End-to-End Safety Gate

Lifecycle

1. pre-gen - run the lesson 83 detector on the prompt
   • if confidence >= block_threshold: return refusal, emit trace, stop
2. during-gen - stream from the model, buffer two chunks, scan for known harmful continuations
   • if matched: terminate iterator, mark trace, treat as medium severity
3. post-gen - if no early termination, run the lesson 85 classifier router and the lesson 86 rules engine on the completed output
4. aggregate - take the maximum severity across pre, during, post.classifier, post.rules
5. apply - map to block, redact, warn, or allow

Aggregation table

Signal state	Action
any high severity	block
any medium severity	redact
any low severity	warn
nothing	allow

Trace structure

RequestTrace
  request_id: str
  prompt: str
  pre_gen: { category, confidence, fired[] }
  during_gen: { terminated_early, matched_pattern, partial_chunks }
  post_gen: { classifier_action, classifier_severity, rules_max_severity, rules_violations[] } | null
  final_action: block | redact | warn | allow
  final_output: str
  latency_ms: float

Artifact

outputs/gate_trace.json contains the summary and one trace per request, including 50 taxonomy fixtures and 10 benign prompts.