Contract: Evidence

Schema: domain/schemas/evidence.schema.json. Produced by agent/ tool calls (via investigation_step.result_refs), stored/assembled by evidence/.

What Evidence is (and isn't)

Evidence is a sourced, falsifiable claim, not an interpretation. "Payment of 842000 MXN occurred on 2026-08-13" is evidence. "This looks like money laundering" is an inference built from evidence, and belongs in investigation_step.decision or the final case.hypothesis, never in an evidence record itself.

Example

{
  "evidence_id": "E-001",
  "type": "TRANSACTION",
  "source_type": "BANK_RECORD",
  "source_id": "TX-8841",
  "claim": "Payment of 842000 MXN occurred on 2026-08-13",
  "strength": "DIRECT"
}

Strength levels

• DIRECT — proves the claim outright (e.g. the transaction record itself).
• CORROBORATING — supports another piece of evidence but doesn't stand alone (e.g. two shell providers sharing a registered address).
• CIRCUMSTANTIAL — consistent with the hypothesis but doesn't prove it (e.g. unusually fast payment timing, alone).

A case.status of SUBSTANTIATED should never rest on circumstantial evidence alone — see docs/contracts/case.md.

Provenance requirement

Every tool in docs/contracts/agent-tools.md returns source_ids, and every Evidence record's source_id/source_type must trace back to one of those. If a judge asks "where does that number come from," the answer is always a specific record ID, never "the model computed it."

Ownership

Evidence Collection Subsystem (backend/src/truelock/evidence/exposure.py).