Student builder and open-source researcher · Granite Bay High School
I build reproducible systems that test what AI verification actually checked, then turn the result into evidence a stranger can inspect.
I work on one question: what did a successful-looking verification result actually check?
My current work is Nobulex, a decision-integrity boundary for automated financial actions. It keeps two questions separate:
- What does the available evidence establish?
PASS,FAIL, orINDETERMINATE. - What may the system do next?
PERMIT,BLOCK, orESCALATE.
That separation matters because missing or unreadable evidence should never quietly become approval.
| 40 | 11 | 27 | 4 |
|---|---|---|---|
| stars on Nobulex | forks of Nobulex | executable research fixtures | merged conformance-harness PRs |
GitHub metrics measured 2026-09-27. Fixture and PR counts link to inspectable artifacts below.
proposed action
│
▼
evidence binding ──► PASS / FAIL / INDETERMINATE
│
▼
bounded policy ──► PERMIT / BLOCK / ESCALATE
│
▼
signed decision receipt
01 · Collect evidence
Bind the proposed action to the exact inputs, sources, timestamps, and identities it depends on.
02 · Establish status
Return PASS, FAIL, or INDETERMINATE. Missing evidence never becomes a clean result.
03 · Apply policy
A bounded deterministic policy returns PERMIT, BLOCK, or ESCALATE.
04 · Preserve the decision
Sign a receipt that binds the action, evidence references, policy version, and outcome.
| Project | What it does | What is inspectable |
|---|---|---|
| Nobulex | Executable research on verification boundaries: skipped coverage, self-selected trust anchors, policy scope, log integrity, and ambiguous evidence. | 27 synthetic fixtures, six paired historical parser cases, and public CI. |
| Decision-integrity gateway | A Python prototype that separates evidence status from execution policy before a financial action can proceed. | PASS, FAIL, or INDETERMINATE evidence feeds a separate PERMIT, BLOCK, or ESCALATE decision. |
| Fail-open corpus | Minimal reproductions of evaluators and safety checks that can report success without completing the check their result appears to certify. | Each published case names its reproduction limits instead of generalizing from one defect. |
- Four conformance-harness pull requests merged - The public research distinguishes merged changes from endorsement and freshly reruns only the case it says it reruns.
- A verifier defect reproduced and fixed upstream - The maintainer reproduced the failure, released a corrected verifier, and confirmed all four negative checks.
- Verification-boundary research runs in public CI - The fixture expectations, historical replay metadata, and deliberate-regression checks are inspectable.
Open the complete Evidence Ledger →
The ledger separates normative changes, merged code, references, listings, open or closed contributions, and self-published research. A merge or listing is never presented as an endorsement.
I publish the command, the expected result, and the limitation beside the claim. If a result cannot establish something, it should say so. If I am wrong, the correction stays visible.
Canonical profile data: data/profile.json. This README is generated from it.