verification

name	verification
description	Verification before completion — evidence before claims, always

Verification Before Completion

Evidence before assertions. No completion claim without fresh verification output.

Quick Reference

Aspect	Detail
Scope	Every status/completion claim before it leaves the agent
Input	A completion claim about to be made (tests pass, build succeeds, bug fixed, etc.)
Output	The claim restated with evidence (command + output), OR a corrected status with evidence
Core rule	Run the proving command → read its full output → cite evidence in the claim

The Gate Function

Before making any status claim, execute these steps in order:

1. IDENTIFY — What single command proves this claim?
2. RUN — Execute it fresh: no --filter, no subset, no cached result.
3. READ — Check exit code. Scan output for failure/error indicators. Count failures.
4. VERIFY — Does the output confirm the claim?
   • YES → State claim with evidence (command run + key output).
   • NO → State actual status with evidence. Do not soften.
5. CLAIM — Only now may you assert completion.

Before / After

Bad — claiming without evidence:

I've fixed the failing test. The build should be green now. Moving on to the next task.

Good — Gate Function applied:

Ran `bun test src/auth.test.ts` — 14/14 passing, exit 0.
Ran `bun run build` — compiled successfully, exit 0.
Auth fix verified: original symptom (401 on refresh) no longer reproduces.

Bad — partial verification presented as full:

Ran the linter — 0 errors. Build succeeds.

(Linter passing does not prove the build succeeds. Two different commands.)

Good — each claim has its own proof:

Ran `eslint src/` — 0 errors, 0 warnings.
Ran `bun run build` — exit 0, bundle output at dist/.
Both linter and build verified independently.

Common Failure Patterns

Claim	Requires	Not Sufficient
Tests pass	Test command output showing 0 failures	Previous run, "should pass"
Build succeeds	Build command with exit 0	Linter passing, "logs look good"
Bug fixed	Original symptom reproduced and now passes	Code changed, assumed fixed
Regression test works	Red-green cycle: test fails without fix, passes with it	Test passes once
Agent completed task	VCS diff shows expected changes	Agent self-reports "success"
Requirements met	Each requirement checked against output/behavior	"Tests passing" (tests may not cover all requirements)

MUST DO / MUST NOT DO

MUST DO	MUST NOT DO
Run the exact proving command fresh before claiming	Rely on a previous run or memory
Cite command + output in the claim	Use hedging words: "should", "probably", "seems to"
Verify each distinct claim with its own command	Substitute one verification for another (linter ≠ build)
Check agent work via VCS diff, not agent self-report	Trust agent/tool success messages without checking
State actual status when verification fails	Express satisfaction ("Done!", "Perfect!") before verifying

Final Checklist

Before any completion claim, confirm:

• Proving command identified and executed fresh
• Exit code checked
• Output scanned for errors/failures — count is zero
• Evidence (command + key output) included in the claim
• Each distinct claim verified independently
• If delegated to agent: VCS diff reviewed, not just agent report