Verify honestly

Verify Honestly

Match every completion claim to evidence from the context that actually runs the work.

Core rule

Verify in the real execution context, then name the verification level reached. Do not silently substitute an easier context:

• source inspection for execution
• mocks for real data
• local behavior for production behavior
• build success for rendered UI correctness
• configuration text for an external probe
• intended output for the actual artifact

If exact context is unavailable, state what remains unverified and what would verify it.

Verification ladder

Use this vocabulary in completion reports:

1. code read – trace logic without executing it.
2. local tests – run relevant unit or integration checks locally.
3. browser pass – exercise rendered behavior in a browser, including relevant viewports.
4. live data check – query real data or service state consumed by the change.
5. production smoke – exercise deployed behavior end to end.

Name highest applicable level per shipped item. Lower levels need mention only when they explain residual risk. Ladder represents increasing runtime proximity, not a demand to run every level.

Proportionality

Choose cheapest probe that would catch likely failure.

• Scale depth with blast radius, irreversibility, and uncertainty – not session length.
• Treat green deterministic gate as evidence. Cite it instead of re-deriving same fact manually.
• Verify each unchanged property once.
• Prefer one consolidated verification pass over many redundant agents or repeated commands.
• Add independent verifier only when it contributes different evidence, failure lens, or required independence.
• If verification costs more than safely redoing work, say so and choose cheaper path.

Surface matrix

Claim	Minimum useful evidence
Code behavior	Focused executable test or exact reproduction
UI behavior or layout	Browser pass at affected viewport and state
Database or data-shape behavior	Representative real rows or production-shaped fixture; name which
Pipeline or scheduled job	Run same command, environment, and working directory when accessible
Infrastructure state	External probe or observed event, not config grep alone
Artifact size, count, or contents	Measure generated artifact itself
Webhook, email, or integration	Captured real payload or test delivery through actual boundary
Production behavior	Deployed revision plus production smoke

Claims from agents and audits

Treat another agent's RED/GREEN verdict, line citation, or confirmed bug as untrusted evidence until checked against current source.

For large finding sets:

1. Re-check riskiest two or three findings plus one random finding.
2. Escalate to full re-check only when sample disagrees.
3. Anchor fixes to live source, not quoted source in audit report.
4. Re-check current revision when HEAD may have moved during audit.

Reporting contract

Lead with outcome and level:

Fixed; local tests passed: <commands>. Browser pass and production smoke not run.

For partial evidence:

Code read confirms new guard is wired. Runtime behavior remains unverified because <blocker>.

Avoid works, verified, all good, or should be fixed without named evidence.

Completion checklist

• Match every shipped item to named verification level.
• Use execution context that runs property being claimed.
• Measure counts and artifacts directly.
• Re-check agent or audit claims relied upon.
• Flag deploy-gated or inaccessible checks explicitly.
• Keep verification proportional and non-duplicative.
• State exact commands, probes, or artifacts supporting claim.

Stop conditions

• Required surface is inaccessible: stop verification and report exact gap.
• User asks to confirm someone else's result: verify from current state rather than inheriting approval.
• Evidence contradicts expected result: report failure before attempting broader work unless task includes fixing it.

Related skills

• gate-engineering builds durable nets after recurring failures.
• session-close uses this vocabulary in final handoff.
• retro-distill turns verification failures into reusable rules.