Skip to content

Backfill canonical_examples batch 1: 5 records filled, 6 honest skips (#444) - #474

Merged
realmarcin merged 2 commits into
mainfrom
feat/backfill-canonical-examples-batch1
Aug 20, 2026
Merged

realmarcin merged 2 commits into
mainfrom
feat/backfill-canonical-examples-batch1

Conversation

@realmarcin

@realmarcin realmarcin commented Aug 20, 2026 •

Copy link
Copy Markdown
Contributor

First canonical_examples batch after the #446 canary, addressing the ~130-record gap tracked in #444. Batch size: 11 records, taken from the top of the #449 priority dashboard's ADD_CANONICAL_EXAMPLES queue. Two adversarial review rounds ran on this PR; the second (independent) round removed two draft exemplars and is documented in #476–#478.

Rules (the canary's, applied unchanged)

  • Exemplars come only from the trait's own deep-research artifact; recalled facts don't qualify — and a cultivation condition is not a stated optimum (Batch-1 review: two of three temperature_optimum_high exemplars asserted optima their artifact never states #476).
  • Every taxon id resolved and label-checked against the local OAK NCBITaxon build; OAK labels used verbatim.
  • Contrast cases excluded by name, with the artifact's own disqualifying language recorded in each record's curation_history entry.
  • A record whose artifact establishes no instance is skipped, not approximately filled.

Filled — 5 records, 6 examples

record exemplar(s) evidence
environment/stenohaline Synechococcus sp. YX04-3 grew at 32 ppt, could not survive 13 ppt; euryhaline HK05 grew at both (note records the two-point assay and that the full interval is unknown)
environment/temperature_delta_high Exiguobacterium chiriqhucha strain RW2 grown 4–50 °C, a 46 °C breadth, inside the >30 °C bin
environment/nacl_delta_high Aspergillus sydowii, Wallemia ichthyophaga 0→saturated 5.13 M NaCl and 10–32 % w/v; both far above the ~8 % w/v threshold
environment/nacl_optimum Wallemia ichthyophaga kinetically determined 15–20 % NaCl optimum — the artifact's own worked example separating optimum from range
environment/temperature_optimum_high Saccharolobus islandicus 76 °C, the one optimum the artifact explicitly states (two draft entries removed on review, #476)

Skipped — 6 records, each with the artifact's own words

temperature_delta_very_low, ph_delta_mid2, ph_delta_mid3, ph_delta_very_low, ph_range_mid2, cell_length_large. Every candidate organism in those artifacts is a mechanism host, a one-sided measurement, or an explicit contrast case; several artifacts state the negative result outright. Best cross-check in the batch: E. chiriqhucha's single 4–50 °C measurement is simultaneously the exemplar for temperature_delta_high and the negative evidence for temperature_delta_very_low.

The cell_length_large skip generalizes: cell-size bins need a taxon-oriented source (#475).

Review trail

Round 1: reference DOIs mechanically verified present in each trait's own artifact; all taxon ids re-resolved. Round 2 (independent reviewer): #476 (two unestablished optima — removed), #477 (EOF placement vs corpus norm — normalized, canary included), #478 (bin-measured exemplar on a family parent — open policy question), plus five wording drifts fixed.

Verification

  • just audit-canonical-examples: 321 examples across 231 records, 0 errors, 1 pre-existing warning (heterocyst label drift, audit-canonical-examples exists (#446) but is not wired into qc or any workflow #445).
  • Corpus re-emitted through write_validated_trait; byte-canonical round-trip holds.
  • Pages regenerated; validate-strict, full qc, and 677 tests green.
  • Batch history record under history/other/canonical-examples-batch1/.

🤖 Generated with Claude Code

…#444)

First batch after the #446 canary, run by its rules: exemplars come only
from each trait's own deep-research artifact, every taxon id resolved and
label-checked against the local NCBITaxon build, contrast cases excluded
by name with the artifact's own disqualifying language recorded in each
record's curation_history, and a record whose artifact establishes no
instance is SKIPPED rather than approximately filled. Targets from the
#449 priority dashboard's ADD_CANONICAL_EXAMPLES queue, top first.

FILLED (8 examples): stenohaline (Synechococcus sp. YX04-3, grew at 32
ppt and could not survive 13 ppt while euryhaline HK05 grew at both);
temperature_delta_high (Exiguobacterium chiriqhucha, strain RW2 grown
4-50 C, a 46 C breadth); nacl_delta_high (Aspergillus sydowii, 0 to
saturated 5.13 M NaCl; Wallemia ichthyophaga, 10-32% w/v);
nacl_optimum (W. ichthyophaga, kinetically determined 15-20% optimum);
temperature_optimum_high (Saccharolobus islandicus 76 C, Pyrococcus
furiosus 90 C, Thermococcus kodakarensis 85 C at mu=0.69/h).

