FAC quality assessment: an accidentally independent AI translation and blind comparison

Why this record exists The project orchestrator did not know that Piotr Achinger and Łukasz Krupa had already translated Jean-Pierre Serre's Faisceaux algébriques cohérents (FAC). An OpenAI Codex production session therefore translated FAC from the French source and had completed, compiled, and source-checked nos. 1–79 before the existing English PDF—and later its editable source—was discovered. That accidental chronology created a genuinely blind overlap rather than a benchmark designed after the fact. What was checked All 79 blind-overlap units were compared and every disagreement was adjudicated from the 1955 French journal authority, not by agreement between English versions. The deposit contains 79 unit reviews, 138 locator-bound findings, 95 frozen input identities, the blind and complete Codex readers, editable project source layers, the full decision logbook, a 219-row self-correction ledger, and a reproducible package validator. Twenty units had substantive agreement without a material finding; 59 had one or more recorded differences. No comparator finding required changing the admitted Codex English or corrected French layer. What this means This bounded result is evidence that a carefully source-controlled contemporary model workflow can produce mathematically useful translation rather than opaque unchecked prose. It gives readers a concrete reason to inspect and use the related SGA/EGA outputs. It is not certification: no mathematician has peer-reviewed this translation, mistakes may remain, neither English version is authority, and citation-critical passages should be checked against the French source. The 138 findings are heterogeneous and are not a scalar score. Provenance The initial French working transcription came from a Claude Opus 5 project lane and was later checked and corrected against the journal authority. The English translation, source alignment, French correction layers, and blind-comparison adjudication were produced in an OpenAI Codex 5.6 session using Ultra reasoning mode, according to project/runtime records. Hosted-model labels are provenance assertions; artifact hashes establish the exact compared byte states and chronology boundary. Open first: 00_READ_ME_FIRST.md, then 05_FAC_Quality_Assessment_Report.md. The Achinger–Krupa PDF and source are credited and hash-bound but not redistributed because no explicit redistribution license was found in the acquired source tree.