When a Result Survives Its Explanation: A Source-Linked Dossier of Errors and Corrections in AI-Assisted Research

An erroneous research statement need not invalidate the calculation it describes. Conversely, an apparently correct result does not establish that its tests, explanation or readiness assessment were sound. This retrospective dossier examines twelve records from five related episodes in three projects of one research program. Nine records have documented artifact chains; three remain qualified as task interpretation, conditional model accounting or insufficient chronology evidence. The records include premature gate approval, ineffective numerical assertions, failure-signaling defects, an inappropriate numerical test, a reporting arithmetic error, unsupported effect and power inference, trajectory overgeneralization, transcript-level provenance contradiction and a specification/implementation mismatch. Several corrections left a primary negative result unchanged as historically reported. A report that prematurely passed a source-point requirement was followed by a separately recorded five-point completion. Corrections, recommendations and qualifications are distinguished. Cases were selected after findings were known; opportunities for error and a complete candidate pool were not enumerated. Codex assisted selection, adjudication and drafting and is finder of record for eight cases, while none attributes an original error to Codex. This asymmetric, exposed collection supports neither model rankings nor an estimated benefit from multi-AI review. The release includes the author-adopted R3 manuscript and a selected-evidence supplement with 73 entries from 26 retained source artifacts. Historical scientific outputs remain audit-reported; no new scientific execution is reported. AI tools assisted curation, drafting and advisory review; known exposure and curator-role asymmetry are disclosed. The author authorized publication on 1 November 2026. The PDF presents the wide case table as case records for readability, with the same text. Full internal source archives and raw scientific outputs are not included.