Module refusal 

Heuristic refusal classifier.

Populates message.refusal on the blocking chat-completion path so safety-aware clients (OpenAI Python SDK, Vercel AI SDK) that branch on message.refusal != null see the expected shape. Atlas does not train safety behavior into its models — this detector only recognizes the text the underlying model emits when it declines to answer.

Honest scope: a prefix-matcher, not a safety classifier. It catches the common refusal openings (“I cannot help with…”, “I’m sorry, but I can’t…”, “As an AI…”) that instruction-tuned models produce. It will miss subtle refusals and will false-positive on content that quotes a refusal. Clients that need real safety-classification should run their own moderation pass — /v1/moderations is a 501 stub on this server.

Set ATLAS_DISABLE_REFUSAL_DETECTION=1 to force refusal: None on all responses, matching pre-PR-4 behavior byte-for-byte.

Functions

detect

Returns the refusal sentence when content opens with one of the known patterns, else None. The returned sentence is the first sentence (truncated at ., ?, or !) with trailing whitespace trimmed. When the kill-switch env var is set, always returns None.