Reading the evidence¶

A run is only as useful as your ability to read what it claims. This page is a practical walk-through of the bundle, in the order you should actually read it.

1. Open `evidence_summary.json` and read `blocked_claims`¶

Start with what the system refused to assert. Each blocked claim has a typed reason and an artifact pointer:

{
  "blocked_claims": [
    {
      "claim": "terminal_scope_claim",
      "reason": "outlet_scope_volume_mismatch",
      "artifact": "routing_flow_gates.json"
    }
  ]
}

If the claim you care about (say, "research-grade skill at the gauge") is in blocked_claims, the reason code tells you which gate stopped it — and the artifact tells you where to look.

2. Read `allowed_claims` and the tier¶

{
  "effective_claim_tier": "exploratory",
  "allowed_claims": ["model_executed", "calibration_improved_baseline"]
}

The tier is computed from the gate table, not asserted. "calibration improved the baseline" and "the model has research-grade skill" are separate claims that gate independently — a run can be allowed the first and blocked on the second.

3. Confirm the metrics are verified, not candidate¶

Cross-check calibration_provenance.json. Only metrics tied to the verification rerun are authoritative:

{
  "baseline": {"nse": -0.015, "kge": -0.177},
  "verified": {"nse": 0.153, "kge": 0.314},
  "authority": "verified_rerun",
  "candidate_metrics_authoritative": false
}

If a number you want to cite is a candidate metric, it is not reportable as a final result. See Locked calibration.

4. Follow the gate artifacts¶

physical_gates.json and routing_flow_gates.json carry the pass/fail detail behind the claim decisions — ET partition, mass balance, volume bias, routed mass closure. When a claim is blocked, this is where the justification lives.

5. Check provenance for reproducibility¶

run_manifest.json carries inputs, artifact paths, and the git SHA; evidence_summary.json carries provenance hashes. Cite the repository and the run's provenance hash when reporting.

A reading checklist¶

[ ] read blocked_claims — is my target claim blocked, and why?
[ ] read allowed_claims and effective_claim_tier
[ ] confirm any cited metric is from the verified rerun, not a candidate
[ ] open the named gate artifact for any claim I rely on
[ ] record the git SHA + provenance hash with the result

The one-line rule

Reporting a metric without disclosing a failed gate is overclaiming — whether a human or an agent does it.