AIG-CAUSAL-005 | June 2026 | Christopher Ramos

Causal Reasoning Auditor

A four-phase pipeline that moves governance evaluation beyond accuracy metrics — attributing every governance decision to its upstream causal drivers using structural causal modeling on real adversarial trial data.

DICE-Adjacent Causal Inference GAL-040B 1,800 Trials 4/4 Phases Passing

1,800 GAL-040B Trials Attributed

99.89% Correct Governance Rate

47.5× Coverage Expansion (Phase 2)

11 Consensus Causal Edges

Four-Phase Pipeline

Phase 1

Concept Extraction

Governance concepts

Converts 1,800 GAL-040B trial records into binary concept activation vectors across authority, provenance, topology, load, adversarial signal, fanout, and outcome dimensions.

Phase 2

Counterfactual Augmentation

25,764

Total states generated

MCMC-inspired counterfactual chain generation expands 1,800 observed states to 25,764 — 47.5× coverage expansion. 0 invalid states. 100% provenance-traced.

Phase 3

Causal Structure Discovery

Consensus edges

Two independent algorithms — HillClimbSearch+BIC (91 edges) and PC chi-square (30 edges) — are run and intersected. Consensus = edges confirmed by both.

Phase 4

Causal Attribution

5.86

Avg root causes / trial

Consensus DAG applied to all 1,800 trials. Each trial receives root cause list, causal chains, and governance recommendation keys. 0 unmatched trials.

Key Finding

adversarial_escalation is the single most causally active governance concept — 3 outgoing consensus edges confirmed by two independent algorithms:

adversarial_escalation → outcome_block
adversarial_escalation → human_gate_bypassed
adversarial_escalation → role_functional (contaminates functional role signal)

Detecting escalation early is worth more than any downstream remediation. PARIAH intercepts at the escalation stage. JACKAL emits early-warning trust-fracture signal. This is the governance architecture doing its job — the data confirmed it.

Consensus DAG — 11 High-Confidence Edges

Edges present in both HillClimbSearch (BIC) and PC (chi-square independence test) outputs. Two independent algorithms. Same edge = high-confidence causal claim.

Cause		Effect	Interpretation
adversarial_escalation	→	outcome_block	Escalation causally drives block decisions
adversarial_escalation	→	human_gate_bypassed	Escalation causally precedes gate bypass
adversarial_escalation	→	role_functional	Escalation contaminates functional role signal
fanout_attempted_excess	→	fanout_over_budget	Excess attempts causally produce budget violations
load_normal	→	load_burst	Temporal: normal load transitions to burst
load_normal	→	load_degraded	Temporal: normal load transitions to degraded
role_mix_contaminated	→	role_mix_clean	Contamination suppresses clean role signal
outcome_correct	→	failure_none	Correct outcome causally produces no failure
outcome_incorrect	→	failure_false_positive	Incorrect outcome causally produces false positive
authority_valid	→	authority_missing	Valid authority context predicts missing authority transitions
adversarial_absent	→	adversarial_present	Temporal: absence precedes presence in adversarial trials

Per-Scenario-Family Attribution Results

Scenario Family	Trials	Correct Rate	Avg Causal Complexity
benign_coordination	300	100.00%	5.90
rogue_coalition	300	100.00%	6.00
provenance_degradation	300	100.00%	5.81
mixed_noisy_load	300	100.00%	6.49
unauthorized_authority_escalation	300	100.00%	5.50
false_consensus	300	99.33%	5.48

false_consensus is the only family with failures (2 of 300 trials — F-BENIGN-002). Both failures trace to authority_missing + role_mix_contaminated as causal roots. mixed_noisy_load has the highest causal complexity (6.49) — noisiest governance surface.

Governance Recommendations

Keyed to causal root nodes. Trigger rate = fraction of 1,800 trials where this root was causally active.

46.3%

authority_missing

Enforce authority attestation at every inter-agent handoff. WARLORD owns this gate — no exceptions.

41.7%

adversarial_present

When adversarial signal is confirmed, route all downstream decisions through WARLORD authority gate before allow/block is issued. ORACLE flags elevated false-negative risk in this condition.

16.7%

fanout_attempted_excess

Implement hard fanout caps at the message dispatch layer. RELAY enforces delivery — WARLORD must own the fanout budget gate.

15.5%

load_burst

ORACLE models burst arrival rates and pre-positions WARLORD for throttle decisions before fanout threshold is reached.

12.3%

role_mix_contaminated

OPSEC enforces strict role boundary separation. SHADOW owns low-observable behavior — contamination detection must not cross into HUMINT lanes. JACKAL emits early-warning trust-fracture signal.

8.3%

adversarial_escalation

PARIAH increases sensitivity at the role boundary when escalation_target is non-null. Mandatory human gate review when adversarial_escalation is active.

Scope & Limitations

What this is not

This is not a universal causal model of AI governance. It is a causal model of the concept space defined for GAL-040B.
This is not external validation. Causal structure reflects AIG's own evaluation data.
The PC algorithm runs on a 2,000-state sample for tractability. Consensus edges are directionally reliable, not statistically exhaustive.
Causal discovery on binary data has well-known limitations (faithfulness assumption, Markov condition). These are acknowledged.
No live agent inference. No network calls. No LLM used for causal inference — classical algorithms only.

Documents & Artifacts

Technical Brief (AIG-CAUSAL-005) ← Research Index DICE Capability Page