Deterministic classical-statistics sensors applied to AI-system log data. No AI in the analytical pipeline -- every finding derives from peer-reviewed statistical methods that predate the AI era by 50 to 125 years. Independent-verifier methodology built to survive the higher Rule 702 admissibility bar that courts started enforcing after the 2025 AI-hallucinated-expert exclusions.
External validation of the “no AI in the analytical pipeline” discipline: Angela Shi’s July 2026 empirical catalog of embedding failure modes documents that stronger AI models do not fix the structural failures — they move the failure line. Her line: “the fix is not a stronger embedding model.” The Rule 702-defensible path is deterministic classical statistics applied to the AI’s output, not more AI applied to audit the AI.
Every sensor rests on the same analytical foundation: deterministic classical statistics applied to AI-decision output streams. The pipeline itself uses no AI. Every measurement uses peer-reviewed statistical primitives that predate the AI era by 50 to 125 years -- Kullback-Leibler divergence (1951), Fisher variance-ratio F-test (1920s), chi-square goodness-of-fit (Pearson 1900), Cohen's kappa (1960), signal-detection ROC analysis (1950s). The unifying pattern is predictor-corrector: a predictor produces the expected distribution (from disclosure, spec, baseline, or legally-mandated equal-treatment), KL divergence measures the departure of actual from predicted, and pre-registered thresholds trigger the corrector-need signal. Same shield across every sensor.
| Sensor / methodology | Best-fit case type |
|---|---|
| Distributional-shape sensor | Disparate-impact in AI-driven hiring, healthcare denial, insurance claims, lending, housing, criminal-justice |
| Retrieval-auditor | LLM copyright / training-data reproduction; RAG-grounded response integrity |
| Prompt-drift sensor | Product-liability where AI behavior changed silently; vendor breach-of-contract on performance; silent-divergence and silent-alignment findings |
| Tool-call-grader | Autonomous-agent tort (self-driving, robo-adviser, autonomous medical devices, trading bots) |
| Predictor-corrector | Fiduciary-breach and disclosure-alignment matters; disclosure-vs-actual departure quantification; automated-decision cases where the corrector layer failed |
| Router-drift sensor | Multi-model AI systems (fintech routing, healthcare triage cascades, agent-orchestration) |
| Trade-execution drift sensor | Robo-adviser fiduciary breach; PFOF matters; order-routing litigation; best-execution-obligation enforcement |
| Prior-authorization decision drift sensor | Healthcare AI (Medicare Advantage post-acute denial, nH Predict shape) |
| Underwriting drift sensor | Fair-lending disparate-impact; P&C insurance AI; AI-driven credit-scoring discrimination |
| Ranking / marketplace-suppression drift sensor | Antitrust, self-preferencing, marketplace-suppression securities matters |
| Ad-targeting-eligibility drift sensor | HUD-shape ad-discrimination; algorithmic-amplification MDL matters |
| Content-moderation drift sensor | Section 230 design-not-content challenges; viewpoint-neutrality claims; MDL 3047-adjacent matters |
| Independent-verifier declaration | Rule 702 admissibility protection across all of the above (load-bearing after 2025 Kohls-family exclusions) |
Additional sensor variants (cohort-eligibility drift, consent-drift, SLA-drift, retention-decision drift, access-control drift) are available on matter-specific inquiry.
Every published sample uses the same methodology structure that would apply to your matter's log data. Different verticals, identical detection shape.
Healthcare Snapshot sample SC P&C insurance Snapshot sample All 11 vertical samples
Bring the case name, docket number, defendants, side, and one paragraph on the AI system in question. I'll bring the arsenal-fit read and honest scoping.
Book on Cal.com