SKIPPED (6): temperature_delta_very_low, ph_delta_mid2, ph_delta_mid3,
ph_delta_very_low, ph_range_mid2, cell_length_large -- each because the
trait's own artifact establishes no organism as an instance, most saying
so explicitly. The strongest cross-check: E. chiriqhucha's 4-50 C
measurement is simultaneously the exemplar for temperature_delta_high
and the negative evidence for temperature_delta_very_low.

323 examples across 231 records, 0 errors, 1 pre-existing warning
(heterocyst label drift, #445); all 7 new taxon ids independently
re-verified; corpus re-emitted through write_validated_trait so the
round-trip invariant holds; pages regenerated; 677 tests, qc green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@realmarcin

Copy link
Copy Markdown
Contributor Author

Adversarial review done as a separate pass:

  • Reference provenance verified mechanically: all 8 reference DOIs grep-confirmed present in the corresponding trait's own research artifact (the rule the batch depends on), and all 7 distinct taxon ids re-resolved against the local NCBITaxon build with labels matching verbatim.
  • One finding filed as Cell-size bins cannot be filled from the mechanism artifacts; canonical_examples needs a taxon-oriented source #475: the cell_length_large skip generalizes — morphometric bins can't be filled from mechanism-oriented artifacts at all; they need a BacDive-style taxon-oriented source. That is a pipeline/source decision, not a fix for this PR.
  • Judgment calls reviewed and accepted, each disclosed in the record's own note or curation_history: the stenohaline exemplar rests on a two-point survival assay (qualitative trait, only direct culture evidence in the artifact); E. chiriqhucha's species-level id stands in for strain-RW2 evidence (same shape as the canary; NCBITaxon has no RW2 node); nacl_optimum is a parent class but an organism with a measured optimum is a real instance of it; W. ichthyophaga's "By contrast" framing in the nacl_delta_high artifact contrasts its high lower bound, not its breadth — the same sentence affirms the wide interval, so it is not a contrast case in the disqualifying sense.
  • Cosmetic note, not filed: DOI case is mixed across the corpus (AEM.00559-14 vs aem.02702-13); DOIs are case-insensitive and the audit doesn't normalize, so left as-is.

Awaiting your go-ahead to merge.

…s, normalize placement (#476, #477)

An independent adversarial content review of this PR (findings filed as
#476-#478) found that two of temperature_optimum_high's three draft
exemplars asserted optima their artifact never states: P. furiosus's 90 C
is the study's CONTROL condition against 98 C "supra-optimal" stress, and
T. kodakarensis's 75-93 C series endpoints are delta-rgy mutant data with
one wild-type point at 85 C. Both removed; S. islandicus (76 C, the one
explicitly stated optimum) stays. The batch is now 5 records, 6 examples.

Also from that review: the A. sydowii note no longer claims the multi-omics
study establishes growth at "both ends" of the 0-to-saturation interval
(it compared 1 M and saturated; even that spans ~24 points w/v, still far
past the ~8% bin); the stenohaline note now carries the artifact's own
"complete tolerated interval remains unknown"; three curation_history
quotes are restored to their conditional originals (Shewanella "without
additional endpoint data", "avoiding unjustified conversion", Halomonas
back in the MAG list); and canonical_examples now sits after `evidence`
in all five records PLUS the #446 canary, matching the 225-record corpus
norm instead of the canary's end-of-file deviation (#477).

#478 (bin-measured exemplars landing on family parents) is filed as a
policy decision, not changed here. Removing the two entries also removed
the PR's only two uppercase DOIs.

321 examples across 231 records, 0 errors, 1 pre-existing warning; round
trip, pages, 677 tests, qc all green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
@realmarcin

Copy link
Copy Markdown
Contributor Author

Second, independent adversarial review round complete (findings filed as #476–#478; fixed in f098513):

Reference provenance and taxon-label checks all cleared in this round too. 321 examples across 231 records, 0 errors; 677 tests, qc green; CI green on the head commit.

@realmarcin
realmarcin merged commit 3fcfa9d into main Aug 20, 2026
5 checks passed
@realmarcin
realmarcin deleted the feat/backfill-canonical-examples-batch1 branch August 20, 2026 19:42
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant