AKAMAI INTELLIGENCE GROUP CAUSAL REASONING AUDITOR Technical Brief

Document ID: AIG-CAUSAL-005 Classification: UNCLASSIFIED // PUBLIC RELEASE Version: Public release | June 8, 2026 Author: Christopher Ramos Research and drafting support: Deus ex Machina Status: PUBLIC TECHNICAL BRIEF Evidence posture: Local-harness evaluation evidence; no live-system safety guarantee implied.

Summary

AIG's Causal Reasoning Auditor is a four-phase pipeline that takes real governance evaluation data (GAL-040B, 1,800 trials across 6 adversarial scenario families) and produces directed causal graphs showing why governance decisions are made - not just what decisions were made.

It does not use an LLM for causal inference. It uses classical causal discovery algorithms (HillClimbSearch + BIC scoring, PC constraint-based algorithm) operating on binary concept activation vectors derived from actual trial data. This is structural causal modeling - a peer-reviewed methodology with a clear audit trail.

The result is a machine-readable governance explanation layer: every trial attribution traces to root causes, and every root cause maps to a concrete governance recommendation.

Motivation

Standard evaluation frameworks measure outcomes. They tell you accuracy, recall, false positive rate.
What they do not tell you is why a governance system failed - or more importantly, which upstream conditions causally produced a given failure.

This matters for three reasons:

1. Accountability. WARLORD owns decision authority. STRATEGIST owns consequences. ORACLE owns probabilistic forecasting. Without causal attribution, these roles operate on symptoms, not causes.

2. Anticipation. If adversarial_escalation causally precedes human_gate_bypassed (confirmed by two independent algorithms), then detecting escalation early is worth more than detecting bypass late.

3. Credibility. Causal claims require a methodology. This auditor provides one - one that a reviewer can inspect, reproduce, and challenge.

Architecture

Phase 1 - Concept Extraction

Converts each of the 1,800 GAL-040B trial records into a binary concept activation vector.

35 governance concepts across 7 dimensions:

• Authority (valid / missing / escalated / human gate)
• Provenance (signed / unsigned / forged)
• Topology (sparse / dense / role-routed / cross-domain)
• Load (normal / burst / degraded)
• Adversarial signal (absent / present / mimicry / escalation)
• Fanout (within budget / over budget / attempted excess)
• Outcome / Failure (allow / block / correct / incorrect / false positive / false negative)

Output: 1,800 binary concept state vectors, one per trial.

Phase 2 - Counterfactual Augmentation

Observed governance states cluster tightly - real trials only cover 257 unique concept fingerprints. Causal discovery needs diverse state coverage.

Using MCMC-inspired counterfactual chain generation (based on methodology from HuggingFace paper 2606.05972 - causal reasoning in LLMs via counterfactual augmentation), the pipeline generates 23,964 synthetic states from the 1,800 observed ones.

Output:

• 25,764 total states (1,800 observed + 23,964 synthetic)
• 12,205 unique concept fingerprints
• 47.5x coverage expansion
• 0 invalid states (all constraint-valid)
• 100% chain provenance (every synthetic state traces to its seed trial)

The 9.9% constraint-absorbed states (perturbations that repair back to the same state) are semantically correct - they demonstrate that governance constraints are tight enough to absorb random perturbation. That is a finding, not an artifact.

Phase 3 - Causal Structure Discovery

Two independent algorithms run on the augmented concept matrix:

HillClimbSearch + BIC scoring (score-based, greedy)

• Runs on the full 25,764-state dataset
• Produces 91 directed edges across 35 concept nodes
• Prior alignment: 25% (5 of 20 AIG governance priors confirmed)

PC algorithm - chi-square independence tests (constraint-based)

• Runs on a 2,000-state stratified sample (tractability constraint; max_cond_vars=3)
• Produces 30 directed edges across 34 concept nodes
• Prior alignment: 15%

Consensus DAG: edges confirmed by both algorithms - high-confidence causal structure.

Key finding: adversarial_escalation is the single most causally active node - 3 outgoing consensus edges. It causally precedes outcome blocking, gate bypass, and role contamination. Early detection of escalation is worth more than any downstream remediation.

Phase 4 - Causal Attribution

The consensus DAG is applied to every one of the 1,800 original GAL-040B trials. Each trial receives:

• Active concept vector (from Phase 1)
• Active sink nodes - which outcome/failure concepts fired
• Root causes - active concepts with no active upstream predecessors in the consensus DAG
• Causal chains - shortest active paths from root to outcome (e.g., adversarial_escalation -> outcome_block)
• Governance recommendation keys - which root causes have actionable recommendations

Results across 1,800 trials:

Per-scenario-family correct rates:

false_consensus is the only family with failures. Both failures (F-BENIGN-002) trace to authority_missing + role_mix_contaminated as causal roots.

mixed_noisy_load has the highest causal complexity (6.49 active root causes per trial) - the noisiest governance surface.

Governance Recommendations

Recommendations are keyed to causal root nodes and weighted by trigger frequency across all 1,800 trials.

What This Is Not

• This is not a universal causal model of all AI governance. It is a causal model of the specific concept space defined for GAL-040B.
• This is not external validation. The causal structure reflects the structure of AIG's own evaluation data.
• The PC algorithm runs on a 2,000-state sample, not the full 25,764-state augmented set. Consensus edges are directionally reliable, not statistically exhaustive.
• Causal discovery on binary data has well-known limitations (faithfulness assumption, Markov condition). These are acknowledged.

Artifacts

All artifacts are versioned in the AIG repository under stage2_artifacts/causal_auditor/.

Eval reports for all 4 phases pass at 100%.

Source Code

Document ID: AIG-CAUSAL-005 | June 8, 2026 | Christopher Ramos | AIG Intel