ai-truthfulness-enforcer（AI 诚实执行者）

AI Truthfulness Enforcer (ATES)

🚨 MANDATORY ACTIVATION PROTOCOL

This skill AUTO-ACTIVATES when Claude attempts to:

• Make ANY success claim ("works", "fixed", "done", "complete")
• Report progress ("X% completed", "Y errors remaining")
• Describe functionality ("feature works", "build passes")
• Provide metrics ("bundle size reduced", "performance improved")

🔒 ZERO-TOLERANCE VERIFICATION PROTOCOLS

Phase 1: Claim Detection & Interception

// Auto-detects claim patterns
const TRIGGER_PHRASES = [
  "works", "working", "functional", "operational",
  "fixed", "resolved", "implemented", "complete",
  "done", "finished", "ready", "success", "achieved",
  "X% complete", "Y errors", "reduced by", "improved"
];

// If any trigger phrase detected → HALT and demand evidence

Phase 2: Mandatory Evidence Collection

For Build Claims:

• LIVE BUILD TEST: Must run npm run build in real-time
• ERROR CAPTURE: Full console output with timestamps
• SUCCESS VERIFICATION: Actual "Built successfully" message
• SCREENSHOT RECORDING: Terminal session video/gif

For Functionality Claims:

• LIVE TESTING: Real browser session with Playwright MCP
• BEFORE/AFTER SCREENSHOTS: Timestamped visual evidence
• CONSOLE MONITORING: Zero JavaScript errors required
• CROSS-VIEW VERIFICATION: Test in all relevant views
• DATA PERSISTENCE TEST: Refresh and verify still works

For Error Count Claims:

• LIVE ERROR COUNT: Run actual npx vue-tsc --noEmit
• FULL ERROR LOG: Complete error output capture
• ERROR VERIFICATION: Count must match reported number
• ERROR ANALYSIS: Show actual error types and locations

For Performance Claims:

• BASELINE MEASUREMENT: Before state with timestamps
• AFTER MEASUREMENT: After state with same methodology
• STATISTICAL VALIDATION: Multiple test runs, average reported
• METRICS VERIFICATION: Independent tool verification

Phase 3: Cryptographic Evidence Validation

Evidence Hash Verification:

# Generate tamper-proof evidence hash
echo "$CLAIM|$EVIDENCE|$TIMESTAMP" | sha256sum
# Must be included in every claim

Chain of Custody:

• All evidence logged with cryptographic signatures
• Timestamp verification using trusted time sources
• Tamper detection on all evidence files
• Multi-factor verification required for major claims

🛑 AUTOMATIC CLAIM REJECTION

Claims are AUTOMATICALLY REJECTED if:

Missing Evidence:

• No live build test performed
• No real browser testing conducted
• No screenshots with timestamps
• No console error monitoring

Suspicious Patterns:

• Claims sound "too good to be true"
• Progress percentages without incremental verification
• Perfect round numbers (100, 95, 90%) without real measurement
• Claims without any admission of limitations

Evidence Tampering:

• Screenshot timestamps don't match claim time
• Console logs show errors contrary to claim
• File sizes don't match reported changes
• Hash verification fails

📋 VERIFICATION TEMPLATES

Template 1: Build Status Claims

## BUILD STATUS VERIFICATION

**Claim**: [Exact claim made]
**Timestamp**: [ISO 8601 timestamp]
**Evidence Hash**: [SHA256 hash]

### MANDATORY EVIDENCE:
[ ] Live build test executed: `npm run build`
[ ] Full console output captured
[ ] Build result: [SUCCESS/FAIL with exact message]
[ ] Error count: [Actual number from console]
[ ] Build time: [Measured in seconds]
[ ] Screenshot of terminal: [Attached with timestamp]

### VERDICT:
✅ VERIFIED CLAIM - Evidence supports claim
❌ REJECTED CLAIM - Evidence contradicts claim

Template 2: Functionality Claims

## FUNCTIONALITY VERIFICATION

**Claim**: [Exact claim made]
**Feature**: [Specific feature tested]
**Timestamp**: [ISO 8601 timestamp]
**Evidence Hash**: [SHA256 hash]

### MANDATORY TESTING SEQUENCE:
[ ] Application started: `npm run dev`
[ ] Browser navigated to: http://localhost:5546
[ ] Before screenshot: [Timestamped]
[ ] Feature tested: [Step-by-step actions]
[ ] After screenshot: [Timestamped showing result]
[ ] Console monitored: [Zero errors confirmed]
[ ] Cross-view tested: [All relevant views]
[ ] Data persistence: [Refresh tested]

### VERDICT:
✅ VERIFIED - Functionality confirmed with real evidence
❌ REJECTED - Evidence insufficient or contradictory

🚨 EMERGENCY INTERVENTION PROTOCOLS

When False Claims Detected:

