgself-79's last, trained with the shaping) settled at ~0.024, ~40% above fgself-79's ~0.0175.

fgself-82 (from fgself-81's last with its Adam state, 21:55; no shaping): a rolling KL anchor (--ref-update 200: the anchor becomes the current policy every 200 updates, ~15 minutes). The clean KL gradient (its critical batch ~2,100 steps, the policy's ~13,000) may have held fgself-81 between the old anchor's habits and what the rewards without shaping want. Evaluation games against insane under the real rules: the current policy against fgself-78's last.

fgself-82 (no shaping, a rolling anchor every 200 updates, from fgself-81's last, 21:55-22:53): worse: after 22:25 (~265 games each) 75% against fgself-78's last's 88% (154% against 181%). Orc collapsed (17 of 69 against 57 of 66) while human (60 of 64 against 55 of 66) and undead rose: without the shaping and with the anchor following the policy, orc drifted. The KL climbed back to ~0.018 within ~170 updates of each move. The shaping helps: the line returns to --shaping 0.5 with a fixed anchor.

fgself-83 (from fgself-79's last with its Adam state, 22:56; the anchor fixed at it, --shaping 0.5, the strategies as in fgself-79): workers counted only up to 20 in the shaping's material (--worker-cap 20). Paused 2026-10-17 23:42 for a machine shutdown after 45 minutes, not judged: 83% against fgself-78's last's 87% (~272 games each, 163% against 172%); after 23:25 86% against 87% (~80 games each). The laptop's WSL failed at 23:10 (Tailscale answered, SSH hung); this machine's fallback played the evaluation games from then.

fgself-83 (resumed 2026-10-18 10:33 with the same recipe, stopped 11:50): level to slightly worse. From the resume, ~780 paired games each: the current policy won 83% (164% of the AI's own record on one matchup mix), fgself-78's last 87% (170%); by 20-minute windows 177/161/155/164% against 191/154/174/170%. Heroes 0.62 a game against 0.77. In its long losses it still made ~24 workers against the AI's ~15 (fgself-78: ~35): the cap took away some of the pile-up but not the losses. The cap goes.

fgself-84 (from fgself-79's last with its Adam state, 11:51; the anchor fixed at it, --shaping 0.5, no worker cap): target strategies drawn only from people's replays (--z-src human: 5,074 of the pool's 8,072, winners first; the built-in AI's 2,998 out). People's plans differ from the AI's: fewer items (17.6 against 21.8 a strategy) and another tech path, so they may pull the policy out of the AI-like rush. Evaluation games against insane under the real rules: the current policy against fgself-78's last.

fgself-84 (people's strategies only, 11:51-13:20, ~1,000 paired games each): level. After 12:50 (~340 games each) the current policy won 89% (by 20-minute windows 176% and 168% of the AI's own record), fgself-78's last 88% (170%, 185%); over the whole run 87% against 88%. Heroes 0.66 a game against 0.82: people's plans did not bring more heroes or tech into the games without a target.

fgself-85 (from fgself-79's last with its Adam state, 13:21; the anchor fixed at it, --shaping 0.5, the strategies from both sources): the opponent curriculum goes on past zero (--curriculum-floor -1, new). When the learner keeps winning, the tax on the built-in AI's income turns into a bonus, up to 90% more of what it gathers (the level times --curriculum-delay 0.9), until the learner wins about half of those games. The levels sat at 0 (the real game) in most matchups, the learner winning ~87%, and almost all of its losses are long games (27 minutes on average) that it seldom meets in training: a richer AI plays longer games and techs. The test games stay against the AI at full strength. Evaluation games against insane under the real rules: the current policy against fgself-78's last.

fgself-85 (the AI's bonus, 13:21-14:50, ~1,020 paired games each): worse, by a collapse of one race. The levels reached the floor within ~25 minutes in 15 of 16 matchups: the insane AI with 90% more income still lost ~85% of its training games (12-13 minutes on average), so its weakness is how it plays, not what it has. For an hour the current policy led (90% against 88%); then its human collapsed: training games as human against the richer AI 75% won, then 37%, test games 21 of 64 against fgself-78's last's 59 of 66, with 42 workers a game against 16. After 14:20 (~320 games each) 79% against 92%: human 44 of 102 against 85 of 96, while orc (77 of 78 against 72 of 76), undead (58 of 68 against 53 of 65) and night elf (95 of 97 against 92 of 94) gained a little. A one-race collapse after a change to the training signal, as orc in fgself-82.

fgself-86 (from fgself-79's last grown to 8 layers, runs/fullgame/grown/fgself-79-8layer.pt, 14:51; the anchor fixed at fgself-79's last, 5 layers; --shaping 0.5, no AI bonus): a deeper network, 8 transformer layers (was 5; 4.51M weights, was 3.18M), the new ones the identity at first (scripts/grow_model.py), Adam's state carried for the old weights. The line has been level since fgself-78; the network is small for four races. The last growth (3 to 5 layers) gained only after ~2 hours (fgself-45, fgself-51): judged after ~2 hours. Evaluation games against insane under the real rules: the current policy against fgself-78's last.

fgself-86 (8 layers from fgself-79's last, 14:51-15:50): stopped after an hour, not judged: 85% against fgself-78's last's 90% (~600 games each), ~1,450 steps/s (an update 5.0 s, was 3.9). Behind as human (52 of 90 after 15:10 against 75 of 91, 29 workers a game against 20) and night elf (84 of 97 against 94 of 95).

Human, behind in every run since fgself-79. The human test games of each run, its current policy against fgself-78's last (first hour, then after): fgself-79 72% against 80%, then 72% against 77%; fgself-80 68% against 87%; fgself-81 67% against 78%, then 67% against 81%; fgself-83 74% against 86%, then 58% against 81%; fgself-84 60% against 80%, then 65% against 78%; fgself-85 76% against 80%, then 44% against 88%; fgself-86 64% against 84% (fgself-82 alone was ahead, 85% against 80%, while its orc collapsed). The human made more workers each time (22-26 a game against 19). fgself-79 raised the strategy reward to 0.05, and fgself-81 and fgself-83 to -86 all started from its last and kept that reward: their "level" verdicts hid the other races' gains behind a weaker human. The line goes back to fgself-78's last and its recipe.

fgself-87 (from fgself-78's last grown to 8 layers, runs/fullgame/grown/fgself-78-8layer.pt, 15:51; the anchor fgself-78's last; the strategy reward back to 0.02; the league of fgself-78): the deeper network again, from the best policy with its own recipe. Judged after ~2 hours. Evaluation games against insane under the real rules: the current policy against fgself-78's last.

fgself-87 (8 layers, 15:51-16:50): stopped at ~50 minutes, level (86% against 86% after ~260 games each, human 49 of 66 against 53 of 68: the human of fgself-78's recipe is back), ~1,380 steps/s. The user prefers no larger network without a clear gain; the line turns to strategic variety.

fgself-88 (from fgself-78's last, 16:51; the anchor fgself-78's last, its league; the user's direction: strategic variety first, a lower win rate accepted for a while): enforced targets on the fast rules. Three quarters of the learner's training games get a target strategy (--z-share 0.75), and while it is active (the strategies' 15 minutes, scaled) the side may only train, build and research the target's items, plus workers and supply (--z-enforce 1, new: strategy.allowed through the availability mask; moves, attacks, harvesting and spells stay free). The reward for following the target's timing stays (0.02). On echofast (Echo Isles with half the hit points and costs and a third of the build times; games ~10 minutes), the targets' times ×0.45 (--z-time-scale, new). Evaluation games against insane on echofast, three policies on the same matchups: the current policy without a target, the current policy with an enforced target (current+z, --eval-z 1, new: drawn as in training, winners of the matchup first) and fgself-78's last (no target). The question: can the policy win with the builds it is given?

fgself-88 (16:47-17:55; started 16:47, not 16:51): the laptop's worker lacked the strategies file until 17:19 (the manager ran older code; its current+z games failed), so the games with a target are from then. After ~1 hour, on echofast against insane: the current policy without a target 44%, fgself-78's last 42% (~540 games each; both far below their real-rules rates); with an enforced target (current+z) 32% (241 games). It barely followed its targets: ~15 of ~21 items missed at a game's end, flat over the run (14.6 in its first 400 training games with a target, 15.0 in its last 400), and ~2 kinds of unit a game with or without a target; tier 2 in ~40% of the games for all three. The mask keeps items outside the target out, but nothing pays for making the target's items beyond 0.02 each.

fgself-89 (from fgself-88's last with its Adam state, 17:55): five times the reward for following the target (--z-reward 0.1 per item), the rest as in fgself-88.

fgself-89 (17:55-19:30, resumed 18:52 to log z_missed and z_extra): it followed no better: ~15 of ~22 target items missed in its first 300 training games with a target and in its last 300. The detailed log (1,209 training games with a target): 9.4 buildings and units missed a game, 3.4 research, 1.4 heroes, the most missed buildings that need a worker sent to a spot (towers, shops, lumber mill, altar, graveyard); 0.6 items a game outside the target. After the resume the enforced targets cost no wins on echofast: current+z 47%, current 48%, fgself-78's last 43% (~130 games each).

fgself-90 (from fgself-89's last with its Adam state, 19:31): the harness builds what the agent leaves undone of its target (--z-build 1, new: TargetBuilder): each due target building goes to the worker nearest the town hall (a spot 450-800 from it, then the harness's spot search) and each due hero to the altar, one a step, when affordable. The strategy reward back to 0.02. The aim: games with many builds and techs, so the agent learns to fight with them. Evaluation games on echofast: current, current+z (with the builder) and fgself-78's last.

fgself-90 (the harness builds the target's buildings and heroes; started 19:23, not 19:31; stopped 20:58, ~620 games each): with a target and the builder 39%, level with fgself-78's last (37%); the agent alone 60% (after ~4 hours of training on echofast, which fgself-78 never had). With a target it used a few more unit types (2.7 against 2.2 a game) and raised more heroes (1.15 against 0.82), and missed ~11 of ~22 target items itself (~14 without the builder). The gap to its own no-target win rate did not close: by 20-minute windows 71/83/103/74/67/80/71% of the AI's record against 128/122/131/128/106/124/138%. Variety in the games, not yet in the agent.

fgself-91 (from fgself-90's last with its Adam state, 20:59): the builder's orders become labels (--z-build-labels 1, new; --opd-coef 0.05): its accepted orders label the agent's step for on-policy distillation's loss, on the units it ordered only (y_only: the other units are not pushed to "no order"). The agent should learn to start the target's buildings and heroes itself.

fgself-91 (20:59-21:10): stopped, a bug in the loss's weighting. The loss averaged over the labelled units alone (~1 in a few hundred steps), so those rare labels carried its full weight: the first update moved the policy by a KL of 0.15 (normally ~0.001), and the KL to the anchor went from ~0.016 to ~0.25 and stayed there.

fgself-92 (from fgself-90's last with its Adam state, 21:12; the league of fgself-90): fgself-91 with each labelled unit weighted as one of all the step's own units (distill_loss: the mean over every own unit of the labelled steps; on-policy distillation's own steps, which label every unit, are unchanged).

fgself-92 (21:12-22:50, ~655 paired games each on echofast): without a target 56% (fgself-90: 60%), fgself-78's last 40%; with a target and the builder 26% (fgself-90: 39%). Its KL to the anchor settled at ~0.03, its order matched the builder's 41% of the time at the end (18% at the start). Without a target its variety rose and faded again: 2.7 unit types and 21.6 buildings a game at 21:30-21:50, then 2.4 and 18.8 (fgself-78: 2.3-2.5, ~19). With a target the agent built on top of the builder (27-31 buildings a game) and lost more. The labels did not make lasting variety.

fgself-93 (from fgself-92's last with its Adam state, 22:51): the builder only teaches (--z-build 2, new): its orders for due target buildings and heroes are labels for the distillation loss and are never carried out, so the agent has to carry out its target itself. current+z is now the agent alone with its target enforced.

fgself-93 paused 2026-10-18 23:44 for a machine shutdown after 52 minutes, not judged. ~340 paired games each on echofast: without a target 55%, with its target enforced (no builder) 30%, fgself-78's last 43%; with a target it missed ~11 of ~22 items on its own (as fgself-92 with the builder) and built ~31 buildings a game; its order matched the teacher's 54% of the time. Without a target it played much like fgself-78 (2.5 unit types against 2.3, ~18 buildings). Idea for later: trim each target to what a game of its length can afford (many targets come from long games and list more buildings than a 10-minute echofast game supports).

Why the targets could not be followed (2026-10-19, from fgself-90-93's episodes). (a) Impossible items: the agent's orders are the built-in AI's (its vocabulary), so tavern heroes, mercenaries and shop units cannot be ordered, yet 54-93% of the targets (by race) had one: ~1.7 certain misses a game. (b) The old builder made only buildings and heroes: of the ~9.4 possible items missed a game, most were tier-2 halls and what needs them, units and research. (c) The times: the old pool's built-in AI sides included the insane AI, which gathers twice as much (engine notes), and every target came from games under the real rules, scaled by a flat 0.45. (d) A target covered 15 minutes whatever its game's length, so items the agent made after a short game's end counted as off-target.

The fix. A new pool, runs/fullgame/strategies-echofast.json (6,271 strategies): 600 games of the normal built-in AI against itself on echofast (native times; 64% decisive, median 18 minutes) and people's 1.29 1v1s with their times mapped per race onto the AI's (x0.72 human, x0.74 orc, x0.61 undead, x0.76 night elf: the ratio of the medians; no item earlier than the AI's 10% quantile; then in tech order, strategy.consistent); only items the agent can order (--vocab), the old pool's item slots kept (--items-from, the input the policy learned), each strategy covering only its game's length (end). The builder now follows the game's tech tree (fullgame/techtree.py, from the game's *Func.txt): it starts the earliest due item that is neither made nor under way (the game's start, finish and cancel events, the agent's own included), or the first thing it needs (a requirement, or the building that makes it); units, heroes, research and hall tiers at an idle building that makes them, buildings with the nearest free worker. The enforcement mask also lets through what the target's items need (techtree.closure).

fgself-94 (from fgself-93's last with its Adam state, 10:17; the league of fgself-93): the new pool at native times (--z-time-scale 1) and the new builder carrying out its orders (--z-build 1, its accepted orders also labels as in fgself-92). The question: can the targets be followed now? fgself-90's builder and agent together still missed ~11 of ~22 items.

fgself-94's first 50 minutes (restarted 10:23 with a log of the builder's orders per item, given and accepted, and each game's target and made times) showed ~4 of ~12 items due a minute before the game's end missed, and the game refused most of the builder's building orders: night elf 3 of ~400 accepted. Four bugs, found with the log and with single games of the builder alone (builder_probe: no policy, the workers sent to the mine first): (1) the nearest worker to the hall was often one inside a mine (a night elf's five wisps live in the entangled mine; the game refuses their build orders): now only workers outside a mine, and with none free a worker is trained first. (2) The strategies took each event's owner from the unit's id over the whole game, but the game reuses a dead unit's id: 240 built-in AI strategies had the other race's buildings (an undead side ordering human farms 67 times). Now the owner at the event's step, and only the race's own items (the old pool's item lists had the same mix-up; its slots stay, the foreign ones always empty). (3) Research ids are upper case like heroes', so a research under way counted as a hero and the second hero asked for the third tier. (4) Expansions: a second town hall went in the main base and the undead's haunted mine (a target order on the mine) was never given; now both go to the free gold mine nearest the hall. Restarted 10:39 with the fixes and the rebuilt pool. In single games the builder alone then made 11-16 of 13-19 due items per race (the rest due in the last minute, research queued behind busy buildings, or an expansion across creeps). The first ~100 training games after it still missed ~4.6 of ~13.6 items due a minute before the end, most never tried: the builder used only idle buildings, and a town hall training workers, or a barracks training units, never is, so the next tier (and the second hero after it) and most research never came; expansions failed (a haunted mine refused ~400 times; a second hall's worker walks into the creeps at the mine). Restarted 10:45: an idle building first, else the order queued; expansions left to the agent; an item refused three times given up for that game. Its first ~170 games still missed the tier-2 halls most (the keep 31 of 37 times, never ordered), and the reason was older than the builder: the order costs had every upgraded building at its table cost, which holds its base's too (a keep 705 gold and 415 lumber, its upgrade 320 and 210; a guard tower 100/70, its upgrade 70/50). The builder waited for money the upgrade does not need, and so did the availability mask of every policy since it came in: a next tier only once the player held a town hall's and a keep's gold together (on echofast 352/208 instead of 160/106). Restarted 10:53 with the costs fixed (costs.py: an upgrade costs the difference to its base). The policies' own tier-2 timing may change with it.

More fixes, each found in the games' logs (what the builder gave, what the game accepted, the refused orders' context, why a due item was not ordered) and checked in single games: the game takes no upgrade or hero behind others in a building's queue (they now wait for an idle building; a hall training workers back to back has their queue cancelled first, its cost refunded); out of supply a unit waits in its building's queue and blocks a next tier behind it (supply first); a human's or orc's building whose worker was taken away stays an unfinished frame (a free worker resumes it after 10 s without progress). Most refusals left were the builder repeating what the agent ordered in the same step (the item made 18 s later: a hero's training). The last cause was money: the agent spends everything on units, so a next tier stayed short of gold ~2,000 steps a game. The builder now keeps the cost of the target's next item that is short of gold or lumber from the agent's other orders (the availability mask, features.affordable; the item itself and free orders stay). Items due a minute before the game's end and missed, per game, with the builder carrying out:

fgself-94 from	due	missed
10:46 (queueing)	13.3	4.5
11:32 (frames resumed)	12.7	4.3
12:06 (the cost reserved)	13.3	2.1

Of the 2.1, ~0.3 are expansions (left to the agent).

fgself-94 (12:06-12:56, ~390 paired games each on echofast against insane): current 72%, current+z (a target, the builder carrying out) 39% (by 20-minute windows 79/73/97% of the AI's record), fgself-78's last 42%. With the builder 82-85% of the due items were made; the agent made 43-47% of those itself, flat for 50 minutes of training on the builder's labels. The targets can be followed; the agent does not follow them on its own yet.

fgself-95 (from fgself-94's last with its Adam state, 12:56): the builder only teaches (--z-build 2): its orders are labels, never carried out; a busy building's decision is labelled too (a hall's next tier while it trains workers). The cost of the target's next item stays reserved from the agent's other orders. The question: does the agent follow its targets on its own, and at what cost in wins?

fgself-95 after 30 minutes (~210 paired games each, not judged): current 72%, current+z (the agent alone with its target) 44%, fgself-78's last 38%. The agent now follows its targets on its own: in training it made 80-82% of the ~13 items due a minute before the game's end (fgself-93, the old pool and builder: ~50%), in current+z games 88%. With a target it plays far more varied games: 2.9 unit types a game (no target 1.6, fgself-78 1.4), 27 buildings (17, 20), tier 2 in 74% of its games (52%, 59%), 1.56 heroes (1.01, 0.89).

fgself-95 judged at 13:56 (~440 paired games each): current 73%, current+z 44%, fgself-78's last 38%; training games made 81% of the due target items in every 20 minutes (no rise yet). Verdict: the agent follows sensible build orders on its own (a question answered), and with a target it wins less: by race with a target / without, night elf 84% / 88%, orc 28% / 84%, undead 30% / 63%, human 26% / 53% (with people's targets 42%, with the built-in AI's 53%). current+z by 20-minute windows 83/82/91% of the AI's record: slowly rising. fgself-95 keeps running: the agent now has to learn to win with these builds (orc, whose rush the targets replace, loses most).

fgself-95 13:56-14:56 (~460 paired games each): current 71%, current+z 50% (by 20-minute windows 92/105/122% of the AI's record: it learns to win with its targets), fgself-78's last 39%; target items made on its own steady at 80%.

Correction (16:04). The laptop's evaluation worker read --z-build as a flag: 2 (a teacher only) became True, which the games compare as 1, so its current+z games had the builder carry out the target. Nearly all of fgself-93's and fgself-95's evaluation games were the laptop's, so their current+z numbers above (wins, and the 88% of due items made) are with the builder carrying out, not the agent alone. The training games ran on this machine and are right: there the agent alone made 80-82% of the due items. Fixed (evalworker.run_config: a number) and the worker restarted at 16:04:55; current+z is the agent alone from there.

fgself-95 16:06-16:26 with the fixed worker (~170 games each): current 78%, current+z (the agent alone) 58%, made 82% of its ~11.5 due target items, fgself-78's last 48%. The corrected numbers are close to the ones with the builder.

fgself-96 (from fgself-95's last, 16:30; head to head with fgself-95's and fgself-78's last): GRPO from mid-game branches (the user's idea; --grpo-share 1 --grpo-group 4 --grpo-coef 0.3). The first game of a launch between agents (seeded) records its commands up to a random step between 1 and 8 game minutes; 4 fresh games of that actor replay them to there (no inference: the recorded commands) and play on with new samples; when the group ends each branch's advantages get 0.3 times its return from the branch less the group's mean (per trained side). The advantages' spread is ~0.05 (adv/std, now logged): return gaps of ~0.1 weigh about as much as a step's advantage; a won and a lost branch ~6 times. More games between agents for it (--ai-share 0.8, from 0.95). Getting the replays exact took three restarts: a check that a branch reaches its source's state (branch_same, the players' units), the first step a replay differs in the log, and then the critters: they wander apart in two seeded games from ~5 s and block workers, so seeded games now start without them (W3S_RemoveCritters). From 16:47, 9 of the first 13 branches reached their source's state exactly (the others differed from steps 150-262: tree harvesting, a hero's skill; a branch that differs still starts from almost the same state).

