Fable 5 biology classifier false-positive on game character creator work (rig/anatomy wording), silent downgrade to Opus 4.8 that /model can't undo

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 was working on the character creator for my voxel RPG sandbox game (Rust/Bevy) in Claude Code. I asked Claude to help design the avatar system: skeletal rig, skinned mesh, morph targets, body proportion sliders. My prompts and design docs described the goal as an "anatomically correct" "simulated body" for the game characters. It's a game character rig, nothing to do with actual biology.

What Claude Actually Did

1. The dual-use safety classifier flagged the session as "biology" and switched the model from Fable 5 to Opus 4.8 mid-session.
2. Running /model to pick Fable 5 again did not work, it reported keeping Opus 4.8 for the session.
3. The flag is sticky in context. Re-opening the same design doc in that session re-tripped it. The only way out was starting a fresh session and rewording everything to rig/morph vocabulary before touching the avatar files.
4. There was no way to appeal or override even though the banner itself admits it may flag safe, normal content.

Expected Behavior

Game character creator work should not trip a biology dual-use flag. Words like "anatomically correct" next to rigs, meshes and morph targets is standard game dev language.

If a switch does happen, an explicit /model selection by the user should restore the model, and the notice should say clearly whether it's reversible. Right now it silently costs me the model I'm paying Max for, for the rest of the session, with no recourse.

Files Affected

None. No files were touched incorrectly. The problem is the model switch itself.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

1. Start a Claude Code session on Fable 5.
2. Ask for help designing a game character creator and describe the avatar as an "anatomically correct simulated human body" (proportions, muscles, skeleton).
3. The classifier banner appears and the session switches to Opus 4.8.
4. Run /model and select Fable 5. It stays on Opus 4.8.
5. Any later message in that session that references the same design doc re-trips the flag, so the session is burned.

Claude Model

Other

Relevant Conversation

Banner said the content may involve cybersecurity or biology topics and that the session was switched to Opus 4.8, and admitted it may flag safe, normal content. After /model: "Kept model as Opus 4.8".

Impact

High - Significant unwanted changes

Claude Code Version

2.1.209 (Claude Code)

Platform

Anthropic API

Additional Context

Related open issues: #74306 (same classifier false-positive on physics-based human modeling research, silent Opus 4.8 fallback, /model can't restore), #67246 (/model does not restore Fable 5 after a classifier switch), #74278 (flag carries via accumulated session context), #66605 and #66916 (biology-bucket false positives on benign scientific work). This report is another variant of the same bucket: game character/avatar work using anatomy vocabulary. Between #74306 (research) and this one (game dev), the classifier is hitting anyone who describes human bodies in technical language. Happens since the 2026-07-01 classifier redeploy.

Platform: Arch Linux (kernel 6.18 LTS), Claude Max 20x subscription.