[BUG] Fable 5 safeguards over-flag routine health/biology work and reroute mid-turn

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Grading the evidence strength of ordinary health claims (e.g. "does green tea reduce stroke risk," "does aged garlic extract affect arterial plaque") repeatedly trips Fable 5's safeguards and forces a mid-turn switch to Opus 4.8. This is standard nutrition/health-literacy work, no dual-use intent. Two separate turns flagged; pure coding turns in the same session never did, so the trigger appears to be health-efficacy content, not the task.

What Should Happen?

Routine health and nutrition literacy work - grading or discussing the evidence behind everyday health claims - should not trip the safeguards or force a model reroute. If a reroute is ever warranted, the trigger should be narrow enough to exclude ordinary biology/health-literacy content, which the safeguard notice itself admits it currently over-flags.

Error Messages/Logs

Steps to Reproduce

In the Claude Code VSCode extension, set the model to Fable (Fable 5).

In any workspace (no project files or special setup required), send a prompt asking the model to grade the evidence behind several everyday health claims. Minimal example, paste verbatim:

Grade the strength of the scientific evidence behind each of these
health claims and assign each an evidence tier (meta-analysis / RCT /
mechanistic / anecdotal):

1. Drinking green or black tea daily reduces stroke risk.
2. Aged garlic extract reduces coronary artery calcification.
3. Six months of HIIT reduces arterial plaque volume.
4. Phytosterols lower LDL cholesterol.

Let the model start generating its response.

Observed: partway through the turn, a banner appears - "Fable 5's safeguards flagged this message" - and the session reroutes to Opus 4.8 for that turn.

Expected: the model grades the claims normally with no safeguard flag and no reroute. This is routine health-literacy content.

Claude Model

Other

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.193

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

<img width="2447" height="104" alt="Image" src="https://github.com/user-attachments/assets/5e679bed-5ca0-4c16-9377-ed9f04742d66" />