[Bug] Safety classifier false-positives on routine code review, forces model downgrade mid-session

Bug Description
Fable 5's safety classifier repeatedly blocks a completely routine engineering task and force-downgrades paid Fable sessions to Opus 4.8, making the work impossible to finish. Task: a correctness / crash-safety code review of MY OWN open-source C inference engine (public repo ariannamethod/yent). Standard software QA — checking that sizes are validated, allocations are checked, array indices stay in bounds. What happens: the classifier reads this as vulnerability/exploit research and pauses the session, offering only "Switch to Opus 4.8." On 2026-07-04 this happened across 7+ consecutive sessions — the task never completes. The flag fires even on empty prompts like "yes, continue" because the accumulated session context carries the review's technical wording. The flag text itself says it is "intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work" — this is exactly that false-positive class. The result: I pay for a Fable audit and cannot get one; every session is redirected to a different model mid-task. Request: narrow the classifier so a developer reviewing the safety of their own code is not treated as an attacker, and do not silently change the model I selected and paid for.

Environment Info

• Platform: darwin
• Terminal: Apple_Terminal
• Version: 2.1.201
• Feedback ID: 3bee9c62-f9a0-4bc3-9d4a-f4d40d6a3945

Errors

[]