[Bug] Fable 5 safeguards incorrectly flag documentation projects at session load before user request evaluation

Bug Description

Fable 5 safeguards flag a documentation project on the session-load turn, > before any request is made. > > I maintain a knowledge-management project over a family member's fourteen years > of laboratory notebooks. He is a retired scientist who spent that time > characterising an antibacterial compound produced by a wood-rot fungus. The > project extracts his notebooks, meeting notes and instrument files into > markdown cards and a claim graph, so the progression of his thinking is > navigable and the unfinished threads are recoverable. The deliverables are a > timeline, a lookback, a proposal for future work and a graph explorer. It is > archival and documentation work. No procedure is being designed or run. > > Every attempt to use Fable 5 in this project is flagged and routed to Opus 5. > Across 19 sessions and 6,353 assistant messages, Fable has served zero. There > are four model_refusal_fallback records, on 2026-07-27 and 2026-08-06. > > The specific problem is when it fires. In the session I am sending this > from, my only input was a one-word greeting, and in an earlier one it was a > command that loads project context. There was no biology question in either. > The classifier is reading the project description in CLAUDE.md out of the > system prompt at turn zero, so I cannot ask anything at all before being > routed away. A per-request classifier is defensible; flagging the act of > opening a directory is not, because no request has been made to evaluate. > > Why I believe this is a false positive rather than a close call. The subject > is antibiotic discovery, which is the therapeutic side of microbiology. There > is no pathogen work, no toxin, no gain-of-function, and no synthesis route to > anything hazardous; the assays are disc diffusion against standard laboratory > test strains and the analysis is LC-MS of a fungal culture filtrate. The > science is 2010s bench chemistry that is already done and mostly already > failed. Nothing here offers uplift to anyone, and the words likely to be > carrying the flag ("antibacterial", "antibiotic", "bioactivity", "bacterium") > are the ordinary vocabulary of pharmacology. > > The concrete cost: the project's quality control was designed as a cross-model > audit, with Fable reviewing cards that Opus wrote, on the reasoning that the > writing model is the worst candidate to check itself. That is now impossible. > 690 cards and 87 claim atoms have only ever been reviewed by the model that > produced them, which matters for a project whose entire purpose is fidelity to > a real person's record. > > What I am asking for: narrow the biology classifier so that historical > antibiotic-discovery records do not trip it, and at minimum do not flag the > session-load turn before a user request exists to evaluate. ## Evidence behind the numbers Reproduce with python3 tools/model_routing_status.py. | fact | value | |---|---| | assistant messages in this project | 6,353, all claude-opus-5 | | messages served by Fable | 0 | | model_refusal_fallback records | 4 | | flag timestamps | 2026-07-27 10:00:38Z, 2026-07-27 10:55:36Z, 2026-08-06 10:33:09Z, 2026-08-06 12:10:21Z | | position of the flag | first assistant turn after Fable was selected, in every case | | Fable's behaviour elsewhere | 11,722 messages across 20 other local projects | The notice, verbatim: > Fable 5's safeguards flagged this message. Our intentionally broad safeguards > allow us to deliver more capabilities faster, but can sometimes flag > legitimate coding, cybersecurity, and biology tasks. Switched to Opus 5. Send > feedback with /feedback or learn more: > https://support.claude.com/en/articles/15363606 That article confirms the design: classifiers run on every user request, covering offensive cybersecurity, biology and life sciences, distillation, and frontier LLM development. It names an appeal route for cybersecurity only, the Cyber Verification Program, with no biology equivalent.

Environment Info

• Platform: linux
• Terminal: vscode
• Version: 2.1.223
• Feedback ID: fab87eb8-8f1c-422a-845d-bfc90bc6b032

Errors

[]