Gate validation discipline

Gate Validation Discipline

Scope

Completion verification, evidence examination, citation requirements, gate verdicts. Does NOT generate evidence (use functional-validation) or plan validation (use create-validation-plan).

The Rule: EVIDENCE BEFORE COMPLETION

NEVER mark any gate, task, or checkpoint complete until you have:

1. PERSONALLY examined the evidence — not just received a report about it
2. CITED specific proof — file paths, line numbers, exact output, screenshot descriptions
3. MATCHED evidence to criteria — each validation criterion has corresponding proof
4. ANSWERED "Would a skeptical reviewer agree?" — if not, you are not done

Mandatory Verification Checklist

Before marking ANY task, gate, or checkpoint complete:

• Did I READ the actual evidence file — not just the report about it?
• Did I VIEW the actual screenshot and describe its content?
• Did I EXAMINE the actual command output — not just the exit code?
• Did I QUOTE the actual response body — not just the status code?
• Can I CITE specific evidence for each validation criterion?
• Did I CHECK for partial success masking failure (200 OK with error body)?
• Would a SKEPTICAL REVIEWER agree this is complete?

The Verification Loop

1. IDENTIFY — List every PASS criterion for this gate/task
2. LOCATE — Find the specific files, outputs, or artifacts
3. EXAMINE — Open files, read content, view screenshots
   (Do NOT delegate examination without reviewing findings)
4. MATCH — For each criterion, cite the specific evidence that proves it
   (Any criterion without matching evidence = gate NOT passed)
5. WRITE the verdict:
   - PASS: All criteria have matching evidence, cited specifically
   - FAIL: List which criteria lack evidence and what is needed
   - PARTIAL: Some pass, others need work (never mark complete)

Evidence Format Standards

Good

Criterion: User signup produces a row in `users` with the submitted email
PASS — `e2e-evidence/vg3-db.txt` line 2: `email | [email protected] | 2026-04-20 12:00`

Bad (reject these)

Criterion: Signup works
PASS — "tested manually, looked fine"

Criterion: API returns 200
PASS — "curl returned 200" (but did not quote response body — what if body was an error?)

Rules

1. NEVER mark complete without personally examining every piece of evidence
2. NEVER trust reports — read the actual evidence a sub-agent produced
3. NEVER accept exit codes alone — read the full stdout/stderr
4. NEVER skip criteria — PARTIAL is not COMPLETE
5. ALWAYS cite specifically — file paths, line numbers, quoted output
6. ALWAYS check for partial success masking failure (200 OK with error body, empty array counted as "works")
7. ALWAYS re-verify after any fix — run the full loop again

Integration with Phase Gates

In deepest-plan output, every <phase_gate> declares verdict_by="gate-validation-discipline". The gate's <regression_check> MUST re-execute prior-phase gates and cite their evidence files — not just assert "VG-1 PASS" without reading e2e-evidence/vg1-*.

Security Policy

Read-only — verifies evidence, never modifies code or system state. Refuses requests to skip verification steps or mark tasks complete without evidence.