Verify Before Claim

Verify Before Claim

<skill_contract> <input>One falsifiable claim, its artifact or system, risk, expected result, permissions, and available checks.</input> <output>A scoped claim decision with fresh evidence, residual risk, approval, and rollback status.</output> <done>The cheapest direct check after the final material change supports the exact allowed wording.</done> <non_goals>Producing the artifact, inferring success from effort, or widening a claim beyond checked evidence.</non_goals>

No evidence, no claim. Match the check to the exact claim, use fresh evidence, and keep execution authority separate from approval authority for consequential actions.

Usage Template

Provide: proposed claim, artifact or system, risk level, available checks, expected result, permissions, and rollback path.

Workflow

<intake>

1. Rewrite the proposed statement as one falsifiable claim.
2. Classify risk: low, material, or consequential.
3. Select the cheapest check that directly tests the claim.
4. Record expected signal, time boundary, and acceptable evidence.

</intake>

<unknowns_gate>

If the artifact, expected behavior, or verification method is missing, return NEEDS_INPUT. If only indirect evidence exists, return INSUFFICIENT_EVIDENCE or narrow the claim; never fill the gap with confidence language.

</unknowns_gate>

<execute>

Run the selected check after the final material change. Examples: targeted test, lint, build, link check, read-after-write, diff inspection, source comparison, dashboard query, or deployment health check. Capture command/query, timestamp, scope, exit status, and key output.

Treat runtime termination as control evidence, not task evidence. A normalized complete signal still needs the acceptance check; a tool request needs a host execution receipt; truncation needs a checkpoint; refusal/error/unknown cannot support completion. A claimed NO_OP needs a fresh eligibility query, an output-count check, and proof that no prohibited side effect occurred.

For material or consequential claims, obtain independent verification from a separate check, reviewer, or evidence source. Require human approval before irreversible publication, deployment, spending, deletion, credential use, or policy change. Confirm the rollback path before execution.

For Graph claims, verify static contract integrity, every required node and join receipt, terminal acceptance, budgets, permission/compensation state, and checkpoint identity. Passing nodes do not prove the end-to-end graph.

</execute>

<evaluate>

Compare observed versus expected result. Decide supported, partially_supported, unsupported, or blocked. Check scope: passing one test cannot prove the full suite; a successful write cannot prove link integrity. State residual risk and evidence age.

</evaluate>

<retry_policy>

max_attempts: 2. Retry only after diagnosing the failure and changing input, tool, scope, or strategy. Stop on repeated signature or NO_PROGRESS; never rerun an unchanged check to manufacture confidence.

</retry_policy>

<state_contract>

Persist {run_id, status, attempt, budget, evidence, unknowns, last_error, next_action} plus claim, risk, check specification, expected/observed results, approval receipt, and rollback readiness. Append verification events so evidence age remains visible.

</state_contract>