[Bug] False-positive safety fallback triggered by biomedical vocabulary in workspace context

Bug Description
Subject: False-positive safety fallback on a published-statistics OSS repository

Summary
-------
Claude Fable 5 is intermittently flagged and auto-falls-back to Opus while I work
in a specific repository, based purely on the auto-loaded workspace context — not
on my prompts.

Repository
----------
Official open-source implementation of a peer-reviewed paper:
"Disentangling Latent Risk Pathways via Bayesian Hypergraph Inference" (ICML 2026),
public preprint at arXiv:2606.07677. The repo is a numerical-software project that
ports a MATLAB reference implementation to a Python package. The content is
published biostatistics / epidemiology methodology — modeling binary co-occurrence
data (multimorbidity) as a latent hypergraph via Polya–Gamma-augmented variational
inference. It contains no dual-use, wet-lab, pathogen, or otherwise hazardous
material; everything is public academic statistics and linear algebra.

Isolation performed
-------------------

• claude --safe-mode --model fable (CLAUDE.md / skills / MCP disabled): Fable

passes reliably.

• Normal mode with the repo's CLAUDE.md / README auto-loaded: Fable is intermittently

flagged and falls back.
=> The trigger is the biomedical-domain vocabulary in the auto-loaded workspace
context, not any actual request. This is a clear false positive on published,
non-hazardous epidemiological-statistics content.

Request
-------
Please tune / allowlist the safety classifier so that published, citable
public-health statistics implementations (named paper + public arXiv/DOI) do not
trigger the biosafety fallback on workspace context alone. Happy to share the repo
URL and paper for review.

Environment: Claude Code v2.1.170+, Max/Enterprise, non-ZDR.

Environment Info

• Platform: darwin
• Terminal: iTerm.app
• Version: 2.1.218
• Feedback ID: 398048f6-2892-470b-8705-d6bceff68b0e

Errors

[]