Judge Verification

Judge Verification

Overview

An independent LLM evaluation layer that verifies whether a task was genuinely completed. The judge reviews the original task goal, the sequence of actions taken, the final state of relevant artifacts, and the claimed completion evidence — then produces a PASS/FAIL verdict with a confidence score and actionable reasoning.

This skill is distinct from verification-before-completion: that skill runs checklist gates within the same agent context. Judge-verification uses a fresh, independent perspective with no access to the executing agent's prior reasoning, catching hallucinated success claims.

When to Use

Skill({ skill: 'judge-verification' });

Invoke when:

• Task claims completion but no test output, diff, or artifact is visible
• High-stakes tasks (security fixes, data migrations, API changes) require independent sign-off
• An agent's completion metadata seems inconsistent with the work described
• The verification-before-completion skill passes but human review is not available
• Loop detection force-done fired (partial completion needs judging)

The Iron Law

JUDGE IS INDEPENDENT — NO SHARED CONTEXT WITH EXECUTING AGENT

The judge must receive only: (1) the original task goal, (2) the list of actions, (3) the final file states. Never pass the executing agent's reasoning or internal notes to the judge.

Judge Evaluation Framework

The judge reviews four dimensions and produces a combined verdict:

Dimension 1: Task Goal Alignment

Question: Does the final state match what the task requested?

Evidence to check:

• Original task subject/description from TaskGet({ taskId })
• Files or outputs explicitly mentioned in the task
• Expected behavior described in the task goal

Score: 0-25 points

Commands:

# Get the original task goal
# (replace task-ID with actual task ID)
node -e "const fs=require('fs');const tasks=JSON.parse(fs.readFileSync('.claude/context/runtime/tasks.json','utf8')||'[]');const t=tasks.find(x=>x.id==='{{TASK_ID}}');console.log(JSON.stringify(t?.subject||'not found'));"

Expected output: Task subject string showing the original goal. Verify: Subject matches what the agent claimed to accomplish.

Dimension 2: Action Completeness

Question: Were the claimed actions sufficient to accomplish the goal?

Evidence to check:

• List of tool calls from task metadata (filesModified, outputArtifacts)
• Were necessary tools invoked? (e.g., Write for file creation, Bash for command execution)
• Are there obvious gaps? (e.g., claimed to run tests but no Bash call with pnpm test)

Score: 0-25 points

Commands:

# Check files modified were actually touched
git diff --name-only HEAD~1 HEAD 2>/dev/null || git status --short

Expected output: List of changed files that should match the agent's filesModified metadata. Verify: At least one file changed; file list is plausible given the task.

Dimension 3: Evidence of Completion

Question: Is there concrete, verifiable evidence the task succeeded?

Evidence to check:

• Test results (exit code 0, pass counts)
• File contents reflect the requested change
• No error output in last bash call
• Build/lint passes

Score: 0-25 points

Commands:

# Check if tests pass (if task involved code changes)
cd /c/dev/projects/agent-studio && pnpm test 2>&1 | tail -5

Expected output: Test summary showing pass/fail counts. Verify: Zero failures for tasks that touched tested code.

# Verify target file content matches task intent
# (judge reads the file and checks against task description)
head -50 {{TARGET_FILE_PATH}}

Expected output: File content consistent with the claimed change. Verify: Content is not placeholder/stub; change is real.

Dimension 4: Final State Coherence

Question: Is the system in a coherent state — no regressions, no broken references?

Evidence to check:

• Lint and format pass
• No new TODO/FIXME introduced without justification
• Referenced files exist
• No circular imports or broken requires

Score: 0-25 points

Commands:

# Quick coherence check
cd /c/dev/projects/agent-studio && pnpm lint:fix 2>&1 | tail -10

Expected output: Zero errors, possibly auto-fix count. Verify: Exit code 0 or only style fixes (no logic errors).

Verdict Calculation

totalScore = dim1 + dim2 + dim3 + dim4  (max 100)

PASS:        totalScore >= 70 AND dim3 >= 15  (evidence gate — cannot pass with no evidence)
FAIL:        totalScore < 70  OR  dim3 < 15
CONDITIONAL: totalScore 60-69 with dim3 >= 15 — requires human review

Output Format

The judge produces a structured verdict:

{
  "verdict": "PASS | FAIL | CONDITIONAL",
  "confidence": 0.87,
  "totalScore": 82,
  "dimensions": {
    "goalAlignment": 20,
    "actionCompleteness": 22,
    "evidenceOfCompletion": 20,
    "finalStateCoherence": 20
  },
  "reasoning": "Task goal was to add input validation. Files modified include auth.ts and auth.test.ts. Tests pass. Validation logic present in auth.ts lines 45-67. No regressions detected.",
  "failureReasons": [],
  "recommendations": ["Consider adding edge case tests for empty string input"]
}

Workflow

Step 1: Collect Task Context

Command:

node -e "
const fs = require('fs');
const taskId = '{{TASK_ID}}';
// Read task context from metadata
const logPath = '.claude/context/runtime/session-gap-log.jsonl';
const lines = fs.existsSync(logPath) ? fs.readFileSync(logPath, 'utf8').split('\n').filter(Boolean) : [];
const relevant = lines.filter(l => l.includes(taskId)).map(l => JSON.parse(l));
console.log(JSON.stringify(relevant.slice(-5), null, 2));
"

Expected output: Recent task log entries with metadata (filesModified, summary). Verify: At least one entry for the task ID.

