scrutiny-validator

name	scrutiny-validator
description	Adversarial scrutiny procedure for engineering validation assignments. Runs hard-gate commands, reviews the current implementation and evidence integrity against assigned contracts, can use feature-reviewer lanes, and returns per-target verdicts.

Scrutiny Validator

Use this skill when the validation assignment asks for implementation scrutiny, hard-gate command review, evidence-integrity audit, or source review for engineering targets.

Scrutiny is not a substitute for real user/caller/operator surface validation. Use user-testing-validator or another surface-specific validator when the contract requires behavior through a real interface.

Inputs

Read:

• Validation assignment, including exact target ids, assigned validation method, and any assignment-level setup or dependency notes.
• Assigned assertions. Prefer compact fields/labels such as Surface, Needs, Behavior, Evidence, and optional Fail, Oracle, or Scope, but accept equivalent headings such as Statement for behavior, Evidence Floor for required evidence, and Non-Goals or Notes for boundaries.
• AGENTS.md.
• Latest worker report for each assigned target when present, plus relevant prior validator reports. Treat reports as claims, not proof.
• Current product checkout files related to the targets. Treat diffs, claimed changes, and worker reports from prior attempts as useful leads when present, not as required inputs or proof.
• Evidence artifacts, setup paths, oracle paths, source baselines, or regression ledgers cited by assignment, contract, prior attempts, or skill.

Procedure

1. Establish scope

   • Identify exact target ids and the current product behavior, files, and evidence surfaces being scrutinized.
   • Parse every assigned assertion into behavior, surface, required evidence, prerequisites, oracle, failure conditions, and scope boundaries. Accept equivalent headings such as Statement, Evidence Floor, Non-Goals, and Notes when they provide the same meaning.
   • If behavior, surface, or required evidence cannot be determined from the assertion and validation assignment, mark the target unverifiable.
   • Map each target's Needs and Evidence requirements before reviewing implementation or evidence surfaces. Do not assume prerequisites, setup, fixtures, services, source baselines, or accepted decisions that are not present.
   • Map each target to implementation files and evidence-producing files.

2. Run hard-gate commands

   • Run tests, lint, typecheck, build, generated-output checks, migration checks, or contract-required commands relevant to the target.
   • Capture command, exit code, focused output, and saved artifact path when the output is too large or important for the report alone.
   • Save durable scrutiny artifacts under <evidence_dir> when useful: command logs, focused outputs, generated-output diffs, fixture/golden review notes, or source-baseline comparison notes.
   • Do not hide exit codes with output truncation pipelines.

3. Review implementation

   • Confirm the current implementation satisfies the contract behavior, surface, required evidence, and any fail, oracle, scope, non-goal, or notes constraints, and accounts for every relevant prerequisite.
   • Check edge cases, error paths, data integrity, auth/authz, compatibility, migration safety, idempotency, and public API behavior when relevant.
   • Flag responsibility drift outside assignment or contract scope.

4. Review evidence integrity

   • Inspect changed or evidence-sensitive tests, fixtures, benchmark scripts, verifier code, golden files, generated expected outputs, source baselines, data files, and cited evidence artifacts when they are allowed and relevant.
   • Restrict that inspection to allowed paths. Do not read hidden verifier internals, hidden tests, holdout labels, forbidden baseline paths, or other off-limits surfaces; use allowed public artifacts and report unverifiable proof gaps instead.
   • Compare the available proof against each target's Evidence field. Missing required evidence owned by this validation task means passed=false.
   • If the contract requires real user/caller/operator surface evidence, do not pass it through scrutiny alone unless the validation assignment explicitly scopes this run as the scrutiny lane and names a sibling validator that owns the real-surface evidence.
   • Fail targets when proof was weakened, mocked away, or made benchmark/test-only.

5. Use feature-reviewer lanes when useful

   • For broad implementation areas or multiple feature areas, spawn feature-reviewer with one bounded review question, assigned target ids, contract paths or bodies, relevant current-checkout files or surfaces, relevant evidence artifacts or command output, claimed changes or prior reports as leads when useful, allowed commands/probes/evidence-write locations, off-limits surfaces, and expected output shape.
   • The subagent report is advisory. You own the verdict.

6. Write regression ledger entries

   • For each failed or unverifiable item, write <regressions_dir>/<item_id>.md with the failing command, review finding, observation, unmet Needs, missing Evidence, artifact paths, and reproduction details.

7. Return verdicts

   • One items[] entry per assigned target.
   • passed=true only when every check owned by this scrutiny assignment passes, relevant contract prerequisites are accounted for, scrutiny evidence artifacts are collected, and no integrity or responsibility-drift issue blocks the target.
   • If required real-surface evidence is outside this assignment's method, name the sibling validator that owns it. Do not claim scrutiny alone proves the full target.

Failure Severity

Blocking findings include:

• Missing implementation for a target.
• Required command failure tied to the target.
• Unmet prerequisite from Needs.
• Missing required evidence owned by this validation task.
• Evidence-sensitive changes that invalidate proof.
• Security, data-integrity, migration, compatibility, or public API break.
• Required oracle/source baseline missing.
• Unverifiable assertion.

Prefer passed=false with concrete limitations over speculative passes.

Report

## Hard-Gate Commands
- `<command>` -> exit <code>. <focused output>

## Per-Target Verdicts
### <target-id>: passed | failed
- Needs: <checked prerequisites, unmet dependencies, or None>
- Evidence: <commands/source review/artifacts collected and any named sibling evidence owner>
- Review: <why contract is or is not satisfied>

## Evidence Integrity
- <changed proof surfaces, missing evidence, and credibility impact>

## Responsibility Drift
- <none or concrete drift>

## Review Limitations
- <unreviewed areas and why>

## Guidance Suggestions
- <optional skill/AGENTS.md/setup guidance for orchestrator>

Call end_node with one item per assigned target, then exit immediately.