By 16:56 only 45 of 89 branches were exact: fights with creeps do not replay alike (a probe: agents attack-moving to the nearest creeps, apart from 13 s), like the critters and the built-in AI. Restarted 17:00: only exact branches get the group's advantage (two needed; the others train with nothing added). Judged from 17:00.

fgself-96 judged at 17:56 (~330 paired games each from 17:01, echofast against insane): current 73%, fgself-95's last 73%, current+z 54%, fgself-78's last 37%: GRPO level after an hour. Its signal was small: 54% of the branches exact, and in most groups the returns hardly differed (median largest gap 0.04; a few groups split won and lost from one state). Target items made on its own in training 76% (fgself-95 80-82%; more games between agents). Next, fgself-97: the same with a stronger signal: branches at 1-4 game minutes (before most fights with creeps: more of them exact) and the group's advantage at full weight (--grpo-coef 1).

fgself-97 (17:58-18:56, ~350 paired games each): current 67% (by 20-minute windows 156/149/114% of the AI's record), fgself-95's last 71% (147/154/143%), current+z 55%, fgself-78's last 39%. Earlier branches were hardly more exact (56%), and most groups' returns stayed within 0.04. GRPO from mid-game branches did not help, at 0.3 or full weight; the line stops (the code stays: --grpo-share). What would make it work: branches the game replays exactly through fights with creeps (it does not: engine notes), or a value-free signal with less noise than one game's outcome.

fgself-98 (from fgself-95's last with its Adam state, 19:06; the league of fgself-95): back to the real game, Echo Isles under the game's rules against insane, the yardstick. Targets from a pool at the real rules' times (runs/fullgame/strategies-echo-real.json, 6,578: the normal built-in AI's sides of the 1,500 real-rules Echo Isles games and people's 1.29 1v1s at their own times, orderable items only, the owners at each event's step), --z-time-scale 1, the builder teaching, the cost reserve, the fixed upgrade costs. Head to head with fgself-78's last (the best real-game policy, ~87-89% against insane) and fgself-95's last. The question: does the line that follows build orders on echofast carry over to the real game, and how does it stand against fgself-78?

fgself-98 at 20:26 (~410 paired games each, the real game against insane): current 87%, fgself-95's last 86%, fgself-78's last 86%: level with the best real-game policy, the races apart: human 80% / 82% / 71%, undead 71% / 67% / 83%, night elf ~97-100%, orc ~92-96%. The echofast line carried over (human up ~10 points, undead down ~12-16). With a target 61% (by 20-minute windows 124/121/126/135% of the AI's record, rising; human 24%, night elf 90%, orc 65%, undead 66%); in training the agent alone made 68-70% of the due items (on echofast 80-82%: the real rules' times are new to it). It keeps training on the real rules for now.

fgself-98 20:26-21:26 (~295 paired games each): current 78% (human 59%, undead 58%), current+z 64% (human 31%), fgself-95's last 90% (human 87%, undead 75%), fgself-78's last 87% (human 76%, undead 84%). Training on the real rules with targets in 75% of the games pulled its play without one down (87% -> 78%), human and undead most. Over both windows fgself-95's last won 617 of 705 (87.5%), fgself-78's 606 of 703 (86.2%): the echofast line's last is the best real-game policy so far, by a little. Next, fgself-99: from fgself-95's last on the real rules with targets in 30% of the games (--z-share 0.3).

fgself-99 (21:29-22:26, ~285 paired games each): current 80% (human 64%, undead 63%; by 20-minute windows 166/178/134% of the AI's record), current+z 47%, fgself-95's last 87%, fgself-78's last 85%. Fewer target games did not stop it: both runs on the real rules pulled fgself-95's last down, while echofast training made it. Next: train on echofast (as fgself-95, targets in 75% of the games) and evaluate on the real game (--eval-map, new: the evaluation games on another map, with its costs and its own strategy pool).

fgself-100 (from fgself-95's last with its Adam state, 22:32; the league of fgself-95): fgself-95's recipe goes on (echofast, targets in 75% of the games, the builder teaching, the cost reserve), its evaluation games the real game (--eval-map "(2)EchoIsles", --eval-z-pool runs/fullgame/strategies-echo-real.json). Head to head with fgself-78's and fgself-95's last. The question: does more echofast training lift the real game past 87.5%?

fgself-100 at 23:56 (~455 paired games each, the real game against insane): current 88%, fgself-95's last 88%, fgself-78's last 84%; with a target (the real-rules pool) 73%, the agent alone making ~70% of its due items. In the last half hour 92% / 90% / 87% / 76%. echofast training keeps the real game where it was (the runs on the real rules pulled it down) and may lift it a little; it goes on.

fgself-100 23:56-00:56 (~350 paired games each, the real game against insane): current 95% (194% of the AI's record; human 90%, undead 92%, orc 100%, night elf 100%), fgself-78's last 88% (human 80%, undead 83%), fgself-95's last 88%; with a target 74% (human 50%, undead 78%, orc 81%, night elf 94%). A new best in the real game: about 2.5 hours more of echofast training, judged on the real game (latest checkpoint at 00:56: runs/fgself-100/checkpoints/0000000012992241.pt). It goes on.

fgself-100 paused 2026-10-20 01:05 for a machine shutdown, still improving: from 23:56 ~400 paired games each, current 95%, fgself-95's last 89%, fgself-78's last 88%, with a target 74%. Its last checkpoint: runs/fgself-100/checkpoints/0000000013649982.pt. Resumed 2026-10-20 10:13 after the reboot (selfplay_restart.py, the same run).

fgself-100 10:15-11:24 after the resume (~390 paired games each, the real game against insane): current 92% (human 91%, undead 82%, orc 97%, night elf 99%), fgself-95's last 87%, fgself-78's last 86% (human 67%); with a target 76% (human 50%), the agent alone making 70% of its ~13 due items. Over the ~800 games since 23:56 current won ~93-95%, fgself-78 ~87%: confirmed, the real game's best. By 20-minute windows 201/188/176/176% of the AI's record: it may be levelling off. Latest checkpoint at 11:24: runs/fgself-100/checkpoints/0000000020063584.pt.

fgself-100 11:24-11:54 (~155 each): current 87%, fgself-78's last 90%, fgself-95's last 86%. After its peak (95% at 01:05, 97% right after the resume) it slid, its KL term still pulling toward fgself-78 (ref_kl ~0.11). Its best checkpoint: runs/fgself-100/checkpoints/0000000013649982.pt (01:05).

fgself-101 (from fgself-100's 01:05 checkpoint, 11:56; the league of fgself-100): the anchor moved to that checkpoint (--init): the KL keeps the policy near its best instead of fgself-78. Head to head with that checkpoint and fgself-78's last. (Earlier anchor moves were level, fgself-40, -46; a rolling anchor was worse, fgself-82.)

fgself-101 judged at 12:54 (~345 paired games each, the real game against insane): current 96% (human 97%, undead 86%, orc 100%, night elf 100%), its start (fgself-100's 01:05 checkpoint) 94%, fgself-78's last 87%; with a target 81% (fgself-100: 74-76%; human 69%, from ~50%). By 20-minute windows 192/172/182% of the AI's record: no slide (ref_kl ~0.015 to the new anchor; fgself-100's was ~0.11 to fgself-78). The anchor at the best checkpoint holds the policy there and the target games got better. It goes on.

fgself-101 12:54-13:54 (~360 paired games each): current 93% (human 83%, undead 90%), its start 93% (human 93%, undead 84%), fgself-78's last 87%; with a target 83% (human 65%). By 20-minute windows 201/194/173%: level with its start, holding (not sliding) but not climbing. Night elf and orc win 96-100%; human and undead are what is left. Next, fgself-102: training games' learner races weighted to them (--learner-races human:0.35,undead:0.35,orc:0.15,nightelf:0.15; evaluation games stay uniform).

A video of fgself-101 (32 minutes, undead against insane night elf, lost) showed the camera at the night elf base the whole game: four of the learner's ghouls stood idle there from ~5 minutes to the end (their hit points 1, 7, 67 and 70 the whole time, unattacked) beside the AI's wisps, and the video's director counted units near enemies as a fight. A fight now needs blows struck (video.Director: hit points lost within ~1,500 of it in the last 4 s); that game's camera then went to real fights ~45% of the time and cycled through both armies and bases otherwise. The idle ghouls are a finding of their own: undead loses most games against night elf, long ones, and parks ghouls (its lumber harvesters) at the enemy base.

fgself-102 judged at 14:54 (~355 paired games each, the real game against insane): current 95% (human 92%, undead 89%, orc 97%, night elf 100%), fgself-100's 01:05 checkpoint 92% (undead 79%), fgself-101's checkpoint (its start) 90% (undead 70%); with a target 83% (human 62%). Training the weak races more lifted undead by 10-19 points and the whole by 3-5; human held. It goes on. Next (proposed to the user: the road past the built-in AI): a league that trains mostly against itself, past snapshots and exploiters, with a yardstick of its own (the policy against frozen snapshots, not only the built-in AI, which it beats ~95%).

The yardstick (--eval-vs, new, 15:58): a share of the evaluation launches (here a quarter) play the current policy against one of --eval-checkpoints (agents on both sides, no tax; rows eval_key "vs /"; the league's records leave them out). fgself-102 restarted with it against fgself-101's and fgself-100's best checkpoints: over 50% against them is progress the built-in AI no longer shows.

First snapshot games (16:11, small): fgself-102's current policy beat fgself-101's checkpoint in 69% of 36 games and fgself-100's best in 21 of 21.

fgself-103 (from fgself-102's checkpoint 11841833, 16:13; the league of fgself-102): AlphaStar's league (the user: "keep improving with self-play to become superhuman", "get inspiration from Dota Five and AlphaStar"; ideas, "Past the built-in AI"). The built-in AI in 15% of the launches (a floor and a curriculum), a main exploiter in 15% (--exploiter-share 0.15), of the rest 40% against itself and 60% against past snapshots by PFSP (--self-share 0.4, --max-past 12): ~29% self, ~43% past, 15% exploiter, ~13% the built-in AI (AlphaStar 35/50/15). The yardstick: the current policy against its start and fgself-100's best in 40% of the evaluation launches (--eval-vs 0.4), beside the built-in AI. The win rate against insane may dip first (fgself-66: -16 points in an hour with half its games against itself); the question is whether it beats its own past more.

Its first 10 minutes read 26% against its start in 65 snapshot games, but the races decided it: the launches' races are drawn once for a launch's games, and the current policy had played human against the checkpoint's orc in 19 (lost all) and undead against night elf in 10 (lost all). Restarted 16:25 with each matchup played both ways round (the next snapshot launch swaps the races); judged from there.

fgself-103 at 18:24 (2 hours of the league; ~600 mirrored snapshot games each, ~55 against insane each): the current policy beat its start in 55% (by thirds of the games 56/49/61%) and fgself-100's best in 55% (55/52/58%); against insane it held 97% (its start 100%, fgself-100's best 95%; with a target 83%). The main exploiter restarted once (it reached 70% against the main agent). The league is a small gain that grows, without the dip against the built-in AI that half its games against itself cost fgself-66. It goes on.

fgself-103 judged at 19:24 (3 hours of the league; ~900 mirrored snapshot games each, ~80 against insane each): against its start by hour 51/62/49%, against fgself-100's best 54/56/48%; against insane 100/93/79% (its start 100/100/93%, fgself-100's best 93/96/96%). The league gained in its second hour, then drifted back, and the built-in AI started to slip: no lasting gain in three hours (AlphaStar's took weeks; the exploiter restarted once). Its second hour's checkpoint (17:56, runs/fgself-103/checkpoints/0000000008386382.pt) is kept as a reference.

fgself-104 (from fgself-102's checkpoint 11841833, fgself-103's start; fgself-102's recipe and league: the built-in AI in 95% of the launches, the weak races trained more): OpenAI Five's reward decay (--shaping-decay 0.6, new: every reward but the outcome times 0.6^(game minutes / 10)). Snapshot games (--eval-vs 0.4) against its start and the league's best hour (17:56): do the decay or the league beat the same start more?

fgself-104 (19:27-20:24, ~250-350 mirrored snapshot games each): against its start 46% (by halves 43/50%), against the league's best hour 44% (49/40%); against insane 97% (34 games; its start 100%). The reward decay did not help (slightly worse than its start after an hour).

fgself-105 (from the same start, fgself-102's checkpoint 11841833, fgself-102's recipe): OpenAI Five's fresh data, each sample used once (--epochs 1, was 2: updates twice as often, staleness lower). The same snapshot games.

fgself-105 (20:24-21:26, ~350 mirrored snapshot games each): against its start 54% (halves 58/49%), against the league's best hour 48% (49/47%); against insane 100% (38 games); staleness ~1.1 (was ~1.5-1.8). Level: fresher data, no stronger policy. The three ideas since fgself-102 (the league, the reward decay, one epoch) all came out about level with their common start.

fgself-106 (from fgself-105's last, 21:27; its recipe): the line goes on, and its evaluation worker also plays classifier-free guidance toward the target (the user's idea from image generation; --eval-guide 2, new: current+zg, current+z's games with the order's logits free + 2 * (with - free), the strategy's block zeroed for "free"). The question: does guidance make the agent follow its targets more (due items made), and at what cost in wins? Snapshot games now 20% of the evaluation launches (against its start and fgself-102's 11841833).

Result (22:08, ~32 games each on the laptop): guidance did not help. current+zg won 50% against insane (16 of 32), current+z 91%; the due target items missed: 5.3 and 5.0 a game. Stopped.

fgself-107 (from fgself-106's checkpoint 3453342, 22:08; its recipe and league): what the agent could not see or do (the user's question). --inputs 3: per hero its abilities (level, can be cast now: learned, cooled down, the mana; cooldown left; a skill point may go there) and items (held, usable, heals, mana, a town portal), what each building makes (a worker, an army unit, a hero, research, its next tier), the time of day, and the enemy units and buildings it saw and no longer sees (where and how long ago; an attack on one is an attack-move there). New channels start with zero weights. --tech-mask: orders the player lacks the tech for are masked (requirements, research levels, the hero limit); --valid-mask: a hero's orders it can't carry out now. Checked first on recorded games: the masks block only what the game would refuse then.

Result (22:55, ~45 minutes): worse than its start: 41% of 225 head-to-head games, flat by thirds (40/43/40%); 89% against insane (54 games), 67% against fgself-102's 11841833. Suspect: a remembered enemy is a new token that looks like one in sight until its marker's weights learn, and an attack on it sends units to a stale place. Stopped.

fgself-108 (the same start, recipe and league, 22:57): the ablation, --inputs 2 (hero state, what buildings make, the clock; no memory), the masks on. Level with the start: the memory hurt (next: fade the remembered enemies in on a schedule); below: the hero inputs or the masks.

Result of fgself-108 (23:23, ~25 minutes): head to head 48% of 88 against its start, 55% of 154 against fgself-102's 11841833, 93% against insane. On the rating scale (more games) the start, fgself-106's last, is 2070 ± 19 (2128 on fewer games), fgself-108's first hour 2024 ± 22, fgself-107's 2044 ± 18, fgself-102's best hour 2062: the new inputs and masks are about level, their start a strong checkpoint. Kept as the base; the memory waits for its fade-in run (--memory-fade, new: older networks and the anchor never see the remembered enemies, attention bias -30; the learner's rises from -8 to 0).

fgself-109 (from fgself-108's last, 23:26; its recipe and league): strategic variety, the user's direction. Narrow targets (runs/fullgame/strategies-echofast-narrow.json, the real game's for evaluation: a game's heroes, its main army and what they need, drawn evenly over 230-307 builds a race; other buildings and research free, other heroes and army units not) and more kinds of conditioning (--z-modes 0.15 --z-mode-reward 0.1: no rush, rush, expand, no expansion, workers, creeping, harass, scouting, army size, low upkeep). Watch the due items made, each kind kept, the win rate and the rating.

Result of fgself-109 (00:24, 1 hour): level on the rating (its first hour 2084.5 ± 15, its start 2084.7 ± 18), 96% against insane, 49% of 347 head to head with its start. With a target it wins 68% (rating 1649 ± 37) and makes 75% of the due target items (80% in training). Kinds kept: workers 88%, army size 84%, no expansion 77%, rush 76%, harass 75%, scouting 71%, creeping 70%, expand 47%, low upkeep 34%, no rush 14%; the hard ones did not improve between the halves (no rush 18% → 10%): a whole violation is worth 0.1 against the outcome's 1.

fgself-110 (from fgself-109's last, 00:25; its recipe and league): rewards by part (GDPO, the user's idea from LLM RL; --reward-parts 1): a value and an advantage per part (outcome; shaping; strategy, the target's items and the kinds), each normalized on its own and summed with equal weights; a part spreading less than a tenth of the outcome's is not scaled up. At the first update the strategy part's advantages spread 0.008 to the outcome's 0.079: now on the same footing. The question: are the kinds and targets followed more, and what does it cost in wins?

Result of fgself-110 (01:00, 31 minutes): worse. Rating 2033 ± 22 to its start's 2084 ± 28, 40% of 150 head to head with it; with a target 51% against insane (68% in fgself-109), the targets and the kinds followed as before. Equal weights also gave the material shaping its full weight back (half had helped, fgself-73); a retry would weigh it less (--part-weights 1,0.5,1). Stopped for the night.

fgself-111 (from fgself-109's last, 2026-10-21 10:10; its recipe and league): what the agent could not do (the user's question): the neutral buildings' wares as orders (--neutral 1: a tavern's heroes, a camp's mercenaries, the goblin merchant's items, bought at the building the pointer picks; the tavern heroes' skills and spells; every network grown to the vocabulary, new orders rare at first), a heuristic teacher (--teach 1, the user's idea), and narrow targets with tavern heroes and mercenaries (strategies-*-narrow-tavern.json: ~2,600 of ~5,300 targets have one, people's builds; the race filter had dropped them). Two restarts in its first half hour: the shops always in the view (the game's minimap shows the neutral buildings; in sight only, the tavern could not be pointed at and no lesson for it counted: none of 21 tavern-hero targets made in 60 games), and a hired unit counted as made (a purchase has no training event). Then, still no tavern hero hired in ~2,400 games, one cause after another: the teacher sent only workers in sight of the tavern (now the nearest worker, however far); its few lessons drowned among the target builder's labels (--teach-coef 0.5, a cloning term of their own); a worker sent to buy was turned back half a second later by its next order (a buy order now holds 40 s, BUY_COMMIT_STEPS); a buy order's pointer seldom picked the shop (it now goes to the nearest building that sells it, as a harvest order goes to the nearest mine). A traced game (12:00) showed the rest: ~36 buy orders a game, all the goblin merchant's items, all sent to workers inside the gold mine and refused. Now the merchant's items are a hero's only (neutral.who_buys) and a unit inside a mine gets no buy order (the hero mask's not-hidden bit); a resumed run takes these tables from the game's data, not its checkpoint (model.trained). Still none: a probe game with a tavern-first target built no altar and held under 100 gold all game (footmen), so the hire stayed unaffordable and masked; the target builder's reserve (the next item's cost kept from the other orders) skipped hires, now it saves for them (12:22). And the root cause (12:45): every buy order from a base was refused by the game. The tavern stands in the fog at the start, and no order can target a unit the player hasn't seen; the few hires (5 in ~700 games) came from units already beside a shop. Now the buyer walks to the shop and right-clicks it in reach (a worker from its start hired a Pandaren 77 steps later), the purchase pending 160 steps and the buyer held 120. Games with the tavern hero due win about a third either way (only long games reach it).

Result of fgself-111 (14:25, 4 hours 15 minutes, 1 hour 40 minutes since the fog fix): level, and it now buys. Rating 2031 ± 20 (its second hour) to its start's 2010 ± 9; 49% of 1,082 head to head with its start, 50% of 1,077 with fgself-102; 93% of 284 against insane (its start 97% of 287, the same hours); with a target 61% of 224 (rating 1624 ± 65 to fgself-109's 1643 ± 37). Due target items made 74% (night elf 82%, orc 78%, human 77%, undead 69%). In the 2,751 training games of its first 40 minutes after the fix: 413 tavern heroes hired (Alchemist 81, Beastmaster 70, Pandaren 65, Naga 63, Dark Ranger 30, Tinker 30, Pit Lord 15, Firelord 9) and ~1,500 mercenaries (forest troll shadow priests 596, ogre maulers 460, troll berserkers 267, mud golems 174), mostly ordered by workers. Kept as the next base: what it can do grew at no cost.

fgself-112 (from fgself-111's last, 2026-10-21 14:30; its recipe and league): self-distillation with privileged information (OPSD, the user's idea from LLM training; --opsd-share 0.2 --opsd-coef 0.05): a fifth of the learner's training games act on a privileged view (every enemy unit, marked when out of sight, and the opponent's resources; a new input path that starts at zero), so the network learns to use what a player can't see; in the other games a KL pulls its orders from the player's view toward the same network's orders given the privileged view (no gradient through the teacher). Never in evaluation. Its first hour went to four bugs: the remembered-enemy columns switched on by the rows' width (every actor), a privileged batch narrower than its own units, the graphed update's statistics, and the teacher pass's pointers past the privileged views' width (a device-side assert); checkpoints are now saved whole or not at all (a stopped learner left a truncated one). And a finding about --neutral: the laptop's evaluation process had predated the neutral orders, so fgself-111's real-game evaluations never bought anything. With them on (from fgself-112), every policy's workers bought ~2 mercenaries a game, never taught (a new order's bias is low, but each worker may give one every half second), and won less against insane: fgself-109 84% of 37 (94-98% before), fgself-111 92% of 36, fgself-112 81% of 47; the games with a purchase won 74-90%, those without 90-94%. Restarted at 15:03 with --neutral 2: the tavern's heroes and the merchant's items, no mercenaries (none in the targets either).

Result of fgself-112 (16:30, 1 hour 25 minutes since the restart): no gain. 44% of 327 head to head with its start (thirds 45%, 39%, 50%); against insane 89% of 72 (its start 95% of 74 in the same games' rules); with a target 56% of 70. On the rating scale its second hour 1937 ± 28 to its start's 1908 ± 16 (both in these rules), its first hour 1857 ± 14 (the mercenaries and the crashes). The self-distillation's KL stayed tiny (~0.004): the teacher, the same network with a privileged input that starts at zero, had barely learned to use what it sees by then. A retry would need the privileged games to learn first (a longer run, or the privileged path trained on its own before the KL). Ratings fixed: a checkpoint from a run without the shops, evaluated with them, is now a player of its own (fgself-109/4933795 +shops: 1730 ± 17, to 2010 ± 9 without); as one player it had made wins over a weakened fgself-109 count as wins over a strong one.

fgself-113 (from fgself-111's last, 2026-10-21 16:28; its recipe and league, --neutral 2): scenarios (the user's idea: the micro curriculum's way kept on; --scenarios clash,raid,rich --scenario-share 0.25): a quarter of the training games start in a situation the harness sets up: an army and a level-3 hero each between the bases (clash), the opponent's army at the learner's base (raid), or 1,500 gold and 700 lumber more each (rich: tiers, tech and heroes early). Evaluation games never have one. The question: does meeting these situations more often make the policy stronger in whole games?

fgself-113 stopped (17:35, 1 hour): confounded. 36% of 186 head to head with its start from its first third (32%), with a target 46% of 50; against insane level (90% to its start's 89%). Looking for why both forks of fgself-111 (fgself-112, fgself-113) began far below their start found two bugs: (1) extend_vocab listed a toggle's two orders (the Tinker's robogoblin and unrobogoblin, and one more pair) in the order of its process's hash seed, so a fork or a restart could give one order for the other, and (2) for the same reason a fork found its parent's snapshots of another vocabulary unusable: both started with a league of one old snapshot (fgself-78/0) instead of fgself-111's own. Now the extension is sorted, a run takes the vocabulary of the weights it trains, and a network of a reordered vocabulary is put in the run's order when it loads. fgself-112's verdict (OPSD, no gain) carries the same confound.

fgself-114 (from fgself-111's last, 17:30; its league of 5 past snapshots, --neutral 2): fgself-113 again, clean: scenarios in a quarter of the training games.

Result of fgself-114 (18:55, 1 hour 15 minutes): level. 50% of 304 head to head with its start (thirds 50%, 43%, 56%), 62% of 290 with fgself-102; against insane 92% of 66, its start 92% of 66; rating 1907 ± 16 (first hour) and 1904 ± 30 to its start's 1910 ± 11. With a target a little worse (52% of 91 against insane; rating 1547 ± 40 to fgself-111's 1591 ± 26). The scenarios themselves were learned slowly (clash 46% → 51%, raid 44% → 55% between the halves; rich ~70%).

fgself-115 (19:00): weight averaging (the user's idea's second half; a model soup): the mean of fgself-111's last and fgself-114's last (the start and the scenario-trained policy, distance 0.79 apart in the order head), trained on with fgself-111's recipe (no scenarios; Adam's state from fgself-114). Evaluated head to head with both parents: is the soup better than either?

