[Bug] Anthropic API Safety Filter False Positives on Mathematics Research Contextfee
Status Open
Reported on v2.1.217
Maintainer reply None cached
Activity 0 comments · opened Jul 22, 2026
Bug Description
Fable 5 dual-use safeguard — false positives on legitimate mathematics research. Working on a formal treatise (Lean 4 / information theory) on the mathematics of bounded intelligence. Fable 5's safeguard intermittently blocks main-loop turns — including neutral ones — apparently scoring the accumulated research context (vocabulary like "intelligence," "critical brain size," "sentience"). Fable subagents with narrow context are unaffected; only the full-context main loop blocks. Request IDs: req_011CdHqSG87Y8tDah8BbADtM, req_011CdHqnFDMMLLdTwC9UDp8h, req_011CdHr9YPJfa9Dc16xT3hHW, req_011CdHsV1rwQPCEru8HVfECU, req_011CdHxkhMypWepcqTwR25KK.
Environment Info
- Platform: darwin
- Terminal: vscode
- Version: 2.1.217
- Feedback ID: 8b5680c5-4bfd-4dde-be09-63d97395211d
Errors
[]