Cookbook

How to audit every AI decision with intelligence traces

Every response carries a trace ID. Fetch the full trace: what it remembered, what tools it proposed, what governance decided, what gaps it found. Hash-chained for tamper detection.

Every response includes a trace

The intelligence object in every response contains the full decision breakdown.

intelligence object

{
  "intelligence": {
    "capability": {
      "total_available": 42,
      "healthy_count": 38,
      "active_integrations": ["stripe", "github", "slack"]
    },
    "reasoning": {
      "intent": "create_invoice",
      "complexity": "simple",
      "confidence": 0.95,
      "approach": "direct_tool_call"
    },
    "memory": {
      "healthy": true,
      "knowledge_queried": 147
    },
    "governance": {
      "active": true,
      "decision": "auto",
      "confidence": 0.94
    },
    "humility": {
      "gaps_detected": 0,
      "blocking": false
    },
    "process": {
      "started_at": "2026-06-21T10:30:00Z",
      "duration_ms": 840,
      "trace_id": "tr_abc123"
    }
  }
}

Query traces from any surface

SDK — list traces

const traces = await cadreen.traces.list({ limit: 10 });
for (const trace of traces) {
  console.log(trace.id, trace.governance.decision, trace.process.duration_ms);
}

CLI

cadreen traces                    # last 10 traces
cadreen traces --limit 50         # last 50

Dashboard

Visit /infra/dashboard/traces to see all traces with:
- Humanized governance decisions ("Handled on its own", "Asked for permission")
- Replay capability
- Filter by decision type, time range, source

Same traces, accessible from dashboard, CLI, and SDK

Hash-chain integrity (Semantic Action Log)

Every action in the Semantic Action Log is hash-chained. Each entry includes the hash of the previous entry. Any tampering breaks the chain — detectable instantly.

Verify chain integrity

// Verify the entire chain
const verification = await fetch("/api/sal/chain/verify");
const result = await verification.json();
console.log(result.valid);     // true
console.log(result.entries);   // 12,847
console.log(result.broken_at); // null (or index if tampered)

Note

The Semantic Action Log records 24 semantic action types with hash-chain integrity. This is your "black box recorder" for compliance.

Note

Next: How to make AI learn from repeated interactions — reusable procedures, pattern detection.