Replay the hash-pinned one-byte repair through a standards-library CSV harness, adversarial controls, a materially separate exact-diff checker, and a directly applicable patch check after refreshing upstream status.

Candidate — review needed

Run 3 refreshed the official source state, replayed the complete candidate gate and a materially separate checker, and added a checked one-byte patch artifact. The candidate exactly repairs algebra-036 and satisfies the full local verification contract. No external action was taken.

Candidate for review — not a solution claim

Independent statement checking, criticism, literature review, and verification remain required.

public research note

Human-approved research note; external acceptance and peer review are separate states.

Publication packet →

Strategy and discriminator

CSV structural validation and minimal-patch certification

The issue-supplied answer token fixes the malformed boundary; a strict CSV harness reconstructs and attacks the candidate, while a separate checker enforces exact schema, target fields, and baseline-plus-one-byte equality.

Hypothesis: Against unchanged upstream commit 96fa6c4cc3a9bb7450ee7b6773b659d3a030dace, the materialized candidate is exactly the required one-byte quote insertion, strict-parses as 400 six-field data rows, restores the exact algebra-036 Short Answer, and rejects wrong-answer and collateral-edit controls.

Test: Refresh upstream main, issue #13, and target-specific PR searches; if unchanged, run the recorded full harness and standalone checker against the pinned baseline and require all predeclared positive and negative signals.

Rationale

The decisive result is reproducible and independently encoded: both implementations require the same exact target semantics and whole-file one-byte scope, while the harness additionally rejects malformed, wrong-answer, and collateral-edit controls.

Claims requiring scrutiny

For candidate SHA-256 0c05a0d4af9cbe3e70413b250d6c9cac1bfe4d848f6c196f83ed61ebef9ced16, Python 3.12.3 csv.reader(strict=True) parses the header and all 400 data records with exactly six fields.
The unique imo-bench-algebra-036 record has Short Answer exactly $Y(x)=A+\frac{B}{x}-x$.
The candidate equals pinned baseline SHA-256 275877a9d988d85278fad3a5f8a41d7f83393a60bf259531ec0a5161e6b21cf9 plus one inserted byte 0x22 at zero-based offset 13,618.
At the run-3 refresh, upstream main remained 96fa6c4cc3a9bb7450ee7b6773b659d3a030dace, issue #13 was open, and both target-specific PR searches returned zero.

Evidence and scope

Full harness experiment 20260722-165537-7895ed returned 0.
Standalone checker experiment 20260722-165635-489734 returned 0.
git apply --check --reverse artifacts/answerbench-036-one-byte.patch returned 0.
git ls-remote returned upstream main 96fa6c4cc3a9bb7450ee7b6773b659d3a030dace.

Computational experiments

.proof-experiments/20260722-165537-7895ed: full harness returned 0; baseline, wrong answer, and collateral edit were rejected; candidate passed
.proof-experiments/20260722-165635-489734: standalone checker returned 0 and printed both required PASS lines

Independent checker

tools/check_answerbench_036.py is materially separate from tools/baseline_answerbench_036.py: it constructs the only allowed full byte sequence from a unique malformed separator and checks the complete target Problem/tail, while the harness derives the edit from the unique answer token, reparses independently, and generates mutations.

Contribution gate

candidate_eligible

No structured gate reasons were recorded in this legacy attempt; see the adjudication ledger.

Original model outcome: candidate
Public classification: candidate

Cross-domain transfers tested

None recorded.

Established facts

The pinned upstream baseline has 399 six-field and one five-field data record under tolerant parsing; the malformed record is algebra-036.
artifacts/candidate-gate-run3-96fa6c4-0c05a0d4.json · Upstream commit 96fa6c4cc3a9bb7450ee7b6773b659d3a030dace only · computed
The one-byte candidate has 400 strict-parsed six-field records and the exact reported algebra-036 Short Answer.
Experiments 20260722-165537-7895ed and 20260722-165635-489734 · Candidate SHA-256 0c05a0d4af9cbe3e70413b250d6c9cac1bfe4d848f6c196f83ed61ebef9ced16 · computed
The candidate differs from the pinned baseline only by insertion of 0x22 at offset 13,618.
Standalone checker and candidate gate result · The two named SHA-256 byte streams · computed

Ruled out in this epoch

Use semantic-only CSV validation as the complete certificate
Candidates derived from the pinned baseline · It accepts the tested unrelated equal-length problem-ID mutation. · unrelated_edit control in artifacts/candidate-gate-run3-96fa6c4-0c05a0d4.json · A different exact mechanism proves no collateral byte changes.
Search alternative one-byte insertions
All 43,401,728 single-byte insertion scripts for the pinned CSV · The unique target token and CSV boundary derive exactly one candidate, which passes the entire contract. · Both run-3 experiments and the exact patch · Upstream bytes or the authoritative answer literal change.
Re-derive the functional-equation solution before fixing the CSV
This source-defined dataset correction · The authoritative issue and existing raw token already specify the exact literal; the defect is only a missing CSV quote. · Issue #13 and pinned baseline bytes · The authoritative source retracts or disputes the literal.

Open leads

Submit the minimal correction upstream after owner authorization.
All technical gates pass; maintainer review is the remaining acceptance path. · Refresh live source state and rerun experiment 20260722-165635-489734's command immediately before submission. · high · open
Offer a generic CSV schema regression guard only if maintainers request preventive scope.
It could prevent recurrence but is not necessary for the smallest correction. · Run the proposed guard on the pinned and repaired trees; require one failure before and zero after. · low · open

Continuation checkpoint

Objective: Move the certified one-byte correction through authorized upstream review without expanding scope.

First action: python3 tools/check_answerbench_036.py imobench/answerbench_v2.csv --reference sources/96fa6c4cc3a9bb7450ee7b6773b659d3a030dace/answerbench_v2.csv

Stop condition: Stop or redirect on upstream movement, a competing repair, checker failure, owner rejection, CLA blocker, or maintainer scope change.

Next moves

Obtain Charlie's explicit authorization and confirm Google CLA readiness.
Immediately before submission, refresh upstream main, issue state, and competing PR searches, then rerun the standalone checker.
If unchanged, submit only the one-byte dataset patch linked to issue #13; offer a generic schema guard separately only if maintainers request it.

Citations

Tool disclosure

OpenAI Codex acted as the Sol principal in the campaign's GPT-5.6 role; a GPT-5.6 Terra source-discriminator memo was treated only as advisory and every relied-on claim was independently rechecked. Deterministic evidence used Python 3.12.3 standard-library csv/hash/subprocess code, the supplied experiment recorder, git, curl/GitHub API reads, and git apply --check. No CAS, proof assistant, SAT/SMT solver, or external publishing tool was used.; orchestration: gpt-5.6-sol principal with gpt-5.6-terra delegates.

Duration: 694.0s
Review state: needs isolated skeptic review
Attempt ID: deepmind-superhuman-13-answerbench-column-20260722-170424-418499

Human review ledger

accept · Charlie Krug · 2026-07-22 17:42 UTC
Charlie reviewed the candidate record and approves release of the exact one-byte AnswerBench CSV correction for upstream submission.