Errata · When the Defect Is in the Specification: Trajectory-Level Validation of Generated Code

This document corrects two statements in version 1 of When the Defect Is in the Specification (10.5281/zenodo.21879238, 10 August 2026). It does not replace the report: everything not named here stands, and both corrections make the report's claims narrower rather than wider. The first correction is to §8.3. The published text describes the surviving defect as two incompatible vocabularies in one contract, one enumerating two classes and the others six. That does not describe any delivered artefact. In the fabricated contract both enumerations are of six and agree with each other; the enumeration of two existed during the run, before the correction of §5, and the report presented that intermediate state as if it were the contract that shipped. The divergence that does exist is in a different instance and is four against six. The second correction is to §10.1. The pending check is described as the one that would have detected the defect of §8.3. That is false for the delivered contract: where the enumerations agree, a check that compares enumerations across declarations passes without an observation. What survived is a transcription that is consistent and wrong, and no comparison internal to the contract can reach it. The pending item is restated in three parts, the third declared not verifiable from the contract alone. Both corrections come from a measurement taken on 13 August 2026 with a read-only checker over the contracts on the device. It found one real divergence, previously unreported by any verification, and it reports its own coverage alongside: three to four per cent of the declarations inspected yielded a recognised statement, so the zeros measure the reach of the extractor rather than the state of the corpus. A check that inspects little and reports zero is indistinguishable from one that does not work. The divergence found is not documentary. The implementation declares four admissible classes and raises on any other value, while the contract signature declares six. The implementation does not admit the classes the contract declares. Unaffected: the trajectories of Appendices B and C reproduced step by step, the negative control, the proportion of 391 generated statements against 30 hand-written signatures, and the finding that every surviving defect was in the human transcription rather than in the generated code.