Fable 5 '(bio)' false-positive declines on biology-metaphor project context and AI-assurance prose

Summary. claude-fable-5 repeatedly declines benign knowledge-work prompts with a "(bio)" classifier tag, auto-falling back to claude-opus-4-8. The fallback works as designed; the declines are the issue. Five occurrences observed on 2026-07-18 (HKT morning) on one machine, all via the Claude Code CLI spawned by the @agentclientprotocol/claude-agent-acp 0.59.0 adapter (ACP), Max subscription auth, macOS Apple Silicon.

Controlled reproduction (three calls, same adapter):

1. Prompt Reply with exactly: PROBE OK, session cwd /tmp → answered normally, no decline.
2. Identical prompt, session cwd = an Obsidian vault whose project context uses cell-biology naming throughout (directories and docs named germline, epigenome, chromatin, ribosome, etc.) → declined "(bio)", fell back to Opus, then answered.
3. Prompt asking whether an email sentence about AI monitoring tooling should be tightened ("a strong T-shirt score cannot be mistaken for demonstrated control effectiveness"), cwd /tmp → declined "(bio)", fell back, then answered.

Inference. Two independent triggers, neither biological: (a) loaded project context dense with cell-biology metaphor vocabulary; (b) AI-governance and assurance phrasing ("monitoring", "control effectiveness"). Both are core daily material for AI-risk consulting work, so Fable sessions in this environment decline structurally, adding a first-message stutter and routing work addressed to Fable onto Opus.

Ask. Tune the bio-risk screen to distinguish biology-as-naming-metaphor and AI-assurance vocabulary from actual biosecurity content, or weight surrounding intent more heavily. Full ACP transcripts of the three-call matrix available on request.