Evidence ledger

FP: Evidence Ledger

Every completion claim needs evidence. The canonical machine record is JSON v1 validated by contracts/evidence-ledger.v1.schema.json; Markdown is only a human view.

Use When

• A medium, debug, live-system, multi-agent, or risky task executed.
• A task failed, blocked, exceeded scope, or needs a durable handoff.
• Metrics or adaptive improvement will consume the result.

Small clear changes may use the Tiny Evidence form instead.

Procedure

1. Record task ID, authority mode, write authorization, and scope.
2. List files read and touched using normalized repository-relative paths when possible.
3. Record each command/check with exact result and evidence; do not equate “ran” with “passed.”
4. For debug work, record hypothesis, probe, actual result, decision, and next step. Before completion, add debug_evidence with the causal chain, first divergence, condition-based wait evidence or a justified fixed wait, and distinct direct/sibling observed checks for any touched shared boundary.
5. Tie every verified claim to evidence and list every remaining unverified claim.
6. Record scope and context-budget violations explicitly.
7. Record remaining risk, decision, and next action.
8. For remote/stateful work, record baseline, rollback, effective runtime, external-client and negative-control evidence.
9. For externally retrieved context, record source, version, structured freshness basis, authoritative page, claim, and unresolved gap.
10. For incomplete cross-session work, compile a continuation block; never use it to auto-replay writes.
11. When a deliberate shortcut is accepted because it has a known safe limit, add deferred_items with its location, shortcut, ceiling, observable upgrade trigger, evidence, and an evidence_ref that resolves to observed command or verified-claim evidence. Do not create debt entries for vague future work or use them to excuse a current acceptance failure.
12. For multi-agent work, record the parent authority ceiling and every delegation envelope, including allowed resources, read-only status, summary budget, artifact path, and observed proposal evidence references.

Completion Gate

• result=pass needs at least one passing check and one evidenced claim.
• decision=complete needs result=pass.
• Required checks come from the Execution Brief and must all pass.
• Report-only work cannot touch files or claim write authority.
• Unknown values stay unknown/null.
• A stale live-system observation is historical evidence, not current state.

Output

Use templates/evidence-ledger.md. Run:

node scripts/lint-contracts.js --ledger <ledger.json> --brief <brief.json>

If the run needs metrics, pass the canonical JSON and its brief to metrics. Use adaptive improvement only when the ledger contains evidence for a reusable change.