Result of fgself-115 (20:25, 1 hour 25 minutes): level, worse with a target. 47% of 329 head to head with fgself-111's last, 51% of 279 with fgself-114's; against insane 93% of 96 (the parents 91% and 93%); rating 1887 ± 17 (first hour) and 1974 ± 25 to fgself-111's 1920 ± 9 and fgself-114's 1882 ± 20. With a target 43% of 90 against insane (rating ~1470 to fgself-111's 1591). Averaging two checkpoints that play alike gave nothing; a soup of specialists that differ (each trained on one scenario family, the user's original plan) remains untried.

fgself-116 (from fgself-111's last, 20:30, --neutral 2): rewards by part, again (GDPO; fgself-110 was worse with equal weights, the material shaping back at full weight): --reward-parts 1 --part-weights 1,0.5,1 (outcome, shaping at half as in fgself-73, strategy). Each part's advantages are normalized on their own, so the strategy reward is no longer drowned by the outcome's spread. The question: are targets followed better, at no cost in wins?

Result of fgself-116 (21:55, 1 hour 25 minutes): worse. 38% of 299 head to head with its start (thirds 43%, 39%, 31%), 41% of 349 with fgself-114's last; against insane 88% of 93 (its start 88% of 90); rating 1862 ± 16 and 1824 ± 26 to its start's 1918 ± 8; with a target 41% of 104 against insane (falling: 47%, 43%, 34%; rating 1451 ± 36 to 1591). It followed the targets a little better in training (due items made 78% → 80%, 2.0 items missed a game to fgself-111's 2.6). Rewards by part are out for now: each time the strategy reward weighs more (fgself-89, fgself-110, fgself-116) the targets are followed a little more and the games are lost more.

fgself-117 (from fgself-111's last, 22:00, --neutral 2): the enemies it saw before, faded in (--inputs 3 --memory-fade 600): fgself-107 added them at once and fell below its start; here the learner's attention bias on them goes from -8 to 0 over its first 600 updates (~35 minutes), the anchor and older networks keep them hidden. Does knowing where the enemy was make it stronger?

Result of fgself-117 (22:55, 55 minutes): worse, as fgself-107. 35% of 130 head to head with its start (thirds 49%, 35%, 23%), 43% of 267 with fgself-114's last; against insane 83% of 69 (its start 95% of 56); with a target 34% of 65. Fading them in did not help: twice now the remembered enemies made it weaker. Before a third try, a look at what it does with them (attack-moves to where an enemy was? the neutral buildings among them?).

fgself-118 (stopped after 8 minutes: a fresh Adam, the privileged path shifting the parameters; now matched by name) = fgself-119 (from fgself-111's last, 23:01, --neutral 2): OPSD again, clean (fgself-112 started with a league of one old snapshot and two spell toggles swapped): --opsd-share 0.2 --opsd-coef 0.05, overnight, so the privileged input (which starts at zero) has time to learn before the distillation can carry anything.

fgself-119 stopped (23:44, 43 minutes): 46% of 240 head to head with its start, 42% of 94 with fgself-114's last; no verdict. Two rules of the game had been wrong for every run, found while answering the user's question about the first hero: (1) the first hero is free (1.29's hero token: only its food, at the altar or the tavern; tested), but the cost mask, the target builder and the teacher held it back until 425 gold and 100 lumber, which may be why the first hero came ~30 s after the targets' time; (2) about half the purchases never went through: a worker's right-click on the tavern could leave the tavern serving none of its player's units for the rest of the game (a fresh worker beside it was refused too). The buyer now walks to the shop and the sale goes through in reach, with no right-click (21 of 22 test hires). Every run from fgself-111 on trained under both.

fgself-120 (from fgself-111's last, 23:45, --neutral 2, its recipe): the corrected rules, overnight: the free first hero and reliable purchases, nothing else changed. A new baseline: does the policy get its first hero sooner, hire tavern heroes as targets ask, and get stronger?

Result of fgself-120 (01:25, 1 hour 40 minutes): level, the new baseline under the corrected rules. 48% of 313 head to head with its start, 55% of 387 with fgself-114's last; against insane 87% of 87 (its start 89% of 79 in these rules); rating 1929 ± 16 and 1922 ± 22 to its start's 1916 ± 7; with a target 42% of 90 against insane. The first hero comes at a median 60 s (79 s before); tavern heroes hired in 7% of the training games, and 32% of the targets that ask for one (153 of 476).

fgself-121 (from fgself-120's last, 8527594, 01:30, overnight): OPSD, clean and under the corrected rules (--opsd-share 0.2 --opsd-coef 0.05; Adam's state matched by name): the first attempt that has its parent's league, one vocabulary and the game's rules right. Head to head with its start and fgself-111's last. Stopped at 01:40 for a shutdown after 10 minutes (682214 steps); to go on from there. Resumed 09:50; result (10:45, ~1 hour of training): no gain: 45% of 182 head to head with its start, rating 1932 ± 17 to its start's 1976 ± 24; against insane 94% of 125 (its start 98%). Self-distillation with privileged information gave nothing in five attempts (the first four confounded); its privileged path, which starts at zero, barely learns in an hour.

The audit (2026-10-22, the user's idea: a subagent looking for bugs like the tavern's). Seven confirmed bugs where the code's assumptions about the game were silently wrong; all fixed and tested, the game mechanics in game: (1) a dead hero was never revived (the mask counted it as owned; and the revival is the altar's revive order on the dead hero, accepted ~7 s after the death, not its training order); (2) undead could not expand: the haunted gold mine's order went down the buy path to the goblin merchant (no undead run had built one), now a target order on a gold mine; (3) on echofast, 44 of Echo Isles' 75 creeps kept their full hit points while every player unit had half (creeping looked twice as costly as in the real game), and the tavern's and camps' units kept full prices; (4) the target builder's and the teacher's labels at idle buildings were dropped (never carried out, they "started nothing"), and their payment check didn't know the first hero is free; (5) a unit sent to buy was held 60 s whatever happened (now freed once it reaches the shop, or at once on a refusal), and a ware could go to a building that doesn't sell it; (6) a building's "being made" input read 0 unless some worker was on lumber; (7) a worker mining gold could not be sent to another mine (expansions could not be staffed). Also: the strategy modes' "expand" counted a second hall at home. And a new measure: the episode rows count real expansions (a hall far from home at a gold mine): 0% of the games of fgself-120 and fgself-121, in training and in the real game; their extra halls (6-11% of games) were all at home.

fgself-122 (from fgself-120's last, 10:50, its recipe): the game as the audit corrected it (revival, haunted mines, mine transfers, buyers, labels, creeps), nothing else changed: the new baseline. Head to head with its start and fgself-111's last.

Result of fgself-122 (11:55, 1 hour 5 minutes): level to a little lower, the base of what follows: 48% of 301 head to head with its start, 43% of 227 with fgself-111's last; against insane 89% of 66 (its start 97% of 65); rating 1917 ± 16 to its start's 1944 ± 15; with a target 52% of 71. Its play: no expansion in any game (training or the real game), a second hall at home in 16% of the training games, ~1.2 creeps killed a game, a tavern hero in 17% of the games, no haunted mine.

fgself-123 (from fgself-122's last, 5838960, 11:57, its recipe; head to head with its start and fgself-120's last): expansions taught. People expanded in 81% of 95 Echo Isles games (at the natural, ~800 north of its mine, by ~6 minutes); no policy ever had. The teacher, when its target's expansion is due (a hall or a haunted mine among its items: 19-52% of people's targets, ~90% of the AI's), has the army clear the natural's camp (6+ units), then the nearest worker build the hall at people's spot (undead: an acolyte haunts the mine), then three workers mine there; a target's hall counts as made only as a real expansion (a second hall at home had counted). Night elves not yet (no entangle order). The question: does it expand, and at what cost in wins?

Result of fgself-123 (12:27, 30 minutes): much worse, stopped. No expansion in 1,259 training games; the camp lesson (the whole army to the natural's camp, ~26 lessons a game) dragged armies away: games with an expansion due won 13% (32% in fgself-122), all training games 45% (59%), 36% of 55 head to head with its start. The hall lessons (1,108) never applied: the camps stood.

fgself-124 (from fgself-122's last, 12:31, its recipe): expansion teaching, take two: no camp lesson; in a third of the training games both naturals' camps are gone from the start (--scenarios natural --scenario-share 0.33, harness op 88), so the teacher's hall and worker lessons can apply when the target's expansion is due. Evaluation games are untouched (the camps stand there).

Result of fgself-124 (13:00, 30 minutes): no expansions, stopped. 3 expansions in 1,698 training games (the naturals' camps gone in a third of them), and a second hall at home in 28% (16% in fgself-122): the 6,890 hall lessons (a worker builds the hall at a point ~6,000 from home) taught the policy "build a hall", not where. 47% of 105 head to head with its start. The spot itself is fine (built by hand in game for human and orc).

fgself-125 (from fgself-122's last, 13:08; its league): an expand order of the agent's own (--macros 1, macros.py): a worker's order "take the natural", the agent builds the race's hall where people build it (undead: haunts the mine); it may be given while the natural is free and the hall affordable. The teacher's expansion lesson is that order, the target builder keeps a due expansion's cost from the other orders, the naturals' camps are gone in a third of the training games as in fgself-124. Every network grows to the new order (its row rarer than any order now).

Result of fgself-125 (14:55, 1 hour 50 minutes): level to a little better, the new base: 52% of 307 head to head with its start (thirds 45%, 52%, 58%), rating 1913 ± 21 (its second hour) to its start's 1871 ± 15; against insane 81% of 68 (its start 91% of 65); with a target 43%. It expands: a quarter of the training games with the naturals' camps gone (1,828), from 16% to 28% within the first hour; those games won 15% (all of them 48%, the others 57%): on echofast's short games an expansion rarely pays off yet. With the camps standing (the real game) the expand order stays masked: no expansions there.

fgself-126 (from fgself-125's last, 8001948, 15:03): a creep order (macros.CREEP): an army unit or a hero attack-moves to the natural's camp while it stands and nobody holds the natural; the teacher shows it to the heroes once the army is 6 strong and the target's expansion is due (not the whole army every 10 s: fgself-123). The naturals' camps are gone in a fifth of the training games (a third before), so it meets the camps more often.

Result of fgself-126 (15:57, 55 minutes): stopped, its creep and expand orders misused. It creeps (3.1 creeps killed a game with the camps standing, ~1 before) and expands with them (12.9% of those games), but the games that expanded won 1%; with a target 14% of 28 against insane; 46% of 154 head to head with its start (recovering: thirds 33%, 39%, 65%). A second audit (the user's request) found why, among others: (1) every worker sampled the expand order on its own: 8-18 walked to the natural in waves, the gold spent meanwhile, no hall built (63 hall orders in one filmed game); (2) past 96 own units the newest were cut from the view (the rows' first-seen order): 12-17% of the games, won 2-5%; (3) the creep order was allowed to any unit from minute 1 (436 attack-moves in a game), not under the teacher's conditions; (4) a worker's build order held only 30 s, the walk to the natural; (5) undead's haunted-mine expansions didn't count in the strategy modes. All fixed: one expand a step and none while one is under way (its hall's cost kept), heroes/army/producers/workers kept past the cap before farms and towers, creep with 6 army units from second 150, build orders held 200 s. (A dead tavern hero still can't be revived: rare.)

fgself-127 (from fgself-125's last, 8001948, 15:59; its league): the expand and creep orders with the audit's fixes, the naturals' camps gone in a fifth of the training games.

fgself-127 at 2 hours (17:55): better than its start, and it expands in the real game. Rating 1850 ± 20 (its second hour) to its start's 1791 ± 14 (its first hour 1742 ± 19: the dip as it learned the new orders); 49% of 369 head to head with its start, rising (thirds 39%, 50%, 57%); against insane 82% of 77 (its start 83% of 81); with a target 22% of 73 (the weak point: many targets ask for an expansion). In the real game it expanded in 14 of 150 evaluation games (won 3) and kills ~3 creeps a game (~1 before the creep order); in training ~12% of the games expand. Kept running.

Result of fgself-127 (19:25, 3 hours 25 minutes): better, the new base (its best hour's last checkpoint, 13485938 at 18:59): rating 1874 ± 19 in its third hour to its start's 1810 ± 12 (then 1817 ± 27: it peaked); 50% of 438 head to head with its start; against insane 81% of 95 (level with its start). With a target it stays weak: 24% of 126 in the real game, worst with people's build orders (21 of 106; the built-in AI's 9 of 20), and not because of expansions (28% of the targets with one, 20% without). In training, games with a target win 49% (51% against the AI), without one 77% (82%).

fgself-128 (from fgself-127's 13485938, 19:30; Adam's state the nearest saved one's): targets not enforced (--z-enforce 0). The mask that lets a side with a target make only the target's items (narrow targets: no hero or army unit outside them until the horizon) may keep it from answering what the opponent does. Following then rests on the target's input, the builder's labels and the strategy reward. Does target play win more, and is it still followed?

Result of fgself-128 at 2 hours (21:25): better, the new base; target play more than doubled. With a target 58% of 95 against insane in the real game (fgself-127: 24%), the target's due items still made (70% there, 71% in training to 75%); its with-target rating from 1522 to 1670 in its second hour. Without one 93% of 92 (its start 85% of 94). Rating 1893 ± 20 (second hour) to its start's 1864 ± 21; 49% of 250 head to head with its start, 58% of 259 with fgself-125's last. In training, games with a target won 67% (49% in fgself-127). By race with a target: night elf 28 of 28, undead 11 of 18, orc 12 of 23, human 4 of 26 (the weak one). In the real game it expanded in 11 of 187 games and won none of them. Kept running.

A silent refusal behind the human weakness (21:40): with a target, human made 40-67 peasants and only footmen and missed its keep. A hall's tier order is refused while the hall trains anything (tested in game: the keep order behind queued peasants refused; with the queue cancelled first, accepted and built), and the policy trains workers back to back, so its tier orders were refused for all races (2% of the games reached tier 3). The target builder cancelled the queue already; now the agent does too.

fgself-129 (from fgself-128's last, 8716868, 21:31, its recipe; fgself-128 stopped there, 2 hours 20 minutes): tier orders that go through (the hall's worker queue cancelled first), nothing else changed. Do targets that tech get followed and won, and does the policy tech more on its own?

fgself-129 at 1 hour 15 minutes (22:55): the best yet. 64% of 253 head to head with its start (thirds 62%, 58%, 71%), 56% of 230 with fgself-127's best; rating 2029 ± 32 in its second hour (1941 ± 19 its first) to its start's 1873 ± 22; against insane 96% of 52, with a target 70% of 56 (fgself-128 58%, fgself-127 24%). With a target by race: night elf 12 of 12, undead 18 of 22, orc 10 of 16, human 5 of 14. In the real game 63% of the games reach tier 2 and 13% tier 3; 7 expansions in 126 games, ~2 creeps a game. Its army is still tier 1 (footmen 31%, ghouls 30%, archers 15%, grunts 12%): the next variety to work on. Kept running.

Result of fgself-129 (00:15, 2 hours 35 minutes): better, the new base. 59% of 471 head to head with its start; rating 1930-1952 ± 19-31 by hour to its start's 1875 ± 16; against insane 94% of 72, with a target 70% of 71. Why its army stays tier 1: the policy seldom asks for anything else (in 12 filmed games, 401 ghoul and 300 footman train orders to 3 for riflemen and 2 for sorceresses): a choice, not a refusal. A target asks only for each item once.

fgself-130 (from fgself-129's last, 11011701, 23:57): a target's army mix (--z-mix 1 --z-mix-reward 0.1): each item's share of the target's army units as an input (the strategy block grows by 64, the new columns at zero), and the strategy potential minus 0.1 times how far the army trained is from that mix (total variation, from 5 army units). Do the armies follow their targets' mixes, and what does it cost? Stopped at 00:40 for a shutdown after 25 minutes (4193900 steps; the army-mix gap 0.62, not yet moving); to go on from there. Resumed 11:38. Result (12:05, ~55 minutes of training): the mix not followed, a little better otherwise. The army-mix gap stayed at 0.62-0.63 over 3,800 games with a target; with a target 74% of 81 against insane (fgself-129 70%), 76% of the training games with one won; 55-57% of 163-178 head to head with its start; without a target 91%.

fgself-131 (from fgself-130's last, 5534555, 11:58, its recipe): the army mix taught: the teacher has a building train the unit the target's army has the largest share of more than its own (from 120 s and 5 army units, by 0.15 at least). At 50 minutes the gap had not moved (0.62-0.64); counted, only 1% of the army-mix lessons taught anything (10% of all lessons): the rest named a unit the policy could not order then (money, tech) and the loss drops those. Restarted 13:09 with lessons only for units it can train then (75% of them kept since).

Result of fgself-131 (14:30, 2 hours 20 minutes): level, the mix not followed. 49% of 292 head to head with its start since the fix, 55% of 361 with fgself-129's; rating 1904 ± 11 (its first hour) to its start's 1941 ± 15; with a target 72% of 64. The army-mix gap stayed at 0.62-0.64 even with the lessons kept: ~4 lessons a game against the hundreds of train orders the policy gives itself.

fgself-132 (from fgself-130's last, 5534555, 14:28; fgself-131's league): a mix order of the agent's own (macros.MIX): a building's order "train the target's mix": the unit the target's army has the largest share of more than the side's own, of those it makes and can pay for now; available while one lacks (with a target that names an army). The teacher's army-mix lesson is that order. As with the expand order, the policy only decides when.

Result of fgself-132 (16:25, 1 hour 50 minutes): worse overall, the mix followed more. The army-mix gap fell for the first time (0.63 to ~0.58); with a target in the real game 25 of 52, then 36 of 53 (its with-target rating 1500, then 1635); but 44% of 408 head to head with its start and its rating 1876 ± 18 (second hour) to its start's 1929 ± 11. As with expansions on echofast, a more varied army costs wins in its short, fast games.

fgself-133 (from fgself-130's last, 5534555, 16:26; fgself-132's recipe with the mix order): halfway rules (--map echo50: three quarters of the hit points and costs, two thirds of the times; the targets people's real-game times scaled by 0.72, --z-pool strategies-echo-real-narrow-tavern.json --z-time-scale 0.72). On echofast tier-1 rushes win, and expanding and varied armies cost games; the real game (the evaluation) has longer games. Real-rules training pulled the policy down months of fixes ago (fgself-98/99); does the halfway now pay?

Result of fgself-133 (17:54, 1 hour 30 minutes): worse. 45% of 312 head to head with its start (50% of 341 with fgself-129's 11011701), its rating 1902-1917 to its start's 1942; against insane 96% (its start 97%), with a target 49% of 104 (from 38% to ~50%). In the real game no more varied without a target (tier 2 in 50% of the games, its start 49%; expansions 1%, 0%); with a target tier 2 in 88%, tier 3 in 45%, an expansion in 14%, 3.2 camps a game. The halfway rules don't pay: echofast stays the training map.

fgself-134 (from fgself-132's last, 9752412, 17:56; fgself-132's recipe and league): items off the ground, night elf expansions, a tavern hero's revival (--macros 2). The harness now reports the items lying on the ground that each player sees (Echo Isles' camps drop one each: a permanent item, a potion or a tome) and has a hero's pick-up (a right-click on the nearest within a radius); the agent's pick-up order is available while a hero has one within 1,000 (and a free slot, but for a tome), the hero kept on it until it holds one, and the teacher shows it when no enemy is near. Night elves expand too: the expand order builds a tree of life at the natural and the agent has it entangle the mine (the policy has no entangle order; night elves had never expanded). A dead tavern hero, lost for good before, is revived at the player's altar with its hire's order. All three checked in the game. Head to head with its start and fgself-130's 5534555 (the best rating).

Result of fgself-134 (19:55, 2 hours): worse head to head, more varied. 41% of 468 head to head with its start and 38% of 446 with fgself-130's 5534555; its second-hour rating 1945 ± to its start's 1967; against insane 99%, with a target ~55-60% (its with-target rating 1547, then 1622, above fgself-132's 1500/1613). It picks up items (0.28 a game without a target, 1.8 with one; a game where it gave the order was won 56-68% head to head against 22-36% of the same length without). A diagnostic (its frozen start under the new code, head to head with the same two checkpoints: 49% and 54% of 123 games) showed the code fair: the drop is learned. By race the loss is human's: 17% of 116 head to head with its start (night elf 46%, orc 44%, undead 53%). Without a target the human policy builds an altar in only 17-24% of its games, so it has no hero (with a target 99% do): it rushes the AI down (74-93%) and loses to policies.

fgself-135 (from fgself-134's last, 9372762, 20:00; its recipe and league): a first hero in games without a target: the teacher's lesson (its altar trains one of the race's usual first heroes, the game's draw; with no altar the worker nearest its hall builds one 450-800 from it, as the target builder places buildings), each when the policy could order it; the tavern's hero then only the second in those games. Head to head with fgself-134's 9208380 (its last but one) and fgself-130's 5534555.

fgself-135 at 1 hour 25 minutes (21:24): level, kept on. 45% of 327 head to head with its start (human 34%, from fgself-134's 17%), 34% of 341 with fgself-130's 5534555; its second-hour rating 1920 to its start's 1898 (fgself-130's 5534555: 1974); against insane 90%, with a target 62%; targets followed 72% (orc 82%, night elf 84%, human 74%, undead 62%). Without a target human now builds an altar and a hero in ~96% of its games. The line since fgself-132 (the mix order, items, a first hero) is ~55-75 points below fgself-130's best: the variety costs wins at first. It trains on for hours (the user: a gain can take a while to show).

Result of fgself-135 (22:00, 2 hours): level to a little better. 47% of 455 head to head with its start, 37% of 484 with fgself-130's 5534555 (by thirds 32%, 38%, 42%: rising); its second-hour rating 1929 to its start's 1898 and its own first hour's 1863. Human without a target now plays a hero; the line stays below fgself-130's best (1974).

fgself-136 (from fgself-135's last, 9440449, 22:01; its recipe and league): games between policies taxed by who is expected to win (--win-balance 0.5, the user's idea), and those games 20% of the launches (--ai-share 0.8, was 0.95). Against the built-in AI a tax cannot hold 50%: the levels sit at 0 (the learner wins ~80% of its training games against the untaxed insane AI), and fgself-85's richer AI still lost 85%. Between policies a target costs a side most of its games (35% won in fgself-134's self-play), and the races' balance (a level per pair of races) does not see it. balance.py fits a logistic model of a game's outcome from both sides' races, whether each plays a target and what it holds (an expansion, tier 3, a tavern hero), a past snapshot, and the tax taken (the newest 3,000 such games, refitted every 25); a game's tax, decided once the targets are drawn, is the share of the favoured side's income that evens it out (at most 0.5, with a little noise so the tax's effect stays measurable). The races' balance until 150 games are in. Head to head with its start and fgself-130's 5534555.

fgself-136 stopped after 10 minutes (513 games) for audit 3's fixes (below), its model fitted on 175 games between agents (log-odds for a side: a target -0.55, a target's expansion -0.36, its tier 3 +0.29, orc +0.56 against the other races' -0.1 to -0.3; the tax's -4.1, so a share of ~0.25 per log-odds unit of edge).

Audit 3 (2026-10-23, a subagent; the user's request): a resume without --win-balance re-inherited the parent's league and overwrote checkpoint 0 (an else taken by the new block); heroes left skill points unspent (51 of 128 agent heroes ended their games with points left; the policy spent a point on 2.8% of the steps it had one) and almost never learned their ultimate (4 of 42 heroes at level 6, the AI 68 of 68), nothing in our code blocking it; the view never forgot a dead building's production (a research masked for the rest of the game in 8 of 400 demonstrations), recycled ids took the enemy's production, and enemy production marked its buildings busy (seen through the fog); undead expansions never made the target's unpl (40 of 123 such games read missed); a keep or a tree of ages at the natural read as a free natural; night elves had no expansion reserve; the mix order counted units no train order makes; scripted camp clears dropped their items for free; filmed games' traces leaked into the episode files (176 MB). All fixed.

fgself-137 (from fgself-136's 657780, 22:10; its recipe and league): audit 3's fixes, and a hero's skill points spent (--macros 3: a point left unspent 10 s is spent as the built-in AI's order for the hero would, its ultimate at 6; herodata.next_skill). Head to head with fgself-135's last (9440449, fgself-136's start) and fgself-130's 5534555.

