Historical version
This page records design, operation, or validation facts from an earlier release. It is not current V7.17.3 guidance. Start from the documentation hub for current instructions.
V7.13.0 review record¶
Preimplementation data-contract review identified card provenance, phase planning and one-use transaction gaps, then issuer/consumer separation and out-of-order lifecycle gaps. The coordinator recorded resolutions; this does not mean the independent Reviewer returned PASS.
Independent postimplementation review covered frozen-policy isolation, publication/result ownership, atomic budgets, permit replay, unknown receipts, Desktop identity linkage and bilingual delivery. Feedback led to explicit unsuccessful-outcome constraints, evidence-backed retries after blocking repair results, and a separate Astra inflight limit.
The final narrow recheck confirmed the Astra snapshot and atomic-reservation fix. Its concern about arbitrary internal API callers forging not-started proof was adjudicated against the supported ingress: the current Desktop adapter accepts no such proof and exposes no automatic refund command. Raw feedback and the coordinator's scope decision are preserved separately; internal APIs do not claim to authenticate arbitrary local writers.
Implementation review used six actual calls and 32 frozen-V3 proxy units, including one narrow third-round exception without increasing total-call or unit ceilings. This is policy accounting, not billing or OS isolation. Source findings are resolved or explicitly adjudicated; failed native admission and model screening remain separate unfinished work.
The later newline correction affects release staging and related documentation only. It passed 20 bilingual regressions and complete CI without changing model/budget runtime code or adding an independent Reviewer. Compatibility support retains the old default; source review is not evidence of GPT-6 quality or native-budget acceptance.
Source, package, installation, loading, dispatch, commit, push and public publication need separate confirmation. Model quality and reasoning gains retain empirical gates; fixtures and model names cannot replace them.