[MODEL] Fable 5 safeguard classifier false-positives on physics-based human modeling research, silently falls back to Opus 4.8

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

I selected Fable 5 as my model and was doing computer vision research on physics-based human body modeling — reviewing the DuoMo paper and related human-motion-reconstruction methods (center of mass, ground reaction force, pressure-mat data).

What Claude Actually Did

1. Fable 5's safeguard classifier flagged the session for "cybersecurity or biology topics"
2. The session silently fell back to Opus 4.8
3. The classifier appears to pattern-match vocabulary like biomechanics, center of mass, pressure, and human mesh onto the biology category
4. Even a follow-up meta-conversation purely about how to report this false positive was itself flagged
5. /model could not restore Fable 5 within the same session

Expected Behavior

Fable 5 should have handled this request directly. Physics-based human modeling and motion reconstruction is standard CV / graphics / animation research, not life-sciences or bioweapon-adjacent work, and does not fall into any of the four blocked areas (offensive cyber, biology/chemistry, distillation, frontier LLM dev). When a fallback does fire, /model should be able to restore the user-selected model in the same session.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Set model to Fable 5
2. Discuss physics-based human body modeling / biomechanics-adjacent CV research (center of mass, ground reaction force, pressure mats, human mesh)
3. Observe the safeguard classifier flag the session and fall back to Opus 4.8
4. Run /model to select Fable 5 again → it keeps Opus 4.8

Claude Model

Other

Relevant Conversation

The switch notice shown: "Fable 5's safeguards flagged this message. Switched to Opus 4.8."
The work being discussed was purely academic CV research on human motion reconstruction. No offensive security, no wet-lab biology, no bioweapon-relevant content.

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.201 (Claude Code)

Platform

Anthropic API

Additional Context

The switch notice itself states these safeguards "may flag safe and routine coding, cybersecurity, or biology work," so false positives are expected by design. But there is currently no way to override a flagged switch back to Fable 5 within a session. This makes Fable 5 hard to use for legitimate physics-based human modeling and biomechanics-adjacent CV research. Requesting: (1) narrow the biology classifier so this category of research isn't caught, and (2) make /model authoritative so users can restore their chosen model after a false-positive switch.