Auditable threshold · EON AI Ventures
Safe vs unsafe is a published policy, not a vibe.
A named steward must own these numbers. Until POLICY_OWNER is set in the host secret store, the steward stays pending — we do not invent a person.
- Policy ID
- compound-iq.classifier.v1
- Version
- 1
- Steward role
- Compound IQ classifier steward
- Steward
- Pending appointment — set POLICY_OWNER in the host secret store
- Safety frames
- OSHA · API · ATEX
- Min occurrences
- 5
- Min unique workers
- 3
- Min unique shifts
- 2
Unsafe if any of: Assess IQ intervention fired, hold point skipped, PPE non-compliant, a control was removed, or Assess IQ hinted safety_violation. Unsafe never writes the twin — it becomes a micro-lesson and a supervisor alert.
Safe if it is not unsafe, meets the three repetition thresholds, and (on a safety-critical step) adds a control, sample point, step, or dwell. Then it is a twin + SOP refinement, Verdict-gated.
Insufficient evidence otherwise. Hold. The AI does not get the final say on a thin sample.