[Model behavior] Verdicts issued before the user's written review procedure was completed

Under a high volume of pull requests, the model repeatedly gave accept verdicts before finishing the mandatory review steps loaded at session start: it read only the code part of diffs (filtering out tests), skipped review threads and review comments, and skipped caller searches and duplication checks. The steps were back-filled only after the user asked "are you following the procedure?". The procedure explicitly says the complete review must precede the verdict.

Model: Claude Opus 5.5 (claude-opus-5-5[1m]), Claude Code CLI, one long working session on 2026-10-02.
Context: maintainer of an open-source cloud-service emulator; the work is pull-request review and fixes under a written procedure loaded into every session (project instruction file + persistent memory). No personal data, repository names, PR numbers or handles are included.

Raised with Claude Code.