Result of fgself-137 (23:06, 56 minutes): better. 51% of 198 head to head with its start (ties half: 100 won, 1 tie, 97 lost) and 44% of 291 with fgself-130's 5534555, by thirds 36%, 45%, 51%: the first of the line since fgself-132 to come level with it; against insane 94%, with a target 60%. The win balance held the games between policies at ~42% won (9% ties), a mean tax of 0.13 of the favoured side's income; heroes' skill points spent for the policy 1.3 a game.

Audit 4 (a subagent; the user's request): the win-balance model froze once its 3,000-game window was full; the "% of the AI's record" yardstick had taken duelfast's AI games for Echo Isles runs since fgself-100 (fgself-135: 194% that way, 170% against Echo Isles'); PFSP read past snapshots from taxed games (42-100% won taxed, 8-33% untaxed); league/past let an old, rarely drawn snapshot dominate (0.73 against 0.41); evaluation games counted in a league member; ties counted as losses in head_to_head.py; the food mask ignored the game's 100 ceiling (2 refused orders a step there); a hall's tier order paired the step's later orders with others' resources; a gone unit's harvest memory went to the next unit with its id; a third of the target mixes held units no building trains; an enemy's illusions and its workers' building orders were inputs (the game shows neither); 77 (type, order) pairs no unit can give were allowed; refused orders were never counted; the ratings' anchor pooled three maps and a resume put every game in hour 0. All fixed (deferred: pick-up and mix available per player, not per unit; the expand and creep masks see the fog).

fgself-138 (from fgself-137's 4769091, 23:07; its recipe and league): audit 4's fixes. Head to head with fgself-137's 4769091 and fgself-130's 5534555.

fgself-138 stopped after 28 minutes (44% of 98 head to head with fgself-137's 4769091, 34% of 143 with fgself-130's 5534555, rising): it had started on the races' balance its parent drifted (below), and two bugs the new refusal counts showed. 30-40% of the policy's orders were refused by the game: as human and orc ~1,000-1,500 repair orders a game on what needed none (11-35% of the policy's orders in every run since fgself-100: its no-op), as undead and night elf 200-430 harvest orders at a gold mine their workers can't mine (undead mine their own haunted mines, night elves their entangled ones). Harvest orders now go to a mine the race can mine, and repair orders only to a damaged or unfinished building or machine of the player's.

Audit 5 (a subagent; the user's request): the income tax had been an absolute set of the taxed side's gold after the harness's pending sales (its walk-and-buy purchases came free) and a share of each step's 10-20 rounded down (0.054 took 0.005); under the win balance the races' balance went on moving on the model's games (to 0.88 of undead's income: 49 of 49 such games lost while the fork's model was unfitted) and a fork started on it; PFSP rested on a few untaxed games, read the cost of the learner's target, and weighed a snapshot beaten twice 0 for good; the headline of real_games.py pooled the training launches on the training map with every evaluated policy; --macros 3 was not noted in the vocabulary; enemy heroes' unspent skill points were inputs; a revival was allowed 7 s before the game takes it; summons counted in the army mix; a pending purchase outlived the unit's next order. All fixed; checked fine: the view's per-step forgetting, the 77 pruned pairs, the C++ inference loop against the model (6e-6), the two observation parsers, the automatic skill order.

fgself-139 (from fgself-138's 2277752, 23:38; its recipe and league, its model seeded from fgself-138's 644 games between agents): audit 5's fixes, harvest at the right mines, no refused repairs. Head to head with fgself-137's 4769091 and fgself-130's 5534555.

Stopped at 23:50 for the night after 12 minutes (46% of 34 head to head with fgself-137's 4769091: too early). In the game since (not yet in a run): a worker repairing already, an orc frame under construction (it takes no helpers) and a building at full health take no repair order, and a wisp mining shows "smart", so a harvest re-sent to it (refused: 318 a game as night elf) is no longer sent.

fgself-139, resumed 2026-10-24 (after the night, two reboots for the GPU unlock, and restarts for speed): better. 56% of 293 head to head with fgself-137's 4769091 (by thirds 50%, 56%, 61%) and 40% of 461 with fgself-130's 5534555 (37%, 41%, 43%); its rating 1923-1938 by hour (h10-h12) to fgself-137's 1881 and fgself-130's 1985; against insane 96%, with a target 63%; targets followed 73%. It trains on.

Throughput, 2026-10-24. The RTX 3090's BF16 rate unlocked and a 400 W power limit (GPU unlock) made matrix multiplies 1.4-1.8 times faster, but the learner's update stayed at ~2.75 s: self-play is set by the CPU (the games, then our Python), not the GPU. A profile of the actors (py-spy) found the tech mask reading the game's archives at every game's start (11% of an actor's time), the shared inference buffer's memmap slices (~5%), the teacher's look at every step (~5%), the natural looked at every step (~4%) and the privileged block's masks per side; the run's video renderer ran at nice 0 (~2 of the 32 threads). With those fixed the CPU was 20% idle, and 7 games an actor (was 6) took the slack: ~1,494 agent steps per second before, ~1,660 after (+11%; 8 games an actor: 1,580, the CPU 9% idle). GNOME's file indexer had crawled the games' Wine prefixes (made and removed all the time): ~/wc3 is now ignored.

fgself-139, the afternoon of 2026-10-24: judged. 58% of 490 head to head with fgself-137's 4769091 (by thirds 51%, 60%, 64%): better; 42.5% of 637 with fgself-130's 5534555 (39%, 40%, 48%): not yet as strong, closing. Against insane 97% of 357, with a target 68% of 320. Restarts on the way: the actors' entity features in C (native/w3feat.c, View.step 1,002 → 556 µs a side-step, the same features), Wine off the system D-Bus (each game prefix's device manager asked udisks for every disk ~0.4 times a second: the daemon had taken ~1.5 cores), a frame's helper not sent without lumber (the game refused a human peasant's power build with none, tested in the game: ~275 refused orders a game as human), and 8 games an actor (the CPU 11% idle at 7: ~1,665 → ~1,730 agent steps per second).

Targets: the first hero is not the target's. The policy makes its first hero on time (~10-20 s before it is due) but the target's one only 25-37% of the time, chance among 3-4 (human, Archmage targets: Paladin 355, Mountain King 321, Archmage 320). The teacher taught a first hero only in games without a target; in target games the builder's lesson came once the hero was due, after the policy had made another one, and a second hero needs a tier. Also missed most: the expansion (~44% for every race), undead's Halls of the Dead (half), tavern heroes (~half). --teach 2 teaches the target's first hero from the altar as soon as the altar stands: fgself-140.

fgself-140 (from fgself-139's 16955575, 13:48): --teach 2, the target's first hero taught at the altar; also round 2 of the game's code made faster (--game-cuts 521721), from 14:19 the update compiled before its CUDA graphs are captured (--update-fuse default: an update 2.8 → 1.5 s in the run) and 9 games an actor. Level in strength, better at following targets. 53% of 314 head to head with its start (by thirds 57%, 56%, 46%), 45% of 355 with fgself-130's 5534555; rated 2015 ± 42 and 1992 ± 35 (h0, h1) to its start's 1978 ± 37; against insane 97%, with a target 64% (fgself-139: 68%). The first hero is the target's in 66% of undead games, 61% of human, 50% of night elf (from 25-37%, chance), orc 33%: its lesson rarely fires (the policy orders a hero the step its altar is done). Kept.

fgself-141 (from fgself-140's 12745403, 16:02): the KL anchor moved to the fork's start, fgself-140's last checkpoint (rated ~2000), from fgself-100's 13649982 (1966, from before the six audits, the macros, the tavern and the free first hero): the forward KL had held the policy to the old policy's habits. Same weight (0.05), direction and heads. Also round 3 of the game's code (--game-cuts 67106297: pathing in C where exact, the units pass's flags in the shim, the step sync over a Unix socket: -6.6% of a game's main thread), the noise scale every 50 updates and graphed, 4 games a launch (was 2: the relaunches' churn was ~2.2 cores). Head to head with its start, fgself-130's 5534555 and fgself-108's 2302196 (rated 2086 under the rules of 2026-10-20: the user's idea of a league of strong checkpoints, measured first).

fgself-141, judged at 17:56: worse. 46% of 265 head to head with its start (by thirds 59%, 43%, 37%: falling), 54% of 165 with fgself-130's 5534555; rated 2046 ± 43 and 1982 ± 44 (h0, h1) to its start's 2030 ± 41; with a target 57% (fgself-140: 64%). Holding the policy to its own start (the KL anchor there) hurt as it trained. fgself-108's 2302196 (rated 2086 under the rules of 2026-10-20, the user's league idea) won 47% of 51 against it: not stronger than today's policies. Its throughput work stays: the actors' Python merged at 17:22 (~2,010 → ~2,170 steps/s).

fgself-142 (from fgself-140's 12745403, 18:00; fgself-141's settings and league): half the KL pull (--ref-kl 0.025, the anchor back at fgself-100's 13649982): 0.2 → 0.05 had gained in fgself-77. Head to head with its start and fgself-130's 5534555.

fgself-142, judged at 19:26: better with targets, but countered by fgself-130. 59.5% of 242 head to head with its start (by thirds 49%, 65%, 65%), but 34.5% of 288 with fgself-130's 5534555 (28%, 43%, 33%); rated 1972 ± 32 and 1981 ± 60 (h0, h1) to its start's 1970 ± 28 and fgself-130's 2019 ± 10; against insane 95%, with a target 70% of 240 (fgself-140 64%, fgself-141 57%). The first hero the target's: undead 74%, human 67%, night elf 56%, orc 36%. Half the pull kept. Its games ran longer (10.4 against 8.4 game minutes) and larger (87 against 80 units at most): ~10% more CPU a step in the games.

fgself-143 (from fgself-142's 9858508, 19:31): fgself-130's 5534555 in the league for good (--league-add: a "fixed:" member, drawn by PFSP as the past snapshots are, never replaced by newer ones): the user's idea of a league of strong checkpoints, with the one that beats us. Otherwise fgself-142 (--ref-kl 0.025), and the actors' Python round 3. Head to head with its start and fgself-130.

Audit 7 (2026-10-24, the day's 72 commits). Two bugs that matter:

• Evaluation games read high while the laptop's worker relaunched. It relaunched on any code edit saved here (58 times that day: every 1-10 minutes from 12:17 to 14:06, five times 14:41-14:55, four 15:14-15:17), its ~20 games in flight lost, and a game's row is written only at its end. On the laptop a win takes 0.7 minutes, a loss 2.1, a tie 3.3: the games cut were mostly losses and ties. A simulation on the day's relaunch times read a true 84% as ~89.5% and a true 60% as ~71%, ties counted half. Verdicts against insane from those windows read high: fgself-139's 97% most of all, fgself-140's with-target 64% too; fgself-143's by about a point. Now the worker finishes its games in flight before a relaunch (SIGUSR1: drained), and relaunches only for committed code (it gets git archive HEAD, not the working tree).
• --teach 2's lesson waited until 45 s. Its altar stands at ~33 s and the policy orders a hero then: the lesson fired in 29% of the 4,020 target games it was for. fgself-140 to fgself-143 measured a quarter of it. Fixed: the target's hero as soon as the altar stands (the altar itself still from 45 s, the builder teaches it when due).
• The tavern hero's lesson almost never fired (found after the user saw barely any tavern heroes or expansions). It taught a target's tavern hero only as the first or second hero, from 135 s: 68 lessons in 3,000 target games, ~1,300 of whose targets have a tavern hero (due at a median 7 minutes, mostly as the second or third). By 135 s a first hero was the altar's, and a third was never taught. Hires fell from fgself-120's 55% of the targets with a tavern first hero and 19% with a later one to fgself-143's 13% and 5%. Now taught when the side has as many heroes as come before it in the target, with the tier that count needs (2 for the second, 3 for the third) and the money (the mask's). Without a target the policy never hires a tavern hero nor expands (0% of every run's evaluation games since fgself-130); with one it expands in ~19% of the training games whose target asks it (36% in evaluation), though the expansion lessons fire (~12,000 in 3,000 games).

Smaller ones, fixed:

• The win-balance model's restore at a resume or fork took scenario games (21% of its rows).
• A fork's league trim kept a fixed member as "the first" and dropped a snapshot; fixed members are now not inherited (the fork's own --league-add adds them).
• The run's own evaluation slots played an unpaired launch at a fork's start, before the laptop's first games (64 games of one checkpoint, 0 of the other).
• Two game threads of an actor raced on the material table: a wrong value or an IndexError now and then, a spurious pair of shaping rewards.
• The dashboard left the fixed opponent's games out; the Render list showed the followed side's result; the System tab's "last ~3 minutes" spanned up to 19; the sampler counted other runs' games as this one's.
• The native libraries rebuilt by file time (stale after a copy that keeps times): now by their source's hash.

Two notes:

• fgself-130 is both a fixed league member and a yardstick. The head to head with it can rise by exploiting it, so read the verdict from the ratings and the other checkpoints.
• The label terms (the builder's and the teacher's) dominate the gradient: |G|² 0.053 to the policy's 0.0006 (noise 2,908 to 6.9), ~40% of updates clipped at 1.0. The PPO signal is ~1% of the squared gradient. Also PFSP is near uniform under --win-balance: results are kept only from ~6% of the past snapshots' games (taxed ones tell nothing), so every snapshot weighs about the same.

fgself-143, stopped at 21:20 for audit 7's fixes, not judged (~80 evaluation games a policy since 20:30):

• 52% of 85 head to head with fgself-130 (fgself-142: 34.5%), 48% of 60 with its start.
• Against insane 89%. With a target 49% of 73, falling (fgself-142 69%, fgself-143 before 20:30 58%; human 5 of 22, orc 1 of 11), though its targets are followed as before (56% of the due items, fgself-142 57%).

Audit 8 (2026-10-24 evening, the behaviour of recent games, the user's ask). It started from what the games show, not from the code's changes:

• An expand order held the hall's cost back a second time. The game takes a building's cost at its order (five build orders in filmed games: the full cost on the next step), but the mask kept the hall's cost until the frame stood, for the 25-35 s walk (60 s when the game refused the order). Every paid order was masked meanwhile: a penalty on expanding. Target games kept it too (the builder's reserve while the hall was due). Now nothing is kept, and a refused expand frees the macro at once.

• Duplicate buildings. Each unit samples its own order, so 2-3 workers picked the same building in a step (26 of 202 build choices in filmed games). In short won games, extra one-of buildings (orc 1.78 a game: two war mills in 52%; a second altar in 25-30% of human and night elf games). Now a building of a kind once a step, and never a second altar, upgrade building or shop while one stands, is a frame, or a worker is on its way to build it.

• About half the deaths were never counted. A dead unit is mostly gone from the step that reports its death, and the owner was looked up in that step's table. This hit lost, kills and creeps in every row (deaths ≈ 48% of what was made in long losses), the harass kind's kill count (its penalty overstated) and the builder's bookkeeping. Owners are now kept as last seen.

• Tavern hires were paid at the sale with nothing kept for the ~33 s walk, and two workers could buy the same hero: 10,475 hire orders made 308 hires (old teacher). A hire on its way now keeps its money from the other paid orders, and only one worker hires a hero at a time.

• Orders the tech mask could not see:

  • troll berserkers: the berserker upgrade isn't in the vocabulary, so they can never be trained; ~59 refused a game as orc.
  • abilities a research teaches: footmen's defend (~23 a game), crypt fiends' burrow (~31) and web (~9).

  These are now masked until the research is done.

• Smaller ones:

  • Undead and night elf repairs of full-health buildings (restoration, renew: ~214 refused an undead game) are now filtered like human repairs. Undead and night elf frames take no helpers.
  • The night elf expansion lesson never stopped: wisps in an entangled mine are hidden, so it never saw 4 there, and it fired 22 times in a long game.
  • The laptop's with-target rows carried no version.
  • The food cap above 100 (the game counts farms past its ceiling) was shown in the overlay and used by the teacher's checks: both now clamp at 100.

• Real game behaviour, not bugs: food used up to 112 (the built-in AI does it too); necropolises rebuilt in losses; 4.5% ties at 40 minutes.

• Open: human repairs of frames are refused ~900 times a human game in fgself-143, in the real game too (215), with money at hand. Two probe games found every such repair accepted (from mining, carrying, hidden or idle workers, the builder itself, five helpers, all at once, an abandoned frame, with fgself-143's and fgself-144's game cuts). The cause is still unknown.

• Expansions, the user's question: without a target the policy takes the expand order when the natural is free (38-81% of the scenario games with the camp gone), but never with the camp standing. No lesson teaches clearing the camp without a target, and the policy doesn't creep on its own. With a target that asks for one: 21-57% with the camp standing, 53-89% with it gone.

fgself-144 (from fgself-143's newest, ~21:45): audit 7's fixes, the target's first hero taught as soon as the altar stands (the lesson fgself-140 meant), its tavern hero at its place among the target's heroes, and evaluation without the relaunches' bias; also round 4 of the game's code (--game-cuts 17179866617: the harness's tokens, commands and items in the shim, the name hash in one pass, soft float as SSE where it is IEEE; −3-6% of a game's main thread). Otherwise fgself-143 (fgself-130 fixed in the league, 25% of the past snapshots' games). Head to head with its start and fgself-130.

fgself-144, judged at 23:30: level (on the games of its own code, 21:27-22:21; the laptop then played audit 8's unpushed commit, see below):

• 50.3% of 143 head to head with fgself-130, 48.3% of 117 with its start.
• Against insane 95% of 130 (its start 98.5%), with a target 49% of 133.
• The target's first hero is the policy's first 70% of the time (fgself-140: 50-66%). Tavern lessons fire (443 in 1,769 target games, from 68 in 3,000), hires still ~8% of the targets with one.
• Kept: its lessons are the corrected ones.

Audit 9 (2026-10-24 night, the behaviour of recent games again). It found:

• Ghost units: a unit that leaves the game without dying stayed in the table for good. Cancelled frames (all races, the built-in AI's too) and the wisps an ancient takes keep their last state. The units passes write a removed record only for a unit seen dead or with type 0, and the game's enumeration never yields one removed outright.
  • Every repair at a ghost frame is refused: the ~900 refused repairs a human game, money never the cause (the probes had used real frames).
  • The builder's resume lesson retried them every 20 steps, for good: 83,347 of fgself-144's human builder lessons, ~90%. The policy took up repairing (on 521 of 732 steps in one game; orc target games in the third with the fewest repair attempts won 85%, the others 15% and 6%).
  • A ghost altar read as "being built" to the teacher (no hero in 6-9% of those human games). With audit 8's one-of rule it blocked every altar: no hero in 40% of undead games on the laptop since 22:21.
  • The ghosts come from the policy cancelling its own frames (cancel is allowed on buildings for their queues): undead 9.8 a filmed game, orc 6.7.
  • Stopgap now: a cancel event drops the frame from the table (native/w3obs.c and protocol.merge_observation). Consumed wisps still stay.
• The laptop evaluated with HEAD, not the run's commit: a local commit relaunched it onto other rules than the run trained with. Now it gets run.json's launch commit.

Open, for the next runs:

• the units passes reporting units gone without dying (wisps an ancient took);
• no cancel on the policy's own frames;
• the builder's resume lesson: a retry cap, and not ahead of due items;
• the builder's reserve voiding its own labels: a later affordable item labelled but masked; the supply building should be exempt;
• supply-first by the item's own food, not +2;
• a running budget of the step's money in the agent's orders (several paid orders in one step overspend);
• the revive mask's 60 s (revivals take up to 110 s in the real game);
• full 7-slot queues;
• later research levels priced at level 1;
• War Stomp, Hex, Banish, Far Sight and Scout missing from the vocabulary;
• targeted items used untargeted;
• the tavern lesson's exact hero count (269 of 289 human games with a tavern first hero made an altar hero first);
• the evaluation pool's raw click times + 45 s (tier items due ~95 s before that player had them);
• bookkeeping:
  • z_build's accepted counts and z_refusals are empty under --z-build 2;
  • echofast games against the untaxed AI are labelled "(real)".

Real game behaviour, not bugs:

• With a target the policy plays a long economic game and loses it (human: 20% won, 23 minutes, 35 workers; 90%, 12 minutes without a target).
• Lumber starvation: human 53 lumber a minute and undead 45 (fgself-120: 120 and 71). Lumber is the main blocker of target items (the float penalty may push against saving it: a suspicion).
• The night elf lumber hoard grew from 960 to ~3,200 held on average since fgself-139.
• Heroes cast ~0.5 spells a hero-minute (0.9-2.0 in fgself-120-125).

fgself-145 (from fgself-144's last, 16264963, 2026-10-25 09:28, after a reboot): audits 8 and 9's fixes:

• no money held for an expand on its way;
• one building of a kind a step, one-of buildings once;
• deaths counted after the unit is gone;
• tavern hires keep their money;
• research-gated orders masked;
• cancelled frames gone from the unit table.

Otherwise fgself-144. Head to head with its start and fgself-130.

Fixed since, for the next run:

• the policy's cancels of its own frames are not sent;
• the builder:
  • its resume is capped at 3 tries a frame;
  • an item after a reserved one is labelled only when affordable net of the reserve;
  • supply first by the next item's food;
• a step's money is spent order by order (a running budget net of the kept reserves);
• revivals are not offered again for 120 s;
• orders to full queues (7) are not sent;
• the tavern lesson also fires once an altar hero came first.

People's targets timed by their items' build times (a separate experiment). People's replays give an order's click, and the pools put every item 45 s after it (CLICK_LAG). A keep takes 140 s, so in the real-game evaluation pool tier items came due ~95 s before the player had them, and units ~25 s late. strategy.py now adds each item's own build time (UnitBalance bldtm, UpgradeData timebase; a building's walk 5 s). Both recipes were reproduced exactly with --click-lag fixed before rebuilding. The new pools are strategies-echofast-narrow-tavern-bt.json and strategies-echo-real-narrow-tavern-bt.json. Median due times of people's targets:

item	real-game pool, old	real-game pool, new	echofast pool, old	echofast pool, new
keep	279 s	374 s	298 s	339 s
footman	125 s	100 s	89 s	69 s

• Units gone without dying are now reported. The units passes (shim/units.c and the harness's JASS pass) keep the units each pass met. One the last pass met and this one did not gets an R record: a cancelled frame, or a wisp an ancient took (at the ancient's end).
  • A probe game: the old shim left both as ghosts, the new one removes both, with and without the game cuts.
  • A built-in AI game of 1,000 steps with 15 units removed without a death played back identical with cuts 0 and 17179866617 (raw bytes too).
  • build/shim must be rebuilt at the next fork, with no games running.
• A lumber lesson (--teach 3). When a target's next item is short of lumber (the builder's) and fewer than 5 work lumber (undead: 3 ghouls), the race's chopper nearest the tree nearest its hall harvests that tree.
  • The policy had all but stopped chopping: human 105 lumber a minute in fgself-120, 59 in fgself-129, 50 in fgself-139, 38 in fgself-145, to ~870-930 gold; orc 138 to 54.
  • People gather 30-50% as much lumber as gold. On echofast gold-only tier-1 armies win, and a target's lumber items wait.

fgself-145, judged at 11:08: slightly better, better with targets:

• 55.3% of 255 head to head with its start (by thirds 54%, 53%, 58%), 44.8% of 261 with fgself-130.
• Against insane 93.5% (its start 96.6%), with a target 57.9% of 305 (fgself-144: 49-54%).
• Rated 1989 ± 37 and 2000 ± 45 (h0, h1) to its start's 1977 ± 38.
• The fixes show:
  • refused human repairs fell from ~900 a game to 22 (the ghost frames);
  • tavern heroes are hired in 17% of the target games with one (fgself-144: ~8%);
  • expansions are still ~16% of the target games that ask for one.
• Kept.

fgself-146 (from fgself-145's 13576303, 11:10, build/shim rebuilt): the fixes above (audit 9's follow-ups and units gone without dying reported). Head to head with its start and fgself-130.

fgself-146, stopped at 12:15: invalid. Its step budget dropped every paid order whenever the target's next item was short of lumber: the kept lumber put the budget below zero, and a footman or a worker needing none was refused too. I first read the large skip counts as the builder's reserve holding (a mistake).

