Multi-Agent Evidence Harness
Umbrella LayerThe Multi-Agent Evidence Harness checks AI outputs and agent actions before production. It prepares context, scans for risk, challenges findings, removes duplicates, seeks proof, repairs weak outputs and decides whether to publish, review, block or return no-answer.
Zij bouwen de modellen. Coherence Engine bestuurt het bewijs, het gedrag en de beslissingen eromheen.
They build the models. Coherence Engine governs the evidence, behavior and decisions around them.
The harness organizes existing governance modules — Model & Context Governance, scanOutput Reliability, Decision Structure Intelligence, AI Behavior Trace, Long-Running Agent Governance, Proof/Evidence Trace, Repair/Review/Block logic — into one clear enterprise pipeline: Prepare → Scan → Debate → Deduplicate → Prove → Repair → Decision.
No modules are duplicated. Each stage links to the existing system component that performs that function.
Context preparation, modality detection, provider selection, reasoning effort assignment.
Risk detection, hallucination checks, provider route checks, cost and context anomaly detection.
Decision structure intelligence, verifier logic, model disagreement signals, counter-checking.
Repair item grouping, duplicate finding handling, repeated issue clustering.
Proof codes, evidence trace, source/retrieval proof, provider trace proof, routeability proof.
Repair/retest logic, rewrite/limitation/block logic, fallback routing, verifier escalation.
Allow, publish with limitations, review, repair, block, no-answer.