Refresh upstream status, materialize the isolated closing quote in imobench/answerbench_v2.csv, and certify strict schema, exact target value, and one-byte-only equality against the pinned source.

CSV structural validation and minimal-patch certification

A standards-library parser checks every record; the issue-supplied token fixes the boundary; an independently written target-aware checker enforces exact fields and baseline-plus-one-byte equality.

Hypothesis: Against unchanged upstream commit 96fa6c4cc3a9bb7450ee7b6773b659d3a030dace, the materialized file is exactly the derived one-byte quote insertion and restores six strict-parsed columns plus the exact algebra-036 Short Answer.

Test: Run the materialized candidate through the strict baseline/control harness and the separate checker with --reference; require baseline rejection, candidate acceptance, exact-byte equality, and rejection of wrong-answer and collateral-edit controls.

The authoritative issue supplies the malformed boundary and intended token; the unchanged hash-pinned source reproduces the defect; two deterministic checking paths agree on the candidate; exact byte comparison excludes collateral edits.

• The materialized candidate is 169,538 bytes with SHA-256 0c05a0d4af9cbe3e70413b250d6c9cac1bfe4d848f6c196f83ed61ebef9ced16.
• It differs from upstream commit 96fa6c4's CSV only by insertion of a double quote at zero-based offset 13,618.
• All 400 data records strict-parse to six columns and imo-bench-algebra-036 has exact Short Answer $Y(x)=A+\frac{B}{x}-x$.
• The checker rejects the malformed baseline, a wrong-answer mutation, and a collateral-edit mutation when --reference is used.

• Experiment 20260722-163326-799c0d returned 0 in 0.531 seconds under Python 3.12.3 with a 256 MB limit.
• artifacts/release-audit-96fa6c4-0c05a0d4.json records baseline widths {5:1,6:399}, candidate strict success, exact hashes, target value, and all controls.
• python3 tools/check_answerbench_036.py imobench/answerbench_v2.csv --reference sources/96fa6c4cc3a9bb7450ee7b6773b659d3a030dace/answerbench_v2.csv returned two PASS lines.
• py_compile and git diff --check returned zero; no additional repository test configuration was present.

• .proof-experiments/20260722-163326-799c0d: materialized release audit returned 0; 400 six-column rows, exact target, exact one-byte diff, and all controls passed.

tools/check_answerbench_036.py is separately written from tools/baseline_answerbench_036.py. It asserts the exact header, full target Problem and tail, and constructs the only accepted candidate directly from a raw-byte separator pattern; the harness separately strict-parses and derives the edit from the unique answer token.

internal_result

• No formal-kernel, repository-CI, external-expert, or independent-third-party validation was recorded.

Original model outcome

progress

Public classification

progress

• Minimal-patch certification -> exact baseline comparison should gate dataset corrections -> the collateral mutation passed semantic checks but failed --reference.
• Strict input validation -> malformed quoting may fail before width validation -> the baseline raised a strict CSV error while tolerant parsing exposed the unique five-field row.

• The workspace candidate equals the pinned baseline plus one 0x22 byte at offset 13,618.
  release audit and exact-diff checker · Pinned upstream commit 96fa6c4 and materialized candidate hash 0c05a0d4 · computed
• All 400 candidate data rows strict-parse to six fields and the target Short Answer is exact.
  artifacts/release-audit-96fa6c4-0c05a0d4.json · imobench/answerbench_v2.csv SHA-256 0c05a0d4af9cbe3e70413b250d6c9cac1bfe4d848f6c196f83ed61ebef9ced16 · computed

• Perform further search over alternative edits
  Single-byte insertion repairs of the pinned CSV · The unique token and CSV boundary derive one candidate, and it satisfies every contract clause. · release audit plus exact-diff checker · Upstream bytes or the authoritative requested literal change.
• Use semantic-only CSV validation as the contribution certificate
  Candidates against the pinned baseline · It accepts the tested unrelated problem-ID mutation. · unrelated_edit control in the release audit · A different checker proves full-file equality outside the target edit.

• Seek upstream acceptance after owner authorization.
  Technical verification is complete; maintainer review is the remaining acceptance path. · Refresh upstream source and PR status, rerun the checker, then prepare the one-byte PR packet if authorized. · high · open
• Offer the generic CSV schema guard only if maintainers want a preventive test.
  It may prevent recurrence but expands the patch beyond the smallest correction. · Ask maintainers during review whether they prefer a separate generic guard. · low · open

Objective: Move the certified correction through human authorization and 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 any upstream movement, competing repair, checker failure, owner rejection, or maintainer request that changes scope.

• Ask the human owner to approve the exact research scope and confirm Google CLA readiness.
• Immediately before any authorized PR, refresh upstream main, issue #13, and competing PR searches and rerun the exact-diff checker.
• If unchanged, submit a minimal upstream PR linked to issue #13 with the verification transcript; await maintainer review before claiming acceptance.

• https://github.com/google-deepmind/superhuman/issues/13 ↗
• https://github.com/google-deepmind/superhuman/commit/96fa6c4cc3a9bb7450ee7b6773b659d3a030dace ↗
• https://github.com/google-deepmind/superhuman/blob/96fa6c4cc3a9bb7450ee7b6773b659d3a030dace/imobench/answerbench_v2.csv#L150 ↗
• https://github.com/google-deepmind/superhuman/blob/96fa6c4cc3a9bb7450ee7b6773b659d3a030dace/CONTRIBUTING.md ↗

GPT-5.6 Sol served as principal investigator. A GPT-5.6 Terra delegate provided advisory reconnaissance only; Sol independently refreshed the GitHub primary source and PR status, inspected and extended the harness, applied the dataset edit, and ran all decisive checks. Deterministic tools were CPython 3.12.3 standard-library csv/hashlib/subprocess/tempfile, Git, curl, GitHub API, py_compile, diff checks, and the computational-researcher run_experiment recorder. No CAS, solver, proof assistant, package installation, cloud lab, or model agreement was used as validation.; orchestration: gpt-5.6-sol principal with gpt-5.6-terra delegates.

Duration

630.5s

Review state

internal result; not a contribution candidate

Attempt ID

deepmind-superhuman-13-answerbench-column-20260722-164036-934de3

No human review recorded.