Step 2: Verify Artifacts Exist

For each file in filesModified, verify it exists and has non-zero size:

Command:

# For each file claimed as modified:
stat "{{FILE_PATH}}" 2>/dev/null && echo "EXISTS" || echo "MISSING: {{FILE_PATH}}"

Expected output: "EXISTS" for each file. Verify: No MISSING entries — missing files = automatic FAIL for dim2.

Step 3: Score Each Dimension

Score dimensions 1-4 using the criteria above. Record each score with one-sentence justification.

Expected output: Four scores totaling 0-100. Verify: Total is consistent with the evidence collected.

Step 4: Calculate Verdict

Apply verdict formula. Check evidence gate (dim3 >= 15 required for PASS).

Command:

const total = dim1 + dim2 + dim3 + dim4;
const verdict =
  total >= 70 && dim3 >= 15 ? 'PASS' : total >= 60 && dim3 >= 15 ? 'CONDITIONAL' : 'FAIL';
const confidence = Math.min(1.0, total / 100 + (dim3 >= 20 ? 0.1 : 0));

Expected output: { verdict, confidence, totalScore }. Verify: Verdict is consistent with the evidence — do not rationalize a PASS without evidence.

Step 5: Write Verdict to Task Metadata

Command:

TaskUpdate({
  taskId: '{{TASK_ID}}',
  status: 'completed', // or keep as-is if just judging
  metadata: {
    judgeVerdict: {
      verdict: '{{VERDICT}}',
      confidence: {{CONFIDENCE}},
      totalScore: {{SCORE}},
      dimensions: { goalAlignment: {{D1}}, actionCompleteness: {{D2}}, evidenceOfCompletion: {{D3}}, finalStateCoherence: {{D4}} },
      reasoning: '{{REASONING}}',
      failureReasons: [{{FAILURES}}],
      recommendations: [{{RECS}}],
      judgedAt: new Date().toISOString(),
    },
  },
});

Expected output: TaskUpdate succeeds with judge verdict in metadata. Verify: TaskGet({ taskId }) returns metadata.judgeVerdict.verdict.

FAIL Handling

When verdict is FAIL:

1. Do NOT mark the task complete
2. Write failure reasons to issues.md
3. Notify the originating agent or Router with the specific failure reasons
4. Suggest concrete remediation steps based on which dimension failed

CONDITIONAL Handling

When verdict is CONDITIONAL:

1. Mark task as blocked with blockerType: 'review'
2. Output: "Judge verdict CONDITIONAL (score {{N}}/100). Human review required before marking complete."
3. List specific gaps that prevented PASS

Integration with verification-before-completion

This skill is complementary to verification-before-completion, not a replacement:

Use both: verification-before-completion first, then judge-verification for sign-off.

Enforcement Hooks

Input validated against schemas/input.schema.json before execution. Output contract defined in schemas/output.schema.json.

Pre-execute hook at hooks/pre-execute.cjs validates that taskId and taskGoal are provided. Post-execute hook at hooks/post-execute.cjs emits a judge-verification event to tool-events.jsonl.

Anti-Patterns

• Never share the executing agent's internal reasoning with the judge
• Never accept a PASS verdict without a non-zero dim3 score (evidence gate is mandatory)
• Never judge a task that hasn't called TaskUpdate(completed) yet — wait for the completion claim
• Never skip this skill for security-sensitive tasks, even when other verification passes

Memory Protocol (MANDATORY)

Before starting: Read .claude/context/memory/learnings.md for past judge verdicts and common failure patterns.

After completing: If verdict is FAIL or CONDITIONAL, append to .claude/context/memory/issues.md:

## Judge Verification FAIL — Task {{TASK_ID}} — [date]
- Verdict: FAIL (score {{N}}/100)
- Failed dimensions: {{DIMS}}
- Root cause: {{REASONING}}
- Recommendation: {{RECS}}

Related Skills

• verification-before-completion — Pre-completion checklist (same-agent perspective)
• behavioral-loop-detection — Detect loops before completion claim
• error-recovery-escalation — Handle errors before reaching judge
• agent-evaluation — Full LLM-as-judge 5-dimension rubric (broader scope)