• Target games' win score against the AI fell (human 0.74 → 0.57, undead 0.74 → 0.51); real-game human peak food 64 → 38.
• With a target 47% against insane; head to head 45% and 35%.

Audit 10 (2026-10-25, today's code and fgself-146's games). Confirmed and fixed:

• The budget: it now refuses an order only for a resource the order needs.
• The army-mix macro skipped the queue and budget checks. It ordered the same unit until the queue was full, then the game refused it: 160-216 a target game.
• Orders at a hall during its tier upgrade are refused by the game (probe): workers ~1,500 a game at a keep, the tier again ~400-580. They are no longer sent, and the tier is not offered again while under way.
• Research-gated abilities now come from the game's own AbilityFunc tables (the race's units' abilities). They had included cannibalize (23,575 refused), sentinel, ensnare, stone form, autodispel and inner fire.
• A tavern hire on its way did not count as a hero. The altar's free hero was trained during the walk, so in 0 of 1,630 targets with a tavern first hero was the first hero the tavern's. It now counts, and there is one hire at a time.
• Smaller:
  • repairs of damaged buildings need the lumber;
  • no orders to loaded units (peons in burrows);
  • the budget covers the tier at a busy hall, the expand macro and haunted mines;
  • owners are updated on every step with events (summons);
  • neutral wares get no build time in the pools' timing;
  • the units passes keep no entries for a dropped unit (its handle may be reused), and the JASS pass writes no second R record (both paths identical over 900 steps, raw bytes too).

fgself-147 (from fgself-145's 13576303 and its league, 12:17, build/shim rebuilt): audit 9's follow-ups and audit 10's fixes. Head to head with fgself-145's 13576303 and fgself-130.

fgself-147, judged at 13:50: level head to head, better with targets:

• 52.2% of 247 head to head with its start (by thirds 54%, 59%, 47%), 44.4% of 270 with fgself-130.
• Against insane 98% (its start 95%), with a target 67% of 294 (fgself-145: 58%).
• Rated 2016 ± 35 and 1989 ± 48 (h0, h1) to its start's 1982 ± 29; with a target 1639 and 1623 to fgself-145's 1575 and 1543.
• Targets' items made: 62% (from 56-59%). Human with a target is the weak race: 12 of 46 won.
• Kept.

Audit 11 (2026-10-25, the learner and the yardsticks, today's code and fgself-147's games). It found:

• The teacher's lessons carry the whole --teach-coef in any minibatch with one. The loss averages them over themselves (tm.sum()), PPO over every own unit (~6,000). One lesson weighs ~2,500 times a PPO unit's decision. The label terms' |G|² is 0.71 to PPO's 0.0006, and updates clipped at 1.0 rose from 8% (fgself-130) to 61% (fgself-147). Now --label-norm units: lessons averaged over the labelled steps' own units, as the builder's labels; and loss/teach logged apart. An experiment of its own.
• The past snapshots are minutes old. One every 20 updates (~1.3 minutes) and 6 kept: the pool spans ~6.6 minutes, close to self-play. An unplayed snapshot weighed 1.0, 4 times an even one (fixed: the smoothing's prior, a half). Snapshots by wall time: an experiment.
• Fixed in fgself-148:
  • a research queued behind units counts as under way (probe: refused when ordered again);
  • the research levels a need asks for (Requiresamount: the night elf Marks need level 2, so do inner fire, polymorph, bloodlust, healing ward);
  • repairs need some of each resource the building costs (probes), and upgraded types are priced from their upgrade's row;
  • a free first tavern hire keeps no money;
  • cancels only at a building with something queued (a cancel undid a fortress tier in a replay);
  • no train or hero orders at frames;
  • the skip counters in the rows;
  • real_games.py splits on evaluation games only;
  • the dashboard's head to head counts ties as half;
  • the ratings' errors come from the whole information matrix (each player's own games alone had read them about half as wide).
• Smaller, not fixed:
  • the anchor's memory runs from the policy's stored state;
  • a builder label over a lesson keeps the lesson's weight;
  • the AI record's ties.

fgself-148 (from fgself-147's 11480163, 13:52): audit 11's fixes. Head to head with its start and fgself-130.

fgself-148, judged at 15:25: level with its start under the same code; the new agent code cost this lineage's human.

• 48.8% of 257 head to head with its start, 31% of 221 with fgself-130 (fgself-147: 44%).
• Against insane 91%, with a target 56%.
• The ratings place it level with its start under the same code. fgself-147's last checkpoint, evaluated with audit 11's agent code, rated 1888, against 1992-2018 as fgself-147's policy under its own. fgself-148 rated 1883-1901.
• The drop is the human's: 52 of 72 won against insane, from 62 of 64. It came with half the peasants (4 against 8), half the lumber, a keep in 17% of games (48%) and an altar in 81% (97%).
• fgself-130's human did not drop under the same code (66 of 70).
• Corrected at 17:40: not the code. fgself-148's last checkpoint won 73 of 76 human games in fgself-149's evaluation, whose agent code is the same (the commits between them touch the evaluation worker's gauntlet only), and the laptop's step times were alike in both windows (38-45 ms a step). The human dip in fgself-148's window has no cause found; it did not reproduce.
• Kept as the base: it is level with its start under equal conditions.

fgself-149 (from fgself-148's 11692643, 15:28): --label-norm units, the teacher's lessons weighed as the builder's labels. Right away, the gradient norm is 0.23 (fgself-147's median 1.27, 61% clipped at 1.0), loss/opd 0.06 (~0.94) and loss/teach 0.003. Head to head with its start and fgself-130.

The gauntlet's games count for the hourly players they stand for (17:40, the user saw fgself-102 h2 still first and greyed): a run's policy by the hour has no file, so the gauntlet plays that run's checkpoint from the hour. Its games had made a player of their own ("fgself-102/11882922 +shops"). Now the manager writes the link (runs/.gauntlet-aliases.json) and the ratings count those games for the hourly player: fgself-102 h2 went from 2148 to 2101 with 185 games, played today; fgself-149 h0-h2 and fgself-148's last checkpoint lead at ~2105.

The gauntlet (from fgself-149; the user's idea). 4 of the laptop's 20 games play the current policy against the top 5 of the ratings (every game alike, no decay; a run's policy by the hour as that run's checkpoint from the hour): fgself-102's 11882922, fgself-100's 15294056, fgself-109's 4991346, fgself-108's 2302196, fgself-106's 3453342. Ratings from old games, under older rules and against weaker opponents, had topped the table; these games rate them again under today's rules.

fgself-149, judged at 17:45: level head to head, better against fgself-130 and with targets.

• 50.6% of 320 head to head with its start, 49.2% of 295 with fgself-130 (fgself-148 31%, fgself-147 44%).
• Against insane 98.5% (its start 99.1%), with a target 68.4% of 326 (fgself-148 56%).
• Rated ~2105 (level with its start); with a target 1618-1656 (fgself-148: 1562-1574).
• Tavern heroes hired in 18% of the target games with one, targets' items made 60%.
• The gauntlet's first verdicts, the current policy's score against the old top rated:
  • fgself-109's 4991346: 86% of 56;
  • fgself-100's 15294056: 69% of 77;
  • fgself-108's 2302196: 64% of 76;
  • fgself-102's 11882922: 63% of 32;
  • fgself-106's 3453342: 53% of 51.
• --label-norm units kept.

fgself-150 (from fgself-149's 19159029, 17:47): --teach 3, the lumber lesson. Otherwise fgself-149. Head to head with its start and fgself-130. Its evaluation worker follows the gauntlet's list as it is made again.

fgself-150, judged at 19:52: level head to head, better with targets.

• 48.9% of 269 head to head with its start, 49.0% of 196 with fgself-130.
• Against insane 96.2%, with a target 75.5% of 269 (by thirds 64%, 81%, 82%; fgself-149 68%).
• Targets' items made 72% (fgself-149 60%).
• Rated ~2070 (fgself-149 ~2090, within the intervals); with a target up to 1771 (fgself-149 ~1640).
• Lumber a minute in target games: human 32 → 204, orc 77 → 257, undead 29 → 225, night elf 360 → 709, from ~0.6 lessons a game.
• --teach 3 kept.

fgself-151 (from fgself-150's 16198012, 19:55): no macro help (--macros-off mix,creep,expand, the user's ask): the army-mix, creep and expand orders masked in training and evaluation. With them off, the teacher's lessons teach the low-level orders: a unit of the target's mix at a building, the hall built where people put theirs (undead: the mine haunted), workers sent to it, the heroes attack-moving to the natural's camp. Otherwise fgself-150.

fgself-151, judged at 00:25 (it ran 4 hours through a usage pause): kept; it adapted.

• 51.7% of 440 head to head with its start (by thirds 45%, 49%, 58%), 45.8% of 427 with fgself-130.
• Against insane 93.8%, with a target 78.4% of 429 (by thirds 62%, 83%, 90%).
• Expansions without the expand order stay rare: 2.9% of the target games that ask for one (16.5% with it). Its halls are built off the site, plausibly from the point bins between the bases (~206 world units in x).

fgself-152 (from fgself-151's 36116579, 00:27): the -bt pools, people's targets timed by their items' build times.

• Also measured at a person's income (--eval-human 1, rated "+human") and against the hand-written orc (--eval-heuristic 0.1); both are measurement only.
• Otherwise fgself-151.

fgself-152, stopped at 00:45 for a shutdown, not judged (~18 minutes of evaluation, ~35 games a policy):

• Against insane 100%, with a target 97% of 34; 56% of 27 head to head with its start, 62% of 16 with fgself-130.
• At a person's income (current+human): 34% of 32 against insane. The yardstick at equal income had read ~95-100%: the agent slots' doubled income (an agent slot takes the advisor's difficulty, insane) had carried much of its play.
• Last checkpoint 2318909.

Game cuts round 5 is in it from 00:30 (the units pass from the unit objects, --game-cuts 1099511625209, build/shim rebuilt).

From 2026-10-26 no agent slot plays at the doubled income (the user: "make sure our agents dont play with the doubled income, in training and in evals etc."). Every agent slot is at normal difficulty, a person's income (selfplay.agent_difficulty): training, evaluation, the laptop's worker, the render requests and heuristic_match.py (--slot-difficulty normal). Only an advisor (--opd-share) takes --opd-difficulty, and never insane. The episode rows carry slot_difficulty. The ratings keep the doubled-income games against insane apart, against a player of their own ("insane AI, equal income", rated ~1,150 at first). The anchor (1500) is the insane AI against a person's income, so every rating before fgself-153 floats on the head-to-head games until new games link it.

fgself-153 (from fgself-152's 2318909, 10:12): a person's income in training and evaluation. Otherwise fgself-152 (the -bt pools). --eval-human 0: the current policy's games are at a person's income now. Head to head with its start and fgself-130/5534555. Judge it by the real game against insane at a person's income (fgself-152: 34% of 35), with a target, and the curriculum's tax on the AI.

fgself-153, kept (13:49, ~300 games a policy; every game against insane at a person's income): 62.6% of 344 head to head with its start, 56% of 302 with fgself-130/5534555; against insane 39.2% of 300 (its start 33.1%, fgself-130 47.7%: the strongest against insane at a person's income, though fgself-153 beats it head to head); with a target 15.8% (flat). Last checkpoint 38318278.

Audit 12 (fgself-153's games at a person's income; read-only):

1. No late game and no expansion. Wins by length: under 12 minutes 60 of 68, 12-18 minutes 43 of 87, over 18 minutes 8 of 137. It expanded in none of 286 real games (the AI in 81 of the 175 losses, ~20 minutes in); in losses it gathered 462 gold a minute to the AI's 957 and 72 lumber to 399, and its main mine ran dry at ~22 minutes. The cause is the training map: echofast halves costs but keeps the real map's mines and the 500/150 start, so in its short games (wins 5-6 minutes, losses 11-16) the mine never runs out and expanding never pays.
2. Lessons that taught nothing: the army-mix lesson labelled the mix order, masked under --macros-off mix (none of ~2.5 a game kept); the hall lesson fired whatever the money (~94% dropped by the mask at a person's income). Both fixed (the unit's own train order; the hall lesson only when the hall can be ordered).
3. Target games (14%): items made 82% but a median 80 s late, tier 2 ~165 s late, expansions almost never; the targets' builds trained 13-14 workers, the policy ~5, so it cannot pay for the tech and an army. (The audit read the evaluation pool's AI targets as builds at double income; they are the normal-difficulty slots' builds, often from games lost to insane: a pool-quality question, not a bug.)
4. Undead: ghouls as both army and lumber; in target games 3+ slaughterhouses (median 6), 13-20 acolytes past 20 minutes with 5 able to mine.
5. Orc: peons left inside burrows (14 orders a minute skipped in the real game, 1.5 in training), tier 2 in 1% of games.

fgself-154 (from fgself-153's 38318278, 13:49; relaunched ~14:05 after its first launch crashed: I merged the fine-points branch while its learner was starting, and its late import of bc.py met the old model.py): bundled, judged by each change's own measure and head to head (split if worse): --z-comp 0.25 (army composition kinds: a cap on the top unit's share, a basic spam unit forbidden, a minimum of unit types), --spells 1 (War Stomp, Hex, Banish, Far Sight, Scout), research levels' own costs, the audit-12 teacher fixes, game cuts round 6 (--game-cuts 4398046508537, -4.2% of the game's main thread).

Merged, not yet in a run: fine target points (--fine-points 7 --fine-centre 0.9: a sub-cell within the chosen bin, 6.9 world units near the bases and 29 mid-map, was 48 and ~206; the agent found the coarse bins do not explain the off-site halls, which come from the bin chosen).

Audit 13 (today's code and fgself-154's games): fgself-154's relaunch resumed from its step-0 checkpoint, saved without Adam's state, so it went on with a fresh Adam, silently (fixed: the step-0 checkpoint saves it, a resume falls back to the start's, a note when Adam starts fresh); the undead hall lesson never taught (its mine out of sight: the acolyte now walks there first); Cloud was offered without a castle (the game's file gives its Requires twice; both count now); the forbid kind was not followed (0-1% for the spam units). Working as meant: research levels' costs (refusals 0.6 → 0.03 a game), the army-mix lessons (kept 100%, was 0%), the hall lessons (100% for human, orc, night elf), the composition kinds drawn.

fgself-154, not kept (16:30, ~230 games a policy): against insane 48% (its start 47.6% in the same window), with a target 20.5%, head to head with its start 42% of 206 and falling by thirds (52%, 41%, 33%). Audit 14: the no-target play drifted toward the target games' styles (tier 3 in 24% of evaluation games, from 4%; more unit types), most likely from the composition kinds' penalties (forbid never kept: its penalty at its most every time) and the army-mix lesson that now teaches, shared through one network; the fresh Adam and the game shares ruled out. Orc's forbidden "spam" unit was the grunt (it spams headhunters): fixed.

fgself-155 (from fgself-153's 38318278, 16:31): the training map echofast2 (mines and start scaled with the costs; the mine feature reads the map's full mine), --econ-masks 1, --z-workers 1, --spells 1, audit 12-13 fixes, game cuts round 7 (--game-cuts 17592186041849); --teach 3 kept (teach 4's no-target lumber lesson would fire nearly all game) and no composition kinds. Judged against its start in the same window, by race; expansions and wins past 18 minutes.

Process rule since: agent branches are merged only while no run is live (at a fork), since a run executes from the checkout (the 13:49 crash; audit 14).

Audit 15 (fgself-155's games): on echofast2 the main mine runs dry at 12-13 minutes and the policy expands in 2-10% of the training games (none of 207 real games); games past 12 minutes are lost ~95% at every tax level (the tax only changes how many end before the mine runs out). A workers mode asked for more workers than the new caps allow (undead never kept it); the spells started 1/2981 as likely as no order and were never cast (9 casts in ~13,800 games). The tax curriculum, the mine feature's scale and the masks against target items checked fine.

fgself-155, kept (19:42, ~340 games a policy): 54.7% of 331 head to head with its start (rising: 50%, 54%, 61%), 64.4% of 378 with fgself-130; against insane 59.4%, +2.5 points over its start in the same windows with the races weighed alike (head_to_head.py --vs); with a target 22%. Last checkpoint 36007395.

The approximation study (the user's question: how much faster with small differences): thinning the range sensors' timers saved ~7-10% of the game's main thread but cut income by 29-48% (each timer is set to when its event can happen, e.g. a worker reaching the mine), and the stuck units' A* wait was safe but saved ~0-2.5%; at most a few percent without breaking the economy. The user: no approximations for now (the bits stay off; evaluation games drop them always).

Since 19:47 every launch runs from its own checkout of its commit (~/wc3/code/<commit>, warcraftsim/codepin.py): merges no longer reach a live run.

fgself-156 (from fgself-155's 36007395, 19:47, the first pinned launch): --expand-lesson 0.4 (the teacher's expansion lessons in every game once the main mine holds less than 40% of a full one), --spells-bias 3, the workers mode capped. Judged by expansions and the wins past 12-18 minutes, spells cast, and against its start in the same windows.

fgself-156, not kept (21:25, ~170 games a policy; stopped early, the gap clear): 35.7% of 171 head to head with its start (27%, 39%, 40% by thirds); against insane 13.5 points below its start in the same windows with the races weighed alike (human 26% to 58%); with a target 10.7% (its start 22%). No expansions and no spells cast in its real games: the expansion lesson cost strength without teaching expansions.

Audit 16 (bad habits in fgself-155/156's real games): one-unit armies in long games (90% the basic unit even at tier 2; against 4+ enemy air units 1 of 232 won), 1.06 upgrade levels a game to the AI's 3.72, no creeping (0-2 camps a game to 13-20), orc peons left in burrows (200+ orders to loaded peons: 0 of 69 won), human militia stuck (back to work refused ~160 times a game), night elf floating lumber with too many wisps, repair/harvest flips. The hero revival hunt: 48% of hero deaths never revived (94% of those steps without the gold; the policy chose a revival it could pay for 2.4% of the time), the mask's price rounded down (the game rounds to 5) and the revival not paid first. Fixed or behind flags: --saturation-lesson, --econ-masks bits 2-32, --revive-lesson, --burrow-lesson, --militia-lesson, the militia's orders in the forms the game takes; lessons for the long game (--army-lesson, --upgrade-lesson, --creep-lesson) are merged, not yet in a run.

fgself-157 (from fgself-155's 36007395, 21:33): the fixes for clearly bad habits only (--saturation-lesson 1 --econ-masks 63 --revive-lesson 1 --burrow-lesson 1 --militia-lesson 1 and the always-on militia and revival fixes), no strategic lessons; those follow one at a time.

fgself-157, not kept (22:15, ~40 minutes, stopped: the gap clear): against insane 37% where its start played 70% in the same windows, under the same code and masks (-28 points with the races weighed alike); 36.8% of 72 head to head. The masks were not the cause (the start checkpoint plays 70% under them); the new lessons were: revival 2.3 a game, burrows 1.3, lumber 1.2 (0.15 before the oversaturation trigger) raised the teacher's loss ninefold at the start (0.054 to 0.006), KL 0.11-0.17 in the first updates, the distance to the anchor 0.15 (0.08). Lessons that fire this often pull the policy off what wins before it learns their point.

fgself-158 (from fgself-155's 36007395, 22:20): the masks and order fixes only (--econ-masks 63, the militia's order forms, the revival's price and payment), no new lessons; then the lessons one at a time, each at a lower weight or rate.

fgself-158, stopped at 23:40 for a shutdown, not judged (~80 minutes, ~130 games a policy; last checkpoint 14858006): against insane 67.5% of 140, rising by thirds (59%, 64%, 80%); its start 72% in the same windows (-6.3 points with the races weighed alike); 49.6% of 122 head to head with its start (46%, 44%, 59%); with a target ~13%. The masks and order fixes alone lifted fgself-155's own checkpoint from ~59% against insane (under its own run's code) to 70-78% (under fgself-157's and fgself-158's).

Replay data (2026-10-26): only 1.29 replays play back in our client (1.26/1.28 fall out of step in 3-6 s even with the ids aligned, 1.30 at ~10 s; 1.31 and Reforged differ more): other versions give build orders only. Downloaded for build orders: 1.26 4,400, 1.27 331, 1.28 4,030, 1.30 4,004 (1v1s; the 1.26 download stopped part way). 1.29 on other stock maps converts in step: 192 of ~608 converted (runs/fullgame/demos-human-129-maps, includes Echo Isles again) when stopped for the shutdown. Details: human replays.

fgself-158, kept (11:07 on 2026-10-27, ~330 games a policy, resumed after a shutdown): 54.3% of 324 head to head with its start (49%, 53%, 61% by thirds); against insane 71.5% to its start's 73-74% in the same windows (-2 points with the races weighed alike, interval -8 to +5); with a target 26%. The first behaviour scorecard (scripts/scorecard.py): workers trained +1.0, orders to loaded peons -12 (orc's burrows), the top unit's share +0.02 (0.93); unchanged and still poor: no expansions, ~0.7 weapon/armour levels, ~1 creep camp, human anti-air 6% of its army against air. Undead weaker against insane than its start (69% to 87%).

fgself-159 (from fgself-158's last, 11:10): the oversaturation lesson alone (--saturation-lesson 1), one lesson at a time.

Audit 17 found why one new lesson made a fork jump: the labels kept training through the 10-update value warm-up, which has no policy gradient, so at a fork's start only the labels and the anchor moved the policy (its labels' gradient at the start weights 0.13 in fgself-159, 0.12 in 157, 0.00 in 158); and a label weighed 9-140x a PPO decision (--label-norm units divides by the labelled steps' units only). Fixed: the labels' weight is 0 in the warm-up and rises to full over --teach-ramp 100 updates after it. Also fixed: the revive lesson's tavern-hero labels widened the allowed table (altars hiring), the militia and burrow lessons used enemies the policy cannot see, a fork's eval list could name an older checkpoint than its start (--set eval-checkpoints=start ... now), resumes record their code. Undead's drop in fgself-158 (69% to its start's 87% against insane) is learned, not a mask: turtling (games 17.9 to 13.3 minutes, 2.5 spirit towers a game) and wasted orders (acolytes to a full haunted mine, Death Coil on own full-health units, orders to burrowed crypt fiends): masks for them in work.

fgself-159, stopped unjudged (11:35): its oversaturation lesson moved the policy in the warm-up (KL 0.07-0.16); against insane 33% of 44, ~36 points below its start. fgself-160 (11:36): the same lesson alone from the same start (fgself-158's 33539153, its league), under the label fix.

fgself-160, not kept (13:20, ~170 games a policy): with the labels off in the warm-up its updates stayed small (KL 0.0001-0.0015 in the first 14), yet the oversaturation lesson still cost strength: against insane -17.6 points below its start with the races weighed alike (human 17% to 50%, orc 37% to 81%), 40.9% of 192 head to head. The lesson itself hurts (a label still weighs 9-140x a PPO decision); oversaturation is to be handled by a mask instead (no gold harvest to a mine with five workers), as every mask so far helped and every new lesson hurt.

