interlock-ops（联锁操作）

Interlock Operations Skill

This skill governs operational behavior for Interlock enforcement.

Core Rules

1. Never Publish Enforcement Heuristics

Enforcement logic is internal. Never commit or document:

• Threshold values for enforcement decisions
• Specific trigger conditions
• Bypass mechanisms or exceptions
• Internal scoring algorithms

Reason: Publishing enforcement heuristics enables gaming.

2. Every Enforcement Action Must Emit a Receipt

No silent enforcement. Every action must produce:

{
  "action": "refusal|degradation|allowance",
  "timestamp": "2026-01-10T12:00:00Z",
  "reason": "quality_floor_breach",
  "incident_id": "int-123456",
  "confidence": 0.45,
  "threshold": 0.80
}

3. Receipt Format Changes Require Validation

Before changing receipt schema:

1. Update the schema definition
2. Create positive test fixture
3. Create negative test fixture
4. Run ./scripts/claude/receipt_audit.sh
5. Verify both positive and negative tests pass

4. Any Failure = Degrade/Refuse

Never allow silent passes:

Receipt Lifecycle

1. EMIT    → Receipt generated at decision point
2. SIGN    → Cryptographic signature applied
3. STORE   → Written to forensic log
4. VERIFY  → Schema validation on read
5. AUDIT   → Periodic integrity check

Validation Commands

# Full receipt audit with negative tests
./scripts/claude/receipt_audit.sh

# Public safety check (no secrets in receipts)
./scripts/claude/public_safety_check.sh

Schema Files

Prohibited Actions

• Publishing enforcement thresholds
• Silent passes on validation failures
• Modifying receipts after signing
• Skipping negative tests
• Committing real receipts to public repo