[Model behavior] Instruction adherence degrades after switching subscription from Pro to Max (second occurrence)

Summary

After switching the subscription from Pro to Max, the model's adherence to loaded instructions in a long code-review session dropped sharply. This is the second time the same switch has been followed by the same degradation. The written procedure, loaded every session from the repository instruction file and the memory directory, was followed reliably before the switch.

Environment

• Claude Code CLI, Opus-class model, 1M context
• Long code-review session on an open-source project
• A written review procedure loaded every session from the repo instruction file and the memory directory

Observed after the switch (one session)

1. Verdicts issued before the procedure steps were done, then reversed as soon as the user asked "sure?". This happened on five separate PRs.
2. A PR touching threads and background pollers was accepted without reading the project's concurrency module or its reapers, which the procedure requires.
3. A "re-review" consisted only of checking that the head commit had not changed. The 322-line new test file was never read.
4. A change was called "breaking" without checking the main branch, where the behaviour was already broken.
5. A new configuration knob was accepted on the grounds that older knobs exist, despite a loaded rule banning new knobs.
6. A git write was run on an ambiguous "yes please", despite a loaded rule forbidding inferred permission for git writes.
7. Output-length and changelog-length rules were ignored throughout.

Expected

The same loaded instructions produce the same adherence regardless of subscription tier. Every procedure step is done before a verdict is given.

Impact

The user has to push two or three times per PR to get work the loaded procedure already requires. That removes the value of the tool for review work.