The human anchor (13:25): bc.py --like-run runs/fgself-158 on 554 converted 1.29 replays on six stock maps (runs/fullgame/demos-human-129-maps, 520 training games, 35 held out), 12 epochs, into runs/bc/human-anchor-1/: the first clone of people's play shaped like the RL runs' network, to become their KL anchor.

The human anchor, runs/bc/human-anchor-2 (15:35): cloned (bc.py --like-run runs/fgself-158) from 575 converted 1.29 replays on six stock maps (539 training games, 36 held out), 12 epochs in 1.9 hours, after audit 18's fixes (a point clipped to the frame's edge left out; the research levels' costs in the encoder): held-out order accuracy 0.747 (0.718 after the first epoch), target 0.672, points 9.1 bins off at the median, its value poor (explained variance -0.14; not used as an anchor). A first run (human-anchor-1) was stopped at epoch 3 for the fixes.

fgself-161 (from fgself-158's 33539153, 15:40): the human clone as the KL anchor (--init runs/bc/human-anchor-2/policy.pt, forward KL over every head at --ref-kl 0.025), in place of the built-in AI's clone (fgself-100/13649982), and every mask on (--econ-masks 127). The clone plays in the evaluation games too. Judged by the scorecard (expansions, upgrades, creeps, army mix, anti-air), head to head, and the KL to the new anchor.

fgself-161, stopped (16:25, ~45 minutes): the human clone as the anchor pulled the policy down (against insane 64%, 36%, 25% by thirds; 31 points below its start with the races weighed alike; training games against the AI won 2%). The clone alone won 0 of 68 against insane and lost 52 of 52 to fgself-161: it barely plays (median 5 workers, no army, 17 food, 157 orders a game, games over at ~8 minutes). Cloning's familiar failure: matching people's orders step by step (74.7% held out) does not make a player once its small errors lead it where people never were. A clone is to be checked in real games before it anchors anything.

human-ft-1 (16:50): fgself-158's last fine-tuned on the human games instead (2 epochs, lr 3e-5; held-out order accuracy 0.703, target 0.562), to keep the RL policy's play and take on people's habits. fgself-162 (16:52, from fgself-158's 33539153, the old anchor): every mask on (--econ-masks 127), with human-ft-1 among its evaluation checkpoints so its own real-game record is measured first.

The human fine-tunes failed for a mechanical reason (2026-10-27/28): human-ft-1 (0 of 92 against insane) and human-act-1 (--acting-only, 0 of 393) were blind to their strategy inputs: the cloning zeroed the init's trained weights for the strategy block's modes and mix parts (zero_z_extras; norm 3.8 to 0), the largest weight change seen. And the order rate printed ignored each unit type's table: as played fgself-158 orders in 12-35% of a unit's steps (not 93%), and human-act-1 ordered 2.4 times as often as its start. Fixed (--z-extras auto keeps the init's weights; the rate "as played"), and --hold-start β (a KL to the start on each unit's act/no-act), --hold-all, --train-heads added. An agent's CPU test (11 GB) beside the live run ran the machine out of memory at 18:30 on 2026-10-27 (everything stopped for 5 hours): agents' jobs beside a run are capped now.

The win kept alive (the user's observation): night elf left undead's last mine alone ~170 s before ending a training game while it finished overdue target items: the strategy potential is not returned at the end (so following the target pays), so target items made before the kill paid (+0.02 each) for little cost (a win 3 game minutes later is worth ~10% less); target wins last ~2 game minutes longer at the median (9.1 to 7.1 minutes). --z-decided D freezes a won game's strategy potential at the moment the opponent's material fell below D of the side's.

fgself-162, level (00:30 on 2026-10-28, ~490 games a policy; resumed after the out-of-memory stop): against insane 69% to its start's 67% in the same windows (+3.3 points with the races weighed alike), but 45.3% of 337 head to head with its start; scorecard: workers +0.8, weapon/armour +0.2, creeps -0.4; undead still weak (34%). The masks act at play: the next fork goes from fgself-158's last again, with them.

human-hold-1 (00:35): fgself-158's last fine-tuned on people's order choices with --acting-only --hold-start 100 and its strategy weights kept; next fgself-163 (--econ-masks 127 --z-decided 0.15) with it among the evaluation checkpoints.

Stopped for a shutdown (00:52 on 2026-10-28): no run live (fgself-162 judged level and stopped); human-hold-1's first try ran the GPU out of memory in its first validation (4 whole sides a batch beside the start network; epoch 1 had reached 68% of people's order choices with the order rate as played 0.229 to the start's 0.270): fixed (2 sides a batch with --hold-*), its rerun stopped for the shutdown. Next: rerun human-hold-1, then fgself-163 from fgself-158's 33539153 with --econ-masks 127 --z-decided 0.15 and human-hold-1 among its evaluation checkpoints; if it plays near fgself-158 against insane, a fork with it as the anchor (--ref-kl-mode forward-acting).

human-hold-1 (10:35 on 2026-10-28): fgself-158's last fine-tuned on people's order choices (--acting-only --hold-start 100, its strategy weights kept, 2 epochs): order accuracy given one 0.02 to 0.676, target 0.22 to 0.548, the act rate as played held (0.270 to 0.270); its value spoiled (explained variance -0.71; no matter for evaluation or an anchor). fgself-163 (10:37, from fgself-158's 33539153 and its league): --econ-masks 127 --z-decided 0.15, human-hold-1 among its evaluation checkpoints.

fgself-163, kept (17:07 on 2026-10-28, ~236 games a policy; a machine reset at 11:13, a CPU machine check, cost it its newest checkpoint, now synced to disk before renaming): 54% of 213 head to head with its start; against insane 62% to its start's 67% in the same windows (-4.7 points, races weighed alike), all from orc (53% to 84%: its orders to peons inside burrows rose from 14 to 95 a game, gold a minute 561 to 516 with more peons); human up (62% to 47%), workers +1.2, lumber +12 a minute, creeps -0.4. human-hold-1 (people's order choices, the act rate held) won 16 of 246 against insane: ruled out as an anchor; copying people's per-unit order choices makes play worse even when the rate holds.

fgself-164 (17:10, from fgself-163's last): --econ-masks 255, adding bit 128 (orc's burrows: battle stations only with a seen enemy near, the agent stands burrows down after 20 safe steps).

Audit 19 (orc's slide, 2026-10-28 evening): no mask or code bug found; fgself-163's orc changed in training. Without a target against insane it fell from 84% to 53% (p = 0.0003), its wins by 12 minutes from 41% to 15%: it lost its early headhunter push and grew greedy (peons 8.1 to 11.9 while gold a minute fell 554 to 507, its barracks at 1.4 minutes instead of 0.8). Burrow orders rose 20 to 103 a game; bit 128 cuts them, but orc still won only 35-42%, and bit 128 masked battle stations only, not a peon's own board order (852043). Suspects: --z-decided 0.15 and orc's falling win-balance tax. So fgself-163's "kept" was wrong for orc: the head-to-head gain hid a collapse of one race.

fgself-164, stopped (19:00, ~99 games a policy): level with its start (61% to 59% against insane, orc 42% to 35%, upgrades +0.3, lumber −15 a minute). Bit 128 now also masks a peon's board order unless a seen enemy is within 1500 of a hall or burrow.

fgself-165 (19:00, from fgself-158's 33539153 with fgself-164's league): --econ-masks 255 without --z-decided, the first half of audit 19's ablation. Judged on orc's wins by 16 minutes (its start: 68%), peons (~8) and barracks time, and head to head with its start and with fgself-163's last. Next: the same with --z-decided 0.15.

Stopped for a shutdown (00:06 on 2026-10-29): fgself-165 at 57.3M steps (checkpoint 57296548, loads), level so far: ~150 games a policy against insane 69% to its start's 67% in the same windows (+1.4 points, races weighed alike), orc 50% to 54% (28 games each), human 81% to 68%, wins past 18 minutes 0.34 to 0.48; orders to loaded units still ~50 a game for orc, so the board mask has not cut them. Next: resume fgself-165 (--commit launch) to ~250 games a policy and judge it; then the --z-decided 0.15 half.

Resumed (09:36 on 2026-10-29, after a clean shutdown): fgself-165 from 57296548. The ~50 orders a game to orc's loaded peons are not a hole in bit 128: the evaluation plays the start under the run's own masks too (50 there; 95 under fgself-163's masks 127). They are orders the policy picks for peons sheltered in a real fight, not sent (the bot skips them), so they waste a pick and nothing in the game.

fgself-165, level (12:00 on 2026-10-29, 262 games against insane to its start's 260): 68% to 69% (−1.5 points, races weighed alike), 55% of 211 head to head with its start. Orc slid again, without --z-decided: 55% to 64%, its peons 8.4 to 11.4; wins past 18 minutes 0.31 to 0.49 (night elf 0.44 to 1.00), anti-air share against 4+ air down 8 points; gold +14 a minute, workers +0.9. So --z-decided did not cause orc's greed (fgself-163 and fgself-165 drift alike); the --z-decided 0.15 half is not run, and the flag stays for its fix of the farmed targets. The drift comes from training itself; candidates: the league's agent games, the curriculum's stronger AI in training, or the other races' updates through shared weights (orc 15% of the learner's games). Not the win-balance tax: the status line's "taxes between races" are the old per-pair balance, unused with --win-balance; the fitted model (3,000 games) taxes orc ~9% against human and its night elf and undead opponents 29% and 11%. In fgself-165's training games orc won 63-65% with 8 peons or fewer against the AI and 42-43% with 11 or more, but against past snapshots 40-43% and 50-53% (whole-game counts, so longer games weigh in).

Next: late-game starts (12:40; the user: improve diversity, playstyle, then win rate and rating; roadmap item 2 in five changes). The clash, raid and rich scenarios were level (fgself-114), so a share of training games will start from mid-game states (minute 8-16) captured from the agent's own games: tier 2-3 armies, mixed units, anti-air and upgrades, where it is weakest (wins past 18 minutes 31%) and where people's build orders lead. Being built (harness ops for research levels and hall tiers, a late scenario, capture to late-states.jsonl); fgself-165 trains on meanwhile as a longer run.

Late-game starts merged (harness ops 74-78: a unit placed as it was, at its hit points and mana, a hero with its level, experience, skills, items, dead or alive; research levels; mine gold; a unit removed; fullgame/late.py, --late-capture, --late-minutes, --scenarios ...,late, --late-states). A real-game check on echofast2 (four states from insane-vs-insane games, all races) matched unit counts by type, hall tiers, heroes, research and creeps. Not kept: trees cut, items on the ground, cooldowns and buffs, queues and research under way, summons, the clock. The built-in AI cannot take over a placed base (its script restarts its build at minute 0: 7,500 gold floated by 5 minutes), so only games between agents start late. Echo Isles is not symmetric enough to mirror a state (43% of its trees match), so a state is used only with the same start locations.

fgself-166 (13:08, from fgself-158's 33539153 with fgself-165's league; fgself-165 stopped at 97.7M steps, its control): late-game starts. 30% of training games capture their state at a minute in 8-16, and 60% of the games between agents (12% of all) start from a captured state (--scenarios natural,late,late,late --scenario-share 0.8: natural stays at 20%). Judged against fgself-165 at equal games: against insane by race, wins past 18 minutes, the top unit's share, anti-air, with a target, head to head.

fgself-166, kept (18:45, 247 games a policy, 2,411 states captured, ~8,000 late starts): against insane 74% to its start's 69% in the same windows (+6.6 points, races weighed alike, 95% interval −1 to +14) and fgself-165's last 69%: the best at a person's income. From undead (71% to 46%) and human (71% to 55%); night elf level (89% to 90%); orc slid a third time (61% to 75%). Head to head 56% of 136 with its start, 47% of 246 with fgself-165's last (trained 60% longer); with a target 28% (fgself-165: 30%). The late game itself did not change: wins past 18 minutes 30% to 31%, the top unit's share 0.93 to 0.91, anti-air against 4+ air 0.26 to 0.33, upgrades −0.16, creeps −0.3; workers +2, gold +18 and lumber +11 a minute. Its states come from its own one-unit games, so they teach more of the same. Next: late starts from people's games (the 95 Echo Isles replays played back), which have mixed armies, expansions, tier 3 and air; fgself-166 trains on meanwhile.

People's late-game states (merged 19:05): 262 states from the 95 Echo Isles replays' playbacks (minutes 8, 10, 12, 14, 16: 78, 64, 49, 41, 30; 15 of 16 race pairs at their starts, night elf mirror none), taken from the demo files (the engine's observations every 0.5 s), scripts/late_from_demos.py to runs/fullgame/late-states-human-129.jsonl. echofast2's layout files are byte for byte Echo Isles'; only mine gold and starting resources are halved, so late.rescaled halves mines' gold and each side's gold and lumber. Per side against fgself-166's learner: top unit 58% of the army to 82%, expanded 34% to 20% (55% at minute 16), 2.0 heroes to 1.3 (tavern heroes, items), air in 12% to 21%, research levels 1.8 to 5.1, tiers 1/2/3 9/52/39% to 14/50/36%. Three real-game checks on echofast2 matched at step 1 (orc against human at minute 16: 39 and 61 units by type, tier 3 both, both expansions, 10 of 10 research, heroes' items, resources, creeps). --late-states PATH:W gives a file a share W of the late starts.

fgself-167 (19:10, from fgself-166's last, 65304585, its league): half of the late starts from people's states (late-states-human-129.jsonl:0.5), the other half from fgself-166's and its own captures. Evaluation checkpoints: its start, fgself-158's 33539153, fgself-130's.

Stopped for a shutdown (23:55 on 2026-10-29): fgself-167 at 52.6M steps (checkpoint 52554108, loads), ~5,980 late starts, half of them people's. Interim at 174 games a policy: level against insane (68% to its start's 69%, the same windows; races weighed alike −0.8 at 148), head to head 49% of 107 with its start. Its play moved the intended way: the top unit's share 0.91 to 0.95 (−0.04, the interval excludes 0), weapon and armour levels +0.31, creeps +0.4, lumber +11 a minute, wins past 18 minutes 44% to 39% (not significant), still no expansions. With a target 19% to fgself-166's 28% (late starts have no target, so fewer target games). Next: resume fgself-167 to ~250 games a policy and judge it.

The games box (14:15 on 2026-10-30): tora-ms-a2, a Minisforum MS-A2, set up and playing fgself-167's evaluation games with inference on its CPU, ~1,085 games an hour at first (the laptop ~790); measurements in scaling up the CPUs. fgself-167 resumed at 14:03 from 52554108. remote_eval.py compared the host's row count with the whole local file, so a second host's rows waited until their count passed the first's, and that many were then lost: now it counts its own host's rows (none lost here, caught at 160 rows).

Stopped to move the box (14:21): fgself-167 at 55.3M steps (checkpoint 55349539, loads). At 222 games a policy, level with its start against insane: 70% to 71% (+0.1 points, races weighed alike); head to head 48% of 139; with a target 22%. The top unit's share 0.91 to 0.94 (the interval excludes 0), anti-air against 4+ air 0.26 to 0.31 (also excludes 0), wins past 18 minutes 45% to 35% (not significant), no expansions. Next: resume it to ~250 games and judge it.

fgself-167, level (15:17, 285 games a policy; half of its late starts from people's games): against insane 68% to its start's 72% (−2.1 points, races weighed alike), human 72% to 77%, orc 60% to 67%, night elf 99% to 93%, undead 53% to 55%; head to head 47% of 178; with a target 24%. Its play improved a little: weapon and armour levels +0.24, creeps +0.3, the top unit's share 0.94 to 0.91, fewer refused orders; anti-air against 4+ air 0.31 to 0.26, and still no expansions, so people's states alone don't teach taking one.

fgself-168 (15:18, from fgself-167's last and recipe): the first real run of remote training games: 42 games on the games box (--remote-games 42 --remote-bind 10.77.0.1, remoteactor --actors 6 --games-per-actor 7) beside the 72 here, over the direct cable; evaluation games on the laptop. Measures steps a second, whether the learner keeps up, and the box's load. (Later the same day, by restarts of the same run: 56 games there with same-page merging, four gateways: ~4,625 steps/s against ~3,280 without the box; the 3090 is now the limit.)

fgself-168, kept (19:30, 252 games a policy, 61M steps in 4 hours): against insane 70% to its start's 64% (+7.2 points, races weighed alike, 95% interval −1 to +15), night elf 91% to 79%, orc 59% to 50%, undead 45% to 30%, human 77% to 86%; head to head 58.5% of 205 with its start (thirds 48%, 55%, 72%), 54% of 178 with fgself-166's last (44%, 58%, 60%), though fgself-166 still scores more against insane (74% in the same windows); with a target 30%. The strongest head to head so far, its play unchanged: no expansions, the top unit's share 0.92, anti-air against 4+ air 0.46 to 0.43, wins past 18 minutes 31% to 29%. People's states put the policy in mixed positions, and with no target there it soon goes back to its one unit. Next: people's late starts with a target, the plan that person went on to play.

Found while building it: the box played no late starts in fgself-168 (0 of 14,413 games, 25% here): the state files were not there, and a late draw with no state fits is a game from the start. So fgself-168's gain came with far fewer late starts than fgself-167's. And since fgself-166 a late start's drawn target stayed in the policy's input, unpaid, and in the balance's features; --late-z 1 fixes both.

fgself-169 (19:53, from fgself-168's last, 67203721): --late-z 1: in a late start from a person's state each side's target is what that person went on to make (late.plan_target, runs/fullgame/late-states-human-129-z.jsonl; items a side holds are left out, times count from the state, nothing due before its build time + 5 s). Plans a side: human 6.5 items, top unit 58%, 10% with an expansion; orc 4.8, 44%, 28%; undead 5.3, 53%, 34%; night elf 4.5, 61%, 35%. About half are the losing player's. The box plays late starts now (RUN=... scripts/box_games.sh sends the states).

fgself-169, not kept as the base (01:09 on 2026-10-31, 278 games a policy, 70M steps; stopped for a shutdown): with a target against insane 38% (fgself-168 ~30% in its own windows, so the change did what it aimed at), without one 72% to its start's 74% (−2.0 points, races weighed alike; night elf 85% to 95%, undead 61% to 55%, human 66% to 69%, orc 73% to 74%), but head to head 42% of 243 with its start (thirds 38%, 49%, 39%), and its play without a target is unchanged at 235 games (the top unit's share 0.94, no expansions, wins past 18 minutes 30% to 33%). It had restarts and ~18 minutes of --infer-past-period-ms tests, none that should change the verdict. The base stays fgself-168/67203721. About half of the plans are the losing player's: next, the plans of the winners only, from that base, judged with a target as well as without.

fgself-170 (10:26 on 2026-10-31, from fgself-168's 67203721 and its league): --late-z 2, only the winner's plan as a target in a person's late start (the other side none); checkpoints and snapshots every 40 updates. fgself-169 is the comparison.

Where fgself-168 loses (2026-10-31, its 271 evaluation games against insane on the real map, wins only counted): before 18 minutes it wins about 90% (179 games), past 18 minutes 26% (92 games), past 25 minutes 6% (36 games), so about 79% of its losses come past 18 minutes. Every game's most-trained unit is one tier-1 type for nearly every side: footmen 80 of 80, archers 70 of 70, headhunters 61 of 63, ghouls 58 of 58. It won none of the 39 games where the AI trained 4+ air units (39 of its 86 losses), though long games and air go together, so air is not shown to be the cause alone. Exposure is not what is missing: 27% of its training games against the AI held 4+ enemy air units. By matchup: undead against orc and night elf 19% each, orc against night elf 18%, night elf 91% overall. The updates move the policy little: KL per update 0.0007, clip fraction 0.6%, lr 3e-5, one epoch. No logged run tries --ref-kl 0 or trains against --heuristic-share; --lr 6e-5 was last tried on 2026-10-14 (fgself-42, worse).

Plan from this: fgself-170 stops at ~150 games a policy if still under 45% head to head with its start. Then one change a run, evaluating only the current policy, with a target, and the start (about 40% fewer evaluation games for a verdict): --ref-kl 0, then --lr 6e-5, then scripted opponents with air armies in training.

fgself-170, level, not kept (16:24, ~250 games a policy): against insane 68% to its start's 65% in the same windows (+2.5 points with the races weighed alike, 95% interval -5 to +10), human 73/67, night elf 89/99, orc 72/62, undead 36/32; 46% of 257 head to head with its start, 51% of 156 with fgself-166. With a target 34%, below fgself-169's 38%. Behaviour unchanged: expansions 0, top unit 0.92, anti-air 0.43, wins past 18 minutes 22% to 34% (not significant); only workers rise, +1.0 (significant). In the same windows fgself-166 keeps 74%. People's plans as targets have now had two runs with no gain in play; dropped. The base stays fgself-168/67203721.

fgself-171 (16:27 on 2026-10-31, from fgself-168's 67203721 and its league): the KL anchor off (--ref-kl 0), the one change against fgself-168's recipe (--late-z back to 0). The anchor is fgself-100, a clone of the built-in AI; 0.2 to 0.05 gained (fgself-77), 0.05 to 0.025 held level (fgself-142), and no logged run had it off. It also takes a network pass out of each update. It evaluates only the current policy, with a target and the start, so a verdict needs fewer games; the box plays 56 games as before (RUN=fgself-171 scripts/box_games.sh caeb6bcea2e7 8 7). Stops early if it falls ~15 points under its start against insane, a race collapses, or it is under 45% head to head at ~150 games.

The stop line moved (17:45): at 20.4M steps and ~175 games a policy fgself-171 is 43% of 256 head to head and -7.7 points against insane (orc 47/68, undead 24/45), under that line. It stays on, because the line was set in games and the smaller evaluation set doubled the games an hour: 150 games used to mean ~50M steps of training, here ~17M. fgself-170 read the same at this age: -9.1 points at 14.5M steps (fgself-171 -9.4 at 14.9M), then +5.4 at 53M. So a fork's first hour reads low, and verdicts need a floor in steps as well as games. New line: judged at ~50M steps, on the games from its second hour on. So far the anchor's removal frees no new play: top unit 0.91 to 0.94 (significant), expansions 0, upgrades -0.33 and income -15 gold and -9 lumber a minute (all significant). The speed gain is +5% (4,497 steps/s to fgself-170's 4,285 at 350 W, from checkpoint times).

fgself-171, worse, not kept (19:30, 49M steps, ~420 games a policy): 37% of 601 head to head with its start, falling by thirds (43%, 39%, 29%); against insane 58% to its start's 69% in the same windows (-10 points with the races weighed alike, 95% interval -17 to -4 from its second hour on). Human 67/68, night elf 94/97, orc 52/68, undead 25/46; with a target 26%. It also fell against the scripted orc: 69%, where fgself-170 had 82%. From its second hour: income -21 gold and -16 lumber a minute, upgrades -0.21, anti-air 0.26 to 0.31, refused orders 4% to 2% (all significant). Its top unit fell, 0.90 to 0.94, orc most (0.72 to 0.86), and night elf went to 2.1 heroes, but expansions stayed at 1%. So the anchor is holding the macro up: without it the policy drifts faster (KL per update 0.0012 to 0.0008) and loses its economy before it gains any new plan. --ref-kl 0.025 stays. The fork's dip does not explain this: unlike fgself-170 it did not recover by 50M steps. A second look at its play (2026-11-01, the user saw more varied replays; games with 10+ army units against insane, first half to second half, ~50 games a race each, the start's ~105 in the same windows): the variety is real for two races, more than the top-unit share showed. Orc: a second unit kind at 15%+ of the army in 70%, then 89%, of games to 32%; grunts the top unit in 31% to 2%; a tier-2 hall 50% to 16%. Night elf: tier 3 in 81% to 4%, 2.3 heroes to 1.0. Human and undead did not gain: still footmen and ghouls in every game, human's tier 2 fell (2% to 47%), and human's wins fell (83%, then 41%). So it was weaker (a big sample, falling) and more varied in part, and --ref-kl 0 remains not kept as the base.

