Evidence schema

The evidence bundle is a set of JSON / JSONL artifacts written to the run directory. This page lists the artifacts and the fields you will most often read. Treat the in-tree files as the source of truth — the schema is evolving toward a versioned, pydantic-owned schema_version: "1.0".

Schema versioning is in progress

A formal versioned schema (with a schema_version field, a typed claim model, and typed diagnostics) is planned. Until it lands, consume the bundle defensively: read the fields documented here, and tolerate additional keys.

Bundle artifacts

Artifact	Format	Contents
evidence_summary.json	JSON	gate results, claim tier vector, allowed/blocked claims
run_manifest.json	JSON	inputs, artifact paths, run summary, git SHA
events.jsonl	JSONL	append-only stage-by-stage execution trace
benchmark/benchmark_lock.json	JSON	sealed baseline metrics + alignment + hashes
calibration_provenance.json	JSON	candidate → lock → verification authority chain
physical_gates.json	JSON	ET / mass-balance / volume-bias gate decisions
routing_flow_gates.json	JSON	routed-flow mass-closure gate decisions

evidence_summary.json — key fields

Field	Meaning
effective_claim_tier	the granted legacy identifier, ordered exploratory → diagnostic → publication_grade → research_grade
effective_claim_tier_label	public workflow label alongside the legacy identifier in new workflow summaries
claim_tier_label	public label for the requested/allowed tier in new workflow summaries
allowed_claims	claims the gates support
blocked_claims	claims refused, each with a reason and artifact
gate entries	per-gate status with a pointer to the gate artifact

A blocked-claim entry has the shape:

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

calibration_provenance.json — authority

Field	Meaning
baseline	locked baseline metrics
verified	metrics from the independent rerun — authoritative
authority	which stage holds authority (verified_rerun)
candidate_metrics_authoritative	always false — candidate metrics are not reportable

Claim tiers

Tier	Granted when
exploratory	the run executed; no quality claim supported
diagnostic	outputs usable for diagnosis; specific gated sub-claims may hold
Calibration verified (publication_grade)	legacy intermediate tier: calibration and skill checks pass, but complete highest-tier fidelity requirements are not established
Gate-verified (research_grade)	applicable provenance, physical, routing, verified-skill, outlet-scope, sensitivity, soil and land-use fidelity requirements pass for the recorded claim and scope

These public labels describe workflow verification. Neither legacy identifier asserts journal approval or a published performance rating. Moriasi criteria must be assessed separately for the specified variable, time step and period; missing required metrics do not establish a pass.

See The evidence bundle and Reading the evidence.

Separate numeric performance assessment

The dashboard embeds streamflow_performance and streamflow_validation_performance as derived assessments, separate from sealed workflow claims. Their schema is moriasi_2015_streamflow_numeric_v1, with status (met, not_met, not_evaluated), R²/NSE/PBIAS, reference DOI, variable, timestep, period, evaluation role, and reasons. Calibration alignment assessment uses every CSV row before plot downsampling and records its source SHA-256. Reported validation metrics are evaluated only when the complete required metrics and period are present; no R² is inferred from KGE.

Moriasi 2015 watershed streamflow numeric criteria are R² > 0.60, NSE > 0.50, and |PBIAS| ≤ 15%. This numeric screen does not replace graphical and contextual assessment, establish independent validation, or change a workflow tier. Historical artifacts are not rewritten to contain these new fields.