holdout-validation

Cross-reference agent self-review claims and evidence-bundle entries against actual file state using hidden holdout scenarios, producing P1/P2/P3 findings mapped to visible acceptance criteria only. C

SKILL.md

GitHub

Holdout Validation

Contract

Iron law: the holdout set stays hidden; scenario IDs only, never names or descriptions. Invoked by /flow:start Phase 4 VERIFY step 4, /flow:address Phase 4 step 3, and /flow:review Phase 3 fan-out (twice in Path A, once in Path B) with self-review findings, an evidence bundle draft, and a file list. Returns Holdout validation: PASS or FINDINGS with P1/P2/P3 rows citing file:line, mapped to visible acceptance criteria, consumed by verdict-judge. Permitted skips: none; with a missing input, evaluate what is available and note the gap, halting only when no file list is given.

Inputs

  1. Self-review findings: the code-reviewer's P1/P2/P3 findings
  2. Evidence bundle draft: per-criterion evidence, including ### Test inputs and expected values and ### Risk map coverage
  3. File list: every file modified or created on the branch

Schema: schemas/holdout-validation/input-schema.json.

Process

  1. Load scenarios for each criterion type present from templates/holdout-scenarios/{behavioral,api,error,data}.md. Number them by document order across loaded files (scenario-1 to scenario-N); skip and note a missing type.
  2. Parse claims. Extract each finding's and evidence entry's claim and file references. A claim with no file/line citation is a bare assertion: automatic P2.
  3. Cross-reference claims. Read each cited location; record CONFIRMED or CONFLICT. "Test added for X" needs assertions that verify X, not a test that names X.
  4. Cross-reference expected values. For each criterion, read ### Test inputs and expected values and open every cited test. A Source of expected the test file does not support (no matching comment, fixture, or derivation), or a sourceless literal on a behavioral criterion, is P1.
  5. Cross-reference risk coverage. Read ### Risk map coverage. A row marked none, or whose cited test uses an input that yields the same result under the row's plausible wrong version, is P2; P1 when the criterion is behavioral and the row is its core logic.
  6. Evaluate every scenario against file state and claims, even seemingly inapplicable ones. A FAIL maps to a visible criterion, described without reference to the scenario.
  7. Emit findings; run the security check.

No evidence bundle: skip steps 4–5 and note it. No self-review findings: scenarios against files only.

Priority mapping

  • P1: claim contradicted by file state; scenario failure the self-review missed; step 4 source mismatches
  • P2: weakness the self-review understated (test covers only the happy path); bare assertions; step 5 coverage gaps
  • P3: minor gap not affecting correctness (cited line number off)

Output

Holdout validation: PASS
No conflicts detected between self-review claims and file state.
Holdout validation: FINDINGS

P1:
- {file:line}: {conflict}

P2:
- {file:line}: {weakness}

P3:
- {file:line}: {minor gap}

Blocking: {Yes — P1/P2 present | No — P3 only}

Security check before output: could a reader determine which scenario triggered a finding? If yes, generalize. Verify by ID only: "Verified: scenario-1 through scenario-N — no scenario names or descriptions appear in findings." Never list names to prove their absence.

Rules

  • Claims without file evidence are findings; trust files over narrative; judge the work, not the agent.
  • Scenario IDs only, in findings, conversation, and every artifact.
  • Every finding cites file:line and maps to a visible criterion only; "test does not cover timeout" beats "coverage is weak".

Path A and Path B

review.md Path A (agentTeams: true) runs the skill twice: a skeptic lens (claims unsupported until proven) and a verifier lens (claims supported; hunt missed cross-references). Findings raised by both lenses get consensus; by one, unchallenged. Holdout findings never enter the A.3 challenge round. Path B runs once, emitting unchallenged. Details: references/holdout-lens-dispositions.md.