Why the long games are lost: income (2026-10-31, 522 evaluation games without a target and 529 with one against insane on the real map, fgself-168 and fgself-170). A mine takes five gatherers, so one mine caps income: the policy's gold a minute stays ~500 whatever its workers, while the AI gathers about twice that (its double income). Without a target it never expanded in these 522 games, so its gold stops at the mine's 13,500:

Game length	Games	Won	Its gold	Share at 11,000+	AI's gold
under 12 min	140	99%	5,600	0%	7,800
12-15	144	92%	7,400	0%	11,400
15-18	59	75%	8,900	0%	14,600
18-21	60	38%	10,200	22%	18,600
21-25	44	20%	11,700	75%	22,600
25-30	37	5%	13,100	97%	25,300
30+	38	3%	13,100	97%	30,400

So its one way to win is to end the game by ~15 minutes, and massing one tier-1 unit is a fitting answer: tier 2 and tech spend half the AI's income and delay the attack. This would explain why many runs moved neither the top unit's share nor the wins past 18 minutes, and why with a target it wins about half as often (34% to 68%): those games have 16 workers to 9 and a tier-2 hall in nearly every game, yet run long (69% past 18 minutes, against 34%) on the same income. Expanding as played does not rescue it. With a target, on the real map, a first new hall at 10-16 minutes won 22% of 55, at 16-22 minutes 8% of 48, none 34% of 415; on echofast2 by 10 minutes 18% of 2,535, none 48% of 10,404. These are correlations: losing sides also expand late, and the targets differ. On echofast2 the main mine holds 6,750 and games past 12 minutes win 2-13%, against 81% before, the same wall at half the time. This puts anti-air second: 27% of training games already hold 4+ air, and air comes in the games already behind on income. The scripted air opponents drop down the list; two-base play is the thing to test.

fgself-172, stopped, no verdict (19:54, 23 minutes, 6M steps): --lr 6e-5 from fgself-168's 67203721, the anchor at 0.025; 37% of 55 head to head and -9 points against insane at this age, the dip every fork shows. Stopped for the higher-priority test below, so the learning rate stays untested at a person's income.

The scenario base2 (fullgame/scenarios.py, built and probed by a subagent): training games start with the natural running: its camp cleared, the race's hall, 5 workers gathering. Against the built-in AI only the agent's side, between agents both. Setup goes through the late starts' machinery, so the first reward holds no gift and nothing placed counts as made (the gift stays in the shaping's potential and is paid back at the game's end, about -0.19 of reward against the AI, a constant). Never in evaluation. Real-game probe on echofast2 over 3 game minutes without policy orders, gold without against with: human and orc 1,780 to ~3,500, undead (3 acolytes on the main) 1,070 to ~2,850, night elf 1,790 to ~3,590. Against the insane AI, human and orc gathered about what the AI did (3,470 to 3,460-3,500). Its games stay out of the win balance and the curriculum as other scenarios do, and out of the captured late states. Decided in the build: undead haunts the mine without a necropolis (as the policy's expand order); a night elf's tree stands 448 from the mine (at people's ~800 or at 650 the instant entangle never took); a target's hall counts as made at its own time; no expand or no-expand mode is drawn.

fgself-173 (19:54 on 2026-10-31, from fgself-168's 67203721 and its league, its recipe): two-base starts, --scenarios natural,base2,base2,late,late,late at a share of 0.8, so ~27% of the training games, with fewer late starts (40% of the training launches drawn, from 60%). It evaluates current, with a target and the start. Judged at ~50M steps and ~250 games from its second hour on: expansions and gold a minute in real games, wins past 18 minutes, the top unit's share, head to head, against insane by race, with a target. A win rate near 100% in base2 games against the AI would mean they put on no pressure.

How early a verdict can be read (2026-11-01, the last seven runs, each cumulative against its start at its own step counts, head to head and points against insane with the races weighed alike):

Run	20M	30M	50M	End
fgself-166	53%, +8	55%, +8	53%, +7	54%, +6
fgself-167	52%, +2	52%, -3	49%, 0	46%, -3
fgself-168	47%, +4	50%, +5	56%, +7	59%, +7
fgself-169	40%, +4	39%, 0	43%, -1	42%, -2
fgself-170	41%, -3	47%, 0	46%, +5	46%, +2
fgself-171	43%, -7	41%, -8	37%, -11	37%, -11
fgself-173	48%, 0	49%, +2	48%, +1	48%, +2

At 30M the points against insane are within 5 of the end, and head to head within 9 (fgself-168 rose after it, so a level reading at 30M must go on). Stopping at 30M at or under 45% with no gain against insane would have cut fgself-169 and fgself-171 and no kept run. This also corrects an earlier note: a fork's first hour does not always read low. At 10M fgself-170 read -11, fgself-171 -19, fgself-174 -16, but fgself-166, 167, 168, 169 and 173 read +2 to +4, so the dip comes with the larger changes. Forks keep their Adam state, so that is not its cause.

Stopped for a shutdown (22:50 on 2026-10-31): fgself-173 at 41.5M steps (checkpoint 41524485, loads; its rows parse), under the 50M floor, so no verdict yet. Level so far at 384 games a policy: against insane 70% to its start's 69% (+0.5 points with the races weighed alike), 49% of 646 head to head, with a target 31%. From its second hour on (258 games to 265): 76% to 71% (+5 points, 95% interval -3 to +12), human 88/76, undead 65/45, night elf 83/94, orc 66/70; workers +1.9, gold a minute +21, lumber +9 (all significant). Play has not changed: expansions 0, the top unit 0.93, anti-air 0.40 to 0.43, wins past 18 minutes 37% to 33%.

The base2 games so far. Against the AI 86% won (of 1,506 at 22:07) at a median of 5.2 minutes: the second mine buys a faster rush, not a long game, so these games put on little pressure. Whether the AI's curriculum tax applies in them is not yet checked; without it they would be a fairer test. Between agents 81% tie at the 40-minute limit (1,199 of 1,488; median 40 minutes, against ~8 in other agent games), with ~26 workers and more tech; the side ahead seems unable to finish, perhaps not finding the other base, not confirmed. They cost steps: speed fell from ~4,400 to ~4,100 steps/s, and a tie reads -0.5 for both sides. Real-game ties have not risen as far as the wins and losses show. By the code the AI's curriculum tax does apply in these games: the draw of a launch's tax does not look at the scenario.

The shops (2026-11-01, the user's observation). The policy does build its race's shop: 42% of 931 evaluation games without a target (human 44%, orc 33%, undead 50%, night elf 42%), 89% with one. But it can never buy what the shop sells: its vocabulary holds only the goblin merchant's 11 items, so healing salves, lesser clarity, rods of necromancy and the like are not orders. It buys ~1-1.8 of the merchant's items a game (mostly invisibility, dust, teleport staffs), so the shop's cost buys it nothing. People, in 889 1.29 replays on all maps: a shop in 89-100% of games (median minute: orc 1.5, undead 0.9, human 5.7, night elf 5.8), an item in 92-99%, 7-15 item clicks a game. Orc: salves 3.8, lesser clarity 3.2, speed scrolls 2.7. Undead: rods of necromancy 2.8. Human: lesser clarity 2.2, regeneration scrolls 1.8. Night elf: lesser clarity 1.3, healing potions 0.9, moonstones 0.7, orbs of venom 0.7. These are clicks, so a refused one counts too. The race shops' wares are being added behind a flag, then tested as one change. Merged (--shop-items 1, default off; fullgame/shopitems.py, scripts/shop_items_probe.py, built and probed by a subagent): each race shop's items as orders, bought by a hero at its side's shop and used in the form the item takes. In the real game on echofast2 every offered item was bought, and most effects were seen (healing and mana potions, salves, scrolls, town portal, ivory tower, staffs, rod of necromancy in 5 of 6 tries); the orbs' numbers are one noisy try each, and the effects of moonstone, anti-magic potion and dust were not visible to the harness. Left out: the mechanical critter (its critter's type is outside the vocabulary) and the tiny great hall (unreliable). Two findings. Agent slots are computer slots, so the engine uses potions, regeneration and salves on its own in fights: a person must click, and item_uses (the policy's orders) differs from items_gone (what left the inventory). And the new orders start rare: in the probe's game at that rarity nothing was bought, so a run may need a lesson before it buys at all. Remote actors, selfplay_fork --set and the evaluation worker were read, not run.

fgself-173, level, kept as the base (11:13 on 2026-11-01, 50.6M steps, ~480 games a policy): against insane 73% to its start's 71% (+1.4 points with the races weighed alike), 48% of 756 head to head with its start, with a target 30%. From its second hour on (356 games to 370): 78% to 73% (+5 points, 95% interval -1 to +11), human 91/79, undead 67/45, night elf 83/95, orc 71/73; workers +1.6, gold a minute +16, lumber +8 (all significant); wins past 18 minutes 38% to 32% (not significant). Play has not changed: expansions 0 of 356, the top unit 0.93, anti-air 0.41 to 0.43. Real-game ties 2.1%, so the long ties in training did not carry over. In training: base2 against the AI 87% won of 2,715 at a median of 5.1 minutes; between agents 73% tied (1,737 of 2,381) at a median of 40 minutes. Cost: 3,964 steps/s (checkpoint times), -7.5% on fgself-170's 4,285. Kept because nothing fell overall, the weakest race rose and the economy with it, though night elf fell, and the policy still does not expand on its own.

fgself-168 on other maps (2026-11-01, the user's question whether Echo Isles favours rushing; a subagent; the tables, 482 games against insane, no training): 65% on Echo Isles, 60% on Last Refuge, 40% on Amazonia, 38% on Terenas Stand LV. On all four it masses one unit, never expands (0 of 482), gathers ~500 gold a minute against the AI's ~900, and wins 0-3% past 25 minutes. This does not say whether Echo Isles taught it to rush (the user's point): a policy trained only on Echo Isles carries the same habits anywhere, so its play elsewhere says little about what training elsewhere would teach. What it does show: one mine's income against the AI's is short on every map, and people's games on Echo Isles are no shorter than on the others, which is weak evidence at most. Only training on the other maps can answer. --map-pool is merged, off by default, to test that.

Why the side ahead doesn't finish (2026-11-01, the user saw undead leave human's natural alone; a subagent, 16 real base2 games between agents with fgself-173's 50855953 on the CPU, 13 tied at 40 minutes). Not a bug or a mask: attack-move at a point stayed open and its bins reach every mine within 27 units. In the last 10 minutes the side ahead has a median 19.6 army units, 9.8 at the dead main and none at the natural, no enemy building in sight, 363 point orders a minute of which 187 go to the dead main and none to the natural (44 of 170,502 in a whole game name the natural). In 8 of the 13 ties it never saw the hall that survived; in 5 it saw it and left. The trailer has no army and no workers from a median minute 16, and its new workers die at its own dead main (228 of 325 logged), which may help keep the army camped there (not tested). --hunt-lesson 1 (merged, default off, needs --teach; labels only, from what the player can know): with no enemy building in view, no enemy army seen for 2 minutes, 3+ seen buildings gone and 6+ army units, half the army is taught an attack-move to the nearest remembered building, else the nearest mine not in sight for 4 minutes, every 30 s at most. Labels-only firing on the same policy: 0 lessons in 6 games from normal starts and 8 against the AI; 18 in 6 natural, 89 in 6 late, 222 in 8 base2; none of 329 with an enemy building in view. A scripted follower uncovered the natural in 141-173 s and won in 163-201 s, for all four races, in a set-up probe. --inputs 3 was not needed. No policy has trained on it yet.

fgself-174 (2026-11-01, from fgself-173's last, its recipe and league): the AI's tax off (--curriculum -1, the user's request), the one change. Training games tax the AI's income and evaluation games do not, so training hides the income gap that loses the long games. Without a curriculum no level is restored from the league (League.load keeps only keys of self.level, empty here) and no launch is taxed; the win balance between agents stays. It evaluates current, with a target and its start. Judged at ~50M steps and ~250 games from its second hour on; its training win rate against the AI should fall towards evaluation's.

fgself-174, worse, not kept (14:58, 50.4M steps, ~440 games a policy): against insane 61% to its start's 77% (-15 points with the races weighed alike), falling; from its second hour 59% to 77% (-17, 95% interval -24 to -11): human 70/67, night elf 85/97, orc 54/78, undead 27/64. Head to head level, 51% of 759; with a target 23%. So its training at the AI's full income made it no weaker against its own start and much weaker against that AI. In training against the untaxed AI from normal starts it won only 14% (undead 3%, night elf 12%, human and orc 25%), so most of its games had nothing to learn from, and undead collapsed: too hard a curriculum, not proof that hiding the gap helps. Its play moved in the right direction: workers +2.3, upgrades +0.31, lumber +10, anti-air against 4+ air 0.57 to 0.46 (orc 0.72 to 0.38), top unit 0.90 to 0.93 (all significant), and it expanded in 23% of those training games; in real games still none. Two-base starts were a real contest there (48% won). The tax comes back; a tax kept lower, or a value that sees it, are the milder forms. The base stays fgself-173.

fgself-175 (15:00, from fgself-173's last, its recipe and league, the tax back): the hunt lesson (--hunt-lesson 1). Judged by the ties in base2 games between agents (73% in fgself-173), hunts found, then as usual. Its videos record on the box at 4× (--video-host, --video-speed 4), which doesn't touch training.

fgself-175 at 50M (18:14, ~420 games a policy): the lesson does what it was for. Ties between agents since its second hour: base2 8% (from 73%, median game 12.9 minutes from 40), late 6% (from 32%), natural 5% (from 19%), normal starts 5% (from 8%), which also lifted training to ~4,600 steps/s (fgself-173 ~4,000). Real-game ties stay 2%. Strength is not settled: 51% of 713 head to head with its start; against insane 65% to 72% (-7.5 points with the races weighed alike; from its second hour -7, 95% interval -14 to 0), human 73/60, night elf 88/98, orc 47/74, undead 52/57. By hour 65/71, 59/76, 71/67, recovering. Workers +2.3 (significant); the top unit, upgrades and wins past 18 minutes level; still no expansions in real games. Not judged yet: it runs on to ~65M, about an hour, since the last full hour was ahead and a run has risen late before (fgself-168).

fgself-175's verdict (19:10, 65M steps, ~525 games a policy): not kept. Against insane 63% to 72% (-8.5 points with the races weighed alike; from its second hour -9, 95% interval -15 to -3), by race human 71/63, night elf 87/98, orc 44/68, undead 53/61. The last full hour was 62/71, so the recovery at 50M did not last. Head to head level, 52% of 887. Workers +3 (10 to 7, at the same 523 gold a minute: one mine is full, so the extra workers cost gold and bring none); the top unit, upgrades, wins past 18 minutes and expansions level. The extra workers don't explain the race results alone: human also went 7 to 11 and rose, while orc went 7 to 11 and fell. The lesson does end the ties, and that's worth having, so it isn't dropped for good: a version that fires less often, or only in games between agents, is a later test. The base stays fgself-173/50855953.

fgself-176 (19:12, from fgself-173/50855953, its recipe and league): --ref-kl 0.01, the first of the queue, no hunt lesson. Does a looser anchor bring some of fgself-171's variety (at 0: orc 2+ kinds 89 vs 32%, night elf tier 3 81 vs 4%) and keep the macro it lost? Read at ~30M and ~50M.

fgself-176's verdict (22:28, 50M steps, ~450 games a policy): kept, level and slightly more varied. Head to head 52% of 633. Against insane 69% to 72% (-2.7 points with the races weighed alike; from its second hour -5, 95% interval -11 to +2), by race human 74/66, night elf 81/96, orc 62/63, undead 61/64. Night elf is the one real loss, though still the strongest race here; its top unit went from 98% to 95% of the army. From its second hour: creeps killed 1.48 to 1.05, heroes 1.11 to 1.00, anti-air against 4+ air units 25% to 16%, the top unit 90% to 92%, all significant; orc upgrades 2.07 to 1.58, orc anti-air 67% to 45%. Income holds, 516 to 524 gold a minute (interval -19 to +1), so the anchor at 0.01 keeps the macro fgself-171 lost at 0. Wins past 18 minutes 25% to 36% (interval -25 to +3), not significant, and no expansions. The move is small, far from fgself-171's. --ref-kl 0.01 stays; the base is fgself-176's last.

fgself-177 (22:29, from fgself-176's last, its league): fgself-171's last as a fixed league member (--league-add), the second of the queue. Does playing against a policy that mixes orc armies and goes to tier 3 as night elf move the army mix and anti-air, level or better? Paused for the night at 12.4M steps (23:17): 57% of 145 head to head, against insane 60% to 66% over ~100 games a policy, too few to read; the fixed member gets games (176, against 189 for fgself-130).

fgself-177's verdict (2026-11-02 12:04, 50M steps, ~400 games a policy, resumed at 09:36 after the night; read from the resume on): kept, stronger head to head. Head to head 58% of 640 (352 won, 246 lost), and 55-60% against every past policy in the evaluation set. Against insane 67% to 71% (-3.9 points with the races weighed alike; from the resume -3, 95% interval -10 to +4), last full hour 79/72; by race human 71/78, night elf 96/90, orc 43/51, undead 53/60. Play: weapon and armour levels 0.80 to 0.56 (orc 2.51/1.89), the top unit 90% to 91%, both significant; lumber 86 to 105 a minute (night elf 169/248), significant; anti-air, heroes, creeps, gold, wins past 18 minutes level; still no expansions. So the fixed member did not bring fgself-171's army mix; it made a policy that beats its own lineage more often. The base is fgself-177's last.

fgself-178 (12:23, from fgself-177's last, its league): a map pool in training (--map-pool "echofast2,lastrefugefast2:0.33,amazoniafast2:0.33,terenaslvfast2:0.34": Echo Isles half the launches, Last Refuge, Amazonia and Terenas Stand LV the other half). Does a policy trained on maps with farther and more bases rush less? Judged on Echo Isles as before (the evaluation map is unchanged), then on the four real maps. Its first minutes crashed under the undervolt test (optimization log); it resumed from 328,833 steps at 12:32.

The queue:

1. (done: in fgself-155) The audit-12 economy fixes: a training map whose mines and start scale with its costs (echofast2), workers labelled up to the target's count, masks against duplicate tech buildings and surplus workers, orc burrow and no-target lumber lessons.
2. Cloning the scripted strategy family as anchor and teacher (AlphaStar's recipe with scripted data). First tune its weak plans, and add night elf and undead expansion plans. Then record thousands of echofast games (scripts/heuristic_match.py --record), train a cloning policy, and use it as the KL anchor and a distillation teacher.
3. --teach-heuristic (live lessons from the matching scripted strategy), the exploiters back on, fine points in a run.
4. Snapshots by wall time (every 15-30 minutes).

Still open: War Stomp, Hex, Banish, Far Sight and Scout in the vocabulary; later research levels' costs.

The hand-written player (2026-10-25, the user's idea: a scripted opponent stronger than the learner shows its weaknesses, then joins the league and serves as a yardstick; branch worktree-agent-ac6cbc5467d415317). fullgame/heuristic.py, HeuristicAgent(side, race, strategy=...), the shape of play.BCAgent (begin, observe, commands, accepted); it sees what its player sees (enemy units while visible, their buildings remembered, the map's creep camps). Matches: scripts/heuristic_match.py (Echo Isles, 0.5 s steps, 40 minutes, the run's cuts and handicap, the learner through play_one with CPU inference, niced games named heur*).

• The tuned plan, orc (orc-grunts): altar and barracks at once, a second and third barracks at ~1 and ~2 minutes, burrows ahead of supply, 5 peons on gold and 3 on lumber, a Far Seer (wolves only while none are out, chain lightning on the most clumped enemies), grunts and headhunters 1:1, melee and ranged upgrades; against undead and orc a war mill at 2:30 and six watch towers by 7 minutes in front of the base; the natural taken once its camp is cleared; attacks with 45 army food or from 11 minutes, retreats when the enemies near its army are clearly stronger (Lanchester: hit points times damage a second), focus fire for its ranged units.
• A family of strategies (the user's direction after the first results: diverse, coherent plans to clone): 14, per race a tier-1 rush, a tier-2 army with a second hero (orc raiders with a Blademaster and Tauren Chieftain, human knights behind a castle, undead crypt fiends and gargoyles, night elf huntresses and dryads), casters (orc shamans and witch doctors, human priests and sorceresses) and for orc and human a fast expansion. Each is a Plan (build order, army mix, tiers, heroes, expansion and attack timings) in heuristic.STRATEGIES.

What the learner (fgself-145's 13576303) does and where it is weak, from its evaluation rows and 438 games against the scripted player's versions of the day (230 won by the script, 184 lost, 24 tied):

• It masses tier-1 units from 3-4 production buildings (human 44 footmen, orc grunts and headhunters, undead 33 ghouls, night elf 33 archers by ~9 minutes) with one hero and ~12-15 workers, and never expands: its one mine (13,500 gold) runs dry at ~13 minutes.
• It gathers twice as much as a normal player: its agent slots carry the advisor's difficulty, insane, and the engine doubles an insane computer slot's income (engine notes). Its 95% against the "untaxed insane AI" is at equal income.
• Against the scripted player, games past 14 minutes were won by the script 61 of 78 (its mine dry, the script's expansion up), games within 10 minutes 7 of 28: the learner wins by its early waves (undead's ghouls from ~2 minutes), loses the long game. 21 of 29 early losses came at the script's base, without an attack of the script's.
• It never builds towers and loses to them: against its undead, six watch towers by 7 minutes took the script from 3 of 8 to 7 of 8 on the same matchups (later batches ~55%).
• Its human is weak (footman spam from one barracks): the script won ~80% of those games.
• It cannot find a last building: ties at 40 minutes where the script's expansion hall survived its base.

What did not help, on the same 16 matchups each (noise ±2 wins): focus fire for every unit (17 of 32 against 16 of 32 without), the hurt hero walking home, a Tauren Chieftain against undead (1 of 8), meeting a wave near the hall with patient attacks (2 of 9), the army waiting by the miners against undead (3 of 8, 3 of 8 without). Spells did (27 of 48 against 6 of 16 without). Bugs on the way, each a measurable loss: the fight orders and spells were dropped for the first ~100 games (marked as given before they were sent), the game boarding the peons into burrows by itself during raids (no income for minutes), lumber workers taken as builders (no lumber), and a budget that a lumber-short reserve made negative (no grunt trained, audit 10's bug in the script).

