26 — Human Review Protocol

26 — Human Review Protocol

When to trigger human review

Trigger review when:

• high chiral tension and high stakes;
• access gaps block strict claims;
• predicate invention proposes high-impact latent context;
• residual contradiction remains high;
• calibration confidence is poor;
• strict claims are impossible but likely claims are decision-relevant;
• critic ensemble deadlocks.

Review packet

A review packet includes:

• input SNOs;
• synthesized SNO;
• Antagonist report;
• proof traces;
• evidence spans;
• access states;
• residual tensor summary;
• latent predicates;
• possible worlds;
• model/run manifest.

Reviewer actions

Reviewer can:

• accept strict claims;
• downgrade likely claims;
• reject unsupported claims;
• mark latent predicate as plausible / unsupported / wrong;
• request evidence collection;
• mark access-state assumptions;
• annotate synthesis quality.

How review affects the system

Human review may be used:

• as post-run annotation;
• as calibration data;
• as training data in future offline runs.

Human review is recorded after runtime unless the run is explicitly marked as a review or retraining step.

Review labels

• accepted
• downgraded
• rejected
• needs_evidence
• access_blocked
• predicate_plausible
• predicate_unsupported
• synthesis_overclaims
• synthesis_preserves_conflict