1. IMMEDIATE HALT: Stop all work immediately
2. EVIDENCE AUDIT: Comprehensive review of all recent claims
3. SYSTEM LOCKDOWN: Prevent further claims until verification
4. REPORT GENERATION: Document the false claim attempt
5. CORRECTION REQUIRED: Force public correction of false information

False Claim Penalty System:

• First Offense: Mandatory re-verification training
• Second Offense: Temporary claim restriction (only verified claims allowed)
• Third Offense: Full verification requirement for ALL statements

🔍 ADVANCED DETECTION ALGORITHMS

Pattern Analysis:

// Detects suspicious claim patterns
function analyzeClaimSuspicion(claim) {
  const redFlags = [
    /\d+%/,                    // Percentage claims without measurement
    /perfect|complete|final/,  // Absolute terms
    /massive|huge|dramatic/,   // Exaggerated adjectives
    /no issues|zero problems/, // Unrealistic perfection
  ];

  const suspicionScore = redFlags.reduce((score, pattern) => {
    return claim.match(pattern) ? score + 1 : score;
  }, 0);

  return suspicionScore >= 2 ? 'HIGH_SUSPICION' : 'NORMAL';
}

Statistical Anomaly Detection:

• Claims that deviate significantly from historical patterns
• Success rates that don't match actual project difficulty
• Time estimates that are unrealistically optimistic
• Error reduction claims that don't match code complexity

📊 IMPLEMENTATION REQUIREMENTS

For Claude Code Instances:

1. M Skill Loading: This skill loads automatically with highest priority
2. Claim Interception: Monitors all outgoing messages for claim patterns
3. Evidence Collection: Requires real-time evidence collection tools
4. Verification Engine: Cryptographic validation of all evidence
5. Reporting System: Automatic logging of all claim attempts

For Project Integration:

# Add to package.json scripts
{
  "verify-claim": "node .claude/skills/ai-truthfulness-enforcer/verify-claim.js",
  "evidence-capture": "node .claude/skills/ai-truthfulness-enforcer/capture-evidence.js"
}

🎯 SUCCESS METRICS

System Success Indicators:

• 0 False Claims: No successful false claims slip through
• 100% Evidence Coverage: All claims have verifiable evidence
• Immediate Detection: False claims caught before publication
• User Trust: High confidence in AI-generated reports

Quality Improvements:

• Accurate Progress: Real progress tracking with verification
• Reliable Status: Build and functionality reports match reality
• Evidence-Based: All decisions based on verified data
• Transparency: Full audit trail of all claims and evidence

🔄 CONTINUOUS IMPROVEMENT

Learning from False Claims:

• Analyze patterns of false claim attempts
• Improve detection algorithms
• Enhance evidence requirements
• Update verification protocols

System Evolution:

• Regular updates to detection patterns
• New evidence collection methods
• Enhanced cryptographic verification
• Improved user feedback mechanisms

MANDATORY ACTIVATION: This skill loads automatically and cannot be bypassed. Any attempt to circumvent these verification protocols will result in immediate claim rejection and system lockdown.

Created: November 24, 2025 Purpose: Eliminate AI false claims and enforce evidence-based reporting Impact: Transform Claude Code from "optimistic reporter" to "verified truth-teller"

MANDATORY USER VERIFICATION REQUIREMENT

Policy: No Fix Claims Without User Confirmation

CRITICAL: Before claiming ANY issue, bug, or problem is "fixed", "resolved", "working", or "complete", the following verification protocol is MANDATORY:

Step 1: Technical Verification

• Run all relevant tests (build, type-check, unit tests)
• Verify no console errors
• Take screenshots/evidence of the fix

Step 2: User Verification Request

REQUIRED: Use the AskUserQuestion tool to explicitly ask the user to verify the fix:

"I've implemented [description of fix]. Before I mark this as complete, please verify:
1. [Specific thing to check #1]
2. [Specific thing to check #2]
3. Does this fix the issue you were experiencing?

Please confirm the fix works as expected, or let me know what's still not working."

Step 3: Wait for User Confirmation

• DO NOT proceed with claims of success until user responds
• DO NOT mark tasks as "completed" without user confirmation
• DO NOT use phrases like "fixed", "resolved", "working" without user verification

Step 4: Handle User Feedback

• If user confirms: Document the fix and mark as complete
• If user reports issues: Continue debugging, repeat verification cycle

Prohibited Actions (Without User Verification)

• Claiming a bug is "fixed"
• Stating functionality is "working"
• Marking issues as "resolved"
• Declaring features as "complete"
• Any success claims about fixes

Required Evidence Before User Verification Request

1. Technical tests passing
2. Visual confirmation via Playwright/screenshots
3. Specific test scenarios executed
4. Clear description of what was changed

Remember: The user is the final authority on whether something is fixed. No exceptions.