Orchestration Pmatch

Procedure

1. Launch dual independent extractors

Start two Task subagents with no shared intermediate context. Each returns:

• extracted claims (id, claim, verification_status, evidence, optional confidence)
• extractor identifier.

2. Build extractor claim sets

Assemble:

• drift_config.mode = "dual-extractor"
• drift_config.extractor_claim_sets with exactly two extractor claim sets.

3. Run runtime adjudication

Call multi-model-review with:

• action.type = "drift-detect"
• document set to source artifact content/type
• drift_config.target_ref set to target artifact path
• drift_config.mode = "dual-extractor"
• drift_config.extractor_claim_sets from Step 2.

4. Lead mitigation pass

For high/critical drift findings:

• attach mitigation action or explicit justification,
• resolve conflicts called out in adjudication.

5. Build artifact

Write drift report to:

• .pipeline/runs/<run-id>/drift-reports/pmatch.json

6. Gate evaluation

Validate:

• schema compliance (contracts/artifacts/drift-report.schema.json)
• no unmitigated critical/high drift findings
• adjudication metadata present (mode, extractors, conflicts_resolved, resolution_policy)
• each claim includes extractor attribution.

Write gate output:

• .pipeline/runs/<run-id>/gates/pmatch-gate.json

Non-negotiables

• No extractor cross-talk
• No high/critical drift accepted without mitigation or explicit human sign-off
• Dual-extractor mode is the default; heuristic mode is fallback-only