Model Reliability Correction Governance

Multi-model outputs are diagnostic suggestions. They never directly revise an accepted MIPVU decision, annotation, analysis result, prior human-reliability artifact, or publication claim.

Four distinct stages

  1. Suggestion. A disagreement becomes a pending review candidate in cases/<case_id>/quality/model-reliability/review-queue/. It records model values, the reference value, and a bounded question, but no accepted value.
  2. Adjudication. A human reviews source context and the codebook through a separate human-governed workflow. Model consensus is not the adjudicator.
  3. Accepted correction. An authorized human may make a deliberate, separately reviewed edit to a canonical artifact. No model-reliability command performs this step.
  4. Downstream regeneration. Ordinary corpus, analysis, status, audit, and publication pipelines are rerun after the accepted correction.

Enforced write boundary

All model-reliability outputs resolve beneath cases/<case_id>/quality/model-reliability/. Review suggestions are further restricted to the dedicated review-queue/ candidate layer.

The following remain immutable during an end-to-end model-reliability command:

  • metadata/, corpus/, and analysis/;
  • every case-local quality/ artifact outside model-reliability/, including prior reliability results, review packets, and adjudication logs;
  • publication/.

The pipeline snapshots these protected paths before execution. If any stage attempts a direct, traversal-based, or symlink-assisted write, original bytes are restored and the command fails with the affected paths. Output helpers also reject destinations outside the writable subtree before writing.