# Evidence: a successful response can describe the wrong operation

Ayobami John Haastrup (ayobamih) · 26 October 2026

Article: https://aiworkaccountability.com/writing/a-successful-response-can-describe-the-wrong-operation/

This is a retrospective of two repairs in the author's Agent Shop application. It is not a new reproduction, independent audit or current production acceptance.

## Files

- [case.json](https://aiworkaccountability.com/evidence/semantic-success/case.json): claim map, selected PR metadata, historical run/job observations, reported head and checked-out revisions, interpretations and limits.
- [ci-excerpts.txt](https://aiworkaccountability.com/evidence/semantic-success/ci-excerpts.txt): selected lines from three decoded GitHub Actions job logs retrieved on 26 October. Omissions are intentional; unrelated environment, setup and account information is excluded. Source job URLs are retained.
- [sources.json](https://aiworkaccountability.com/evidence/semantic-success/sources.json): pinned source URLs and SHA-256 of the complete UTF-8 source text retrieved through the GitHub connector. Hashes are content checksums, not Git object IDs or independent attestations.
- [citation.bib](https://aiworkaccountability.com/evidence/semantic-success/citation.bib): citation for this author-published case study.
- [SHA256SUMS.txt](https://aiworkaccountability.com/evidence/semantic-success/SHA256SUMS.txt): byte checksums of the five files above, excluding the checksum file itself.

## What was checked

The review inspected both PRs, source before and after the relevant repairs, three run records, their verify-job and browser-step conclusions, and decoded historical job logs. For the two PR builds, the checkout logs identify synthetic test-merge commits distinct from the reported head SHA. The source links use those tested revisions. The 12 October push build checked out the actual merge commit.

The initial POST traceback proves failure of the expected status >= 400 check, not the exact returned status or body. The dated validation report supplies the HTML/200 diagnosis. The method matrix checks twelve combinations, not twelve independent incidents. It asserts 405 and Allow: GET, but not the rejection content type. The completion case has a before/after source record and passing merged browser run; no separate failing pre-repair browser trace was recovered.

The completion script's terminal banner says "no price" although its source expects a planned price on the page. Treat the specific assertions as the evidence, not the banner's wording. Historical deployment claims remain labelled as author reports. No current Agent Shop deployment or application replay was tested for publication.

## Inspect the claims

Follow each claim's source identifiers in case.json. Compare the pinned code and tests, then the relevant timestamped CI excerpts. The sources manifest permits later comparison with the retrieved file bytes. A live public endpoint may serve a different revision and cannot substitute for the historical fixture.

This package contains no private workbook URLs, private correspondence, credentials or whole CI environment dumps. The article describes AI assistance and authorship explicitly. No AI-causation, representative failure rate, novelty or customer-impact estimate is claimed.