Causal Chain Attribution

In industrial AI evaluation, understanding that an agent failed is less important than understanding why it failed. AgentV 1.5.0 introduces the Causal Chain, a chronological ledger that correlates forensic triggers with terminal outcomes.

1. Triggers vs. Symptoms

The most common mistake in AI diagnostics is confusing a symptom for a root cause.

Aspect	Type	Example
Trigger	Root Cause	A logic loop where the agent calls ls repeatedly.
Symptom	Terminal Status	INFRA_TIMEOUT (The run ended because of the loop).

The Causal Chain ensures that the reasoning error (the loop) is the classification, rather than the secondary effect (the timeout).

---

2. Reading the Chain

Every failure report now includes a causal_chain field. This is a timestamped record of forensic alerts generated by Analyzers during the diagnostic process.

{
  "root_cause": "logic_planning_error",
  "terminal_status": "infra_timeout",
  "causal_chain": [
    {
      "timestamp": 1712904000.123,
      "trigger": "logic_planning_error",
      "evidence": "Fuzzy loop detected: agent repeated 'ls -la' 3 times.",
      "turn_index": 5,
      "severity": "medium",
      "rank": 0
    },
    {
      "timestamp": 1712904060.456,
      "trigger": "infra_timeout",
      "evidence": "Global timeout (300s) reached.",
      "turn_index": 12,
      "severity": "high",
      "rank": 1
    }
  ]
}

---

3. Real-World Correlations

Advanced analyzers (available via Enterprise extensions) often reveal causal links that standard loggers miss:

Case Study: The “Try-Hard” Loop

1. Turn 5: The agent receives a Permission Denied error for a file.
2. Turns 6-10: The agent attempts 5 different variations of sudo, chmod, and chown, all failing with parity mismatch.
3. Turn 11: The machine reaches a 100% CPU spike (Resource Gradient).
4. Turn 12: The session crashes.

Traditional Result: INFRA_DOCKER_FAILURE AgentV Causal Chain: Identifies the STRATEGIC_PLANNING_LOOP as the trigger, allowing the developer to fix the agent’s persistence logic rather than debugging the sandbox infrastructure.

---

4. Audit Traceability

For compliance audits, the Causal Chain provides a Deterministic Forensic Trail. Every link in the chain is backed by evidence found in the Forensic Ledger, including state fingerprints and hardware telemetry.