Final measurements (the branch's last version, both agent slots at the learner's difficulty, insane):

the script	against	games	won	lost	tied
orc-grunts	fgself-145's 13576303 (master's agent code)	32	20	10	2
	by race: human 8-0, night elf 7-1, orc 4-3-1, undead 1-6-1
orc-grunts	fgself-150's 14800351 (the live run's newest)	16	7	8	1
orc-grunts	the insane AI	16	5	10	1
orc-grunts	the insane AI, an earlier version	8	5	2	1
the family, each of 14 in turn	fgself-145's 13576303	16	8	8	0
the family, each of 14 in turn	the insane AI	16	9	6	1

In the final 32 games every game decided within 10 minutes was the learner's (9) and every one past 14 minutes the script's (15). Against insane the script loses to its towers and tech (undead's spirit towers: 0 of 4): the three are not transitive (the learner beats insane 95%, the script beats the learner ~60% and loses to insane ~55%).

Demonstrations and lessons. heuristic_match.py --record DIR writes each game as collect.py's .npz (collect.Recorder: units, heroes, players, events each step; as orders both sides' orders in the state they were given in, the script's accepted commands by collect.command_row and the built-in AI's issued), with meta["strategies"]; bc.py --data DIR reads it as any collection, and strategy.from_demos makes each side's realized build its target (src: "heuristic", strategy: its name), so strategy.py --demos DIR builds a pool of them and cloning is conditioned on them. --vs heuristic --strategy random --race all plays the family against itself (both sides labelled). In a smoke test of 16 games against insane each scripted side gave ~6,500-11,500 labelled unit orders (the AI's ~2,000). selfplay --teach-heuristic 1: a training side whose target came from a scripted game is taught by that strategy itself (a HeuristicAgent of it shadows the side; its builds, training, research and tiers label the step as the teacher's lessons do). League: --heuristic-share, --eval-heuristic (eval key vs heuristic:orc, rated as a player of its own), --heuristic-strategy random.

At a person's income (2026-10-26, phase 1 of cloning the family; branch worktree-agent-a2242f5c07f72f866). The family had been tuned while every agent slot gathered the insane AI's double income. Measured at normal (a person's) income against the insane AI (which keeps its double), 8 games a strategy on Echo Isles (each opponent race twice) and 4 on echofast2, ~1 game a minute per slot at nice 15 with fgself-155 running. Before: 17 won, 90 lost, 5 tied of 112 (15%); echofast2 0 of 56. Economy: one mine with 5 workers gives ~500-550 gold a minute by 10 minutes (13,500 gold: dry at ~23-25 minutes alone). Undead and night elves never expanded. Lumber piled up (from 4 to 15 minutes: undead 900-1,500 banked, night elf 400-760, human 260-590). Undead and night elves sent a sixth worker to a full haunted or entangled mine every step (~1,860 and ~1,020 refused harvests a game). Changed:

• every race expands: undead haunt the natural's mine (an acolyte's target order with the haunted mine's code), night elves root a tree of life at the natural and entangle its mine (ENTANGLE, every 10 s until it is); new undead-expand and nightelf-expand (16 strategies). Undead and night elves start on their own haunted or entangled mine: a plain-mine-only list had made the natural their "main". Every plan takes the natural (rushes from 8 minutes, the others from 7, the expansions from 3:20-4; 1 minute later against undead and orc, was 2), or when the main mine falls under 5,000 (was 3,000);
• lumber workers to what each plan spends (ghouls 1-2 for the ghoul rush, 2-3 for the others; human and night elf rushes 3; tier-2 and caster plans 4-5); no sixth worker on a haunted or entangled mine;
• money piling up (700 gold and 250 lumber from 8 minutes) buys the race's weapon and armor upgrade levels as the tier allows, the next tier, and more of the first production building below the food cap (one research a building a step; a refused one waits 30 s);
• two army bugs: out creeping or attacking, the army's centre is now the group that left the rally point (the new units there had pulled the median home and the leaders back to it: in half the games, one start, the army bound for the natural's camp 6,000 away swung between home and halfway for 7-11 minutes, the expansion waiting; now ~8.5 minutes from either start); a camp none of whose creeps died for 120 s is given up for 5 minutes, and one seen empty is gone (creeps killed out of sight send no event);
• the timed attack (two thirds of the plan's army from attack_at) only with 1.2x the power of the strongest enemy army met; a hall stops training workers for its tier only once it can pay (undead's one hall had trained no acolyte for minutes).

After (the same matchups, 8 games each; won-lost-tied, then the share of games with a working second mine and its median minute):

strategy	before	after	expansion after		strategy	before	after	expansion after
orc-grunts	2-6-0	1-6-1	5/8, 9.5		undead-ghouls	0-8-0	1-6-1	8/8, 15.0
orc-raiders	1-6-1	1-7-0	6/8, 9.5		undead-fiends	1-7-0	1-7-0	8/8, 12.5
orc-casters	1-7-0	1-7-0	6/8, 9.5		undead-gargoyles	2-6-0	1-7-0	8/8, 11.5
orc-expand	2-6-0	2-6-0	7/8, 6.5		undead-expand	(new)	1-6-1	7/8, 6.0
human-footmen	1-6-1	2-6-0	6/8, 10.5		nightelf-archers	0-7-1	1-7-0	5/8, 9.5
human-knights	2-6-0	2-6-0	8/8, 9.5		nightelf-huntresses	1-6-1	1-7-0	8/8, 9.0
human-casters	2-6-0	2-6-0	7/8, 10.5		nightelf-dryads	0-7-1	2-6-0	6/8, 11.0
human-expand	2-6-0	1-7-0	8/8, 7.0		nightelf-expand	(new)	2-6-0	7/8, 7.0

In all 22 of 128 won (17%; undead and night elf from the last version's 64 games), level with before; echofast2 0 of 64. Every win is against insane human (17 of 28 before, 22 of 32 after); against insane orc, undead and night elf 0 of 84 before and 0 of 96 after (1 in an earlier round). The economy is near its ceiling on one mine; the AI gathers 2-2.5x (e.g. 34,000 gold to 14,000 in 26 minutes), expands twice, and reaches tier 3, and the script's attacks into its base lose 3-10 units for each one killed. Lumber banked from 4 to 15 minutes: undead 300-570, night elf 140-320, human 230-360. Refused harvests ~0. Still weak: undead's haunted natural is raided and its main falls at ~16-20 minutes (6 acolytes at 10 minutes); lumber still piles up late (nothing to spend it on at the food cap but upgrades); a hall ordered at a natural whose camp was given up can be refused again and again; on echofast2 the AI's double income at half costs overruns every strategy by 10-13 minutes.

For phase 2 (recording at scale): an opponent the family can play whole games against (the family itself, --vs heuristic --strategy random --race all, or the AI at normal), since against insane most recorded games are a plan cut short by a collapse at 15-20 minutes; ~240 real-map games an hour on 4 slots (a few thousand games a day, more on the games box); a check of the recordings' labels (refused orders per strategy, games where the army stood still) and of echofast2's timings (half costs: fewer lumber workers, earlier tiers) before recording there.

Skill ratings (the user's idea, the dashboard's Ratings view): a Bradley-Terry fit on the Elo scale over every evaluation game in the real game, the insane AI at 1500. At 22:45: fgself-106's last checkpoint 2128 ± 24, fgself-102's 11841833 1977 ± 8, fgself-100's best 1956 ± 9.

Queued (2026-10-19, the user's idea): GRPO. The game is deterministic (lockstep multiplayer needs it), so a group of games can start from one state and differ only by the policy's sampled actions; each game's advantage is its return against the group's mean (and spread), with no value function (DeepSeek's GRPO). Plan: (1) check determinism: the harness has no seed control yet; the map script would call SetRandomSeed at start, and two games with one seed and the same orders (none, or a replayed order log) must give identical unit rows step for step, the built-in AI included. (2) The main design (agreed with the user): branch mid-game. A game from one start gives little credit for any decision (a whole game's result over hundreds of them), so: play a game, pick a state at step t (e.g. 2-8 game minutes, near a decision that matters: a tier, an attack), replay the same seed and orders to t in G (~8) games side by side, then let each continue with its own sampled actions; each continuation's advantage is its return against the group's mean and spread, the policy loss on the steps after t, the KL to the anchor kept, beside the usual GAE games. Replaying to t is cheap on echofast (a game runs ~13x real time). If replays drift (the built-in AI's threads), the fallback is the game's own save and load (SaveGame / LoadGame in single player; whether the AI scripts survive a load must be checked). (3) Only if branching fails: groups from one start (races, seed, target, opponent). Compare to the current recipe at equal steps. The determinism check (2026-10-19, engine notes): with a seed (GameSetup.seed, new), games between agents with the same orders are identical step for step; games with the built-in AI are not (they differ after 13-18 s). So GRPO branches come from self-play games (against past snapshots and itself); with a seed each player now also starts at its own start location (4 of 4 seeded games identical for 300 steps). Left to build: the branch scheduler (replay a game's orders to step t in G games, then sample) and the group-relative policy loss.

How the best policy wins and loses (fgself-73's policy: its own run's 347 evaluation games and its last checkpoint's 141 in fgself-74: 488 games against the untaxed insane AI, 356 won). One strategy: an early rush of tier-1 units. The median win came at 9.7 game minutes, 206 of the 356 wins within 10 minutes, and only 38% of the wins reached tier 2. The losses were long (23.7 minutes on average), with tier-1 units and many workers (56 a lost game against the AI's 14, army units 89 against 38). Heroes 1.0 a game against the AI's 1.5. Undead's lumber (83 a game minute against the AI's 213, fgself-69/-70) is still its weak point.

The plan (2026-10-17, agreed with the user), in order. 1. Finish fgself-75. 2. An exploiter league against the early rush (AlphaStar's main exploiter is in selfplay.py, --exploiter-share, not yet used in whole games). 3. Strategy conditioning as AlphaStar's z: a target build order and unit mix sampled for each game, from the built-in AI's Echo Isles games and later from human replays, with a small reward for following it, to break the single rush and force tech. Human replays: ~3,000 replays of 1.29 from warcraft3.info, with a converter for their build orders (any version) and their playback in the real game as demonstrations (1.29 only, see human replays).

Next, ranked with what AlphaStar and OpenAI Five found (ideas): forward KL to the anchor on every head (ours is reverse KL on the orders only), a temperature for the targets, features of each unit's own last order and hit-point history (branch features-v2 has the hit-point change and sine/cosine positions), a value with both players' information.

A fairer yardstick: the AI's own record in the same matchups. duelfast is far from balanced even for the built-in AI against itself (3,000 games, mixed matchups): night elf wins 81% (normal) / 75% (insane), orc 65% / 71%, undead 33% / 37%, human 22% / 19% (human vs night elf 8% / 6%, vs orc 15% / 6%). So a race's real-game win rate says as much about the map as about the policy. scripts/real_games.py now also counts, for real games against normal and insane, how many the AI itself would win in the same matchups: from 2026-10-03 18:36 to 2026-10-11 00:17 the learner won 195 of 896, the AI would win ~455 (43% of the AI's record; 45% in the last two hours before the anchor experiment). 100% is the built-in AI's level.

The built-in AI as an advisor (shadow games)

Takeover games label only the states after the takeover, and they soon become the AI's states. On-policy distillation asks the teacher for a label in every state the student reaches. The built-in AI is a script, not a function, so it cannot be asked. In a shadow game it plays the policy's player too: the harness records its orders as labels and undoes them (protocol.ShadowAI, collect.py --shadow).

undo	how
training, research	cancelled when they were paid for (the player's resources before and after the order)
build orders, orders to the agent's builders	the agent's last order again, at once
a hero skill	unlearned, the point back
town bell, burrows	switched off
any other order	stands until the step ends; then the agent's last order again, before the observation
the engine's own (a worker going back to its mine, autocasts), a trained unit's rally	not touched

Each version against the same 16 games without the advisor (fgself-12 at 20.9M, duelfast, normal and insane AI):

version	food at 1 min	food at 2 min	gathered by 2 min	what was wrong
no advisor	16.6	26.4	1,412
undo everything at once	12.7	18.2	1,266	units the AI ordered every tick lost their attack swings: 78 deaths by 2 minutes, not 51
other orders undone at the step	15.9	20.5	1,284	human buildings finished: 1.2 a game, not 4.2. The AI pulled builders away.
builders undone at once	14.4	19.5	1,262	the restored builder's order is "repair", so the policy gave it another order
the policy keeps repairing builders	17.4	25.7	1,402

The labels cover 5.9% of unit-steps (the AI's own games: 3.5%), 770 train or research orders and 816 build orders in the 16 games. demos-fast-shadow-12: 800 games from fgself-12 at 23.7M steps.

Distillation in self-play (on-policy distillation: the teacher labels the student's own states while it learns, so the labels never go stale). From 24.1M steps, fgself-12 --opd-share 0.5 --opd-coef 0.05: in half the games (never the real game) the insane AI advises the learner. Its orders of a step label that step (features.Encoder.step_labels, the same code as the demonstrations' encoder). The update adds the cloning loss on those steps, from the PPO pass's outputs. First updates: 45–60% of a batch's steps advised, the distillation loss 2.6–3.0, the policy's first choice the AI's order 51–59% of the time. Steps/s unchanged (750–800). The policy's KL per update 0.004–0.006 (0.001 before), its distance to the anchor 0.005 (0.002). The 28 offline shadow games collected are kept, unused.

Where a self-play step's time goes (speed, 2026-09-30)

fgself-9 ran at 560 agent steps/s with 32 games. A game thread spent 55% of its time waiting for the policy and 36% for the game. After this round the learner is the limit with the machine's CPU close behind, and it runs at about 840 steps/s with whole games, five to a load of the map (607 before the reset fix, when half the games were the cheap broken ones).

change	agent steps/s	what it showed
start	560	the inference server 90% busy at 3.4 rows a call
a pipe per game thread instead of queues	545	a request took ~32 ms of a 59 ms step, the server's call 4.5 of them: feeder threads, a shared write lock, a reader thread, each a wait for a GIL
games niced (+10)	595	the server and the learner get the CPU first: a call 4.2 → 2.9 ms, an update 11.4 → 9.5 s
a round's calls started together, one wait	606	next to the learner every wait costs a turn of the GPU (a call: 0.8 ms alone, 1.7 next to a synthetic learner, 2.5 in the run); a round 3.9 → 2.8 ms
CUDA graphs captured by hand	607	0.4 s a shape instead of torch.compile's 10-20 s (the games waited); a new league member no longer stalls everything for 25-50 s
every game reloads the map (the fix above)	540	the games between agents are real games now (47 ms a step, not 32)
the loading screen drawn in the game's thread, a video every 10 minutes	644	a reload's CPU 2.1 → 1.4 s
the pass over the units made by the shim (C) instead of the harness (JASS)	719	a step 41 → 36 ms; the learner is the limit again (an update 10.5 s, the actors waiting 1 s)
minibatches of similar entity counts, no GPU waits for the statistics	≈800	an update 8.9 s; the learner is bound by its own Python (the GPU is 36% busy)
five games per load of the map, each with two players of its own	822	the actors now make more steps than the learner trains on: an update 9.3 s, the batches 3 updates old and falling behind
the trajectory queue holds one batch; an inference round at most every 4 ms	837	an update 8.7 s, the batches 1.3 updates old, the learner waiting 1.1 s an update

• The cloning batches. The loader takes its workers' batches in turn, and a worker reading 16 games before it emptied them gave nothing for seconds: the learner waited 26% of an update. bc.Steps is now a shuffle buffer kept full (9%).

• Only 58% of the games' time was play. A reload took 13 s with 32 games running (5 s alone), as long as the game before it.

  • "Allow Local Files" (a registry setting the harness needed when it read its actions from a file) made the game look in its folder for every file before its archives; under Wine each miss scans the directory. Off: 13.2 → 9.9 s.
  • The rush rules were map object data (w3u, w3q, w3a), which the game applies change by change at every load: 2 of 3 s. They are now the game's own tables with the rules' values, in the map: a reload takes 1.0 s alone, as on the plain duel map. Checked cell by cell against the object data (a unit test) and with 240 built-in AI games on each map (2.10 and 2.09 minutes; the races' win rates within noise).

• The machine is full. A game alone steps 68 times a second at 14.6 ms of CPU a step. With 32 games running, the same step costs 35-40 ms (two threads a core, the all-core clock, contention), and 16, 32 and 48 games step 418, 638 and 618 times a second in all. A reload per game is a quarter of the games' CPU.

• Dead ends. The inference call as a thread of the learner (one GPU context): capturing a CUDA graph fails while another thread draws random numbers. Emptying melee's preload lists: the shim's Preload hook already drops them. A slower clock while reloading: the loading screen waits about two game seconds drawing frames, so it costs more.

• The units, in C. Every step the harness went over every unit three times in JASS (collect them, count them for the result, write the records of those that changed): ~35 native calls and a few hundred JASS instructions per unit, 9-13 ms of a step with 150 units (4 ms on an idle machine). The natives are plain functions in the executable, found by name from the code that registers them, so the shim now makes the same calls itself: 4 ms (1.5). A game played with the harness's pass and its replay with the shim's agree in every field of every unit on every step (six games, both directions).

• The learner's minibatches. A minibatch pads to its widest step: 91 entities at random for a mean of 41, and attention costs the square. Cut from groups of 8 minibatches sorted by entity count they pad to 49. The update's KL, clip fraction and value loss stayed the same (0.0033, 0.024, 0.018); the gradient norm rose from 0.15 to 0.19.

• Five games per load of the map. A reload per game was a quarter of a game thread's time. The scripted reset's trouble is per player, so the duel maps are now also built for five pairs of players (GameSetup.pairs): each game of a load is played by two players who have not played, the next game starts 0.1 s after the last one's end, and the fifth game's end reloads the map. Python still sees players 0 and 1. Twelve players crashed the game at load; ten and the observer load.

  • The next pair's game has to start as a load's does. The first version gave the first observation one step late: the workers already on their way, the built-in AI's starting gold spent, the agent's first orders lost. Win rates hardly moved in 25 minutes, but orc built 0.60 altars a game instead of 1.57 and trained 0.4 heroes instead of 1.4, night elf alike: the policy's opening depends on its first observation. The per-game statistics by race showed it within minutes (the lesson of the reset above, applied).
  • The switch now takes three ticks of the game, as a load has them. Everything of the last game is removed (units, corpses, items; trees regrow, blight clears). A tick later, the gold mines the engine put back from under removed haunted and entangled mines are removed too, and the map's mines and creeps and the new players' starting units are made. Another tick later the units are collected as an enumeration finds them, and the first step follows at once.
  • Checked at three levels. Every unit, player and event of a later pair's first six steps against a load's, for all four races against the undead and night elf AI: the differences are those between two loads. Built-in AI games, 240 with five pairs and 240 with a reload each: 2.11 and 2.18 minutes, the races' win rates alike; and 298 games by their place in a load: food at step 80 within 1 of the matchup's mean at every place. The policy's altars, heroes and research per game by race in the run (orc 1.54 altars and 1.40 heroes again).
  • Collecting demonstrations (built-in AI on both sides, no network calls) runs 1.55 times as fast: 2700 against 1750 games an hour, next to the training run.

• The learner and the actors in balance. With the reloads gone the actors made more steps than the learner trained on. The trajectory queue had no bound, so the batches were 3 updates old and falling behind; it now holds one batch and the actors wait. Every round of the inference server takes the GPU from the learner for a turn, so a round now starts at most every 4 ms (--infer-period-ms; 174 rounds a second instead of 229, for the same rows): an update 9.3 → 8.7 s.

• Another dead end. The learner and the inference server on cores of their own (--pin): the update was no faster (9.0 s) and the games lost 4 threads.

What is left, largest first: the learner (an update takes 8.7 s for 8192 steps, so 940 steps/s at most; it is bound by its own Python: a compiled update, or fewer passes such as the clone's logits once per update), the video renderer (3.4 cores while it renders), and wineserver (a tenth of a step: the engine's threads signal each other ~500 times a step).

Memory and a value head in BC

Two additions to cloning:

• A memory core: a minGRU, scanned in parallel (--memory).
• A value head: it learns self-play's returns from the recorded games.

380 games, 3 epochs, the same 20 validation games. Every model is validated on whole game sides.

model	validation loss	order loss	order accuracy	target	point error (bins)	value explained variance
no memory	3.985	0.328	61.4%	76.2%	8.65	0.854
memory, 8 lanes × 32 steps	4.089	0.337	59.8%	75.4%	9.00	0.863
memory, 32 lanes × 8 steps	4.106	0.339	60.2%	74.0%	8.99	0.868
memory, 32 lanes × 16 steps (batch 512)	4.283	0.364	56.3%	73.9%	9.45	0.862

What the table shows:

• Memory predicts the value a little better and the orders worse. Batches of consecutive steps vary less than batches of shuffled steps, and the AI's decisions depend mostly on what it currently sees.
• The value head explains 85% of the variance of self-play's returns in held-out games, so self-play starts from a trained value.
• Clones stay without memory for now. Self-play can add a core that starts as a no-op (--memory 1).

An earlier comparison validated each model on the chunks its own training loader made. Those loaders stop when their first lane runs out of games, which favored early-game steps, so the numbers looked very different (3.39 vs 3.99).

Race balance on duelrush

In 3389 demonstration games night elf won 96% of its games against the other races, undead 68%, orc 26%, human 5%. The cause is the economy: the speed rules make mining and chopping 7 times faster, but walking only 1.3 times (the engine's movement limit). Human and orc workers walk every load of gold to the town hall, so their gold income rose 1.6 times; night elf and undead gold comes from the mine with no walking and rose 7 times. Gold mined in the first minute: human 571, orc 588, undead 3230, night elf 4005 (all four within 10% of each other at normal speed).

rules	human	orc	undead	night elf	game minutes
duelrush	3%	30%	64%	96%	2.1
duelrush, handicap 100 (units at half hit points, not a quarter)	13%	27%	76%	100%	1.7
duelrush, buildings at full hit points	0%	0%	86%	100%	1.9 (24 games)
duelrush, walking workers carry gold ×4.4, lumber ×2.5	25%	50%	29%	89%	2.1
duelrush, gold ×6.5, lumber ×4 (now the default)	37%	57%	29%	78%	2.1
duel (normal rules on the small map; 42% ties at 15 minutes)	31%	52%	16%	9%	12.4

Win rates against the other races, 96 games each unless noted. What a walking worker carries per trip is now a rule (Rules.gold_carry, lumber_carry), and duelrush has gold ×6.5 and lumber ×4 (map version v11; collections record the map file). Even normal rules on the small map are not balanced, so this is where tuning stopped. Self-play starts on mirror matchups, which are balanced by construction.

Other findings

• lr 0.01 (tuned on footmen2) is far too high with 15 action heads: the KL per update was 1.0-1.5 (clip fraction 0.9) and the win rate peaked at 29%. The KL at a given lr grows with the number of heads (≈0.15 with 6, 0.3 with 9, 1.0+ with 15); lr 0.003 keeps it at 0.03-0.14.
• Action masks: before masks, 94% of the from-scratch policy's casts on mirror_mix_abil_hp400 were impossible. With them, it reached 39% at 0.9M steps against 35% without.
• Updates per epoch are replay_ratio × batch / minibatch. With the minibatch equal to the batch (the old default), there was one update per epoch and learning was slow: footmen2 reached 61% wins in 1M steps, against 97% with 16 updates.