When Eligibility Cannot Be Probabilistic

The approval that fails the audit

Utilization management is not a summarization task. A PA request for advanced imaging, a specialty drug, or an elective procedure is a stack of member eligibility, plan coverage, CMS guideline invariants, ICD/CPT/HCPCS relations, site-of-service constraints, and documentation checklists. A UM nurse mid-queue can pend what she cannot prove. A generative agent can invent what it cannot prove — and sound fluent doing it.

That is the compliance failure mode: a hallucinated approval that would not survive CMS review, an appeal, or a medical-necessity challenge. Regulators care whether the gate was invariant. Members care whether the care was correctly authorized. Probabilistic eligibility serves neither.

Medical necessity is not a vibe. It is an invariant. ƐMinds ships a Zero-Hallucination Compliance Gateway for prior authorization: Brain truth bound to CMS and coding ontologies, Reflex judgment on coverage rules (Choice, Score, Noul), Voice only for edge and exception letters — governed in LEAP so clinical and compliance experts verify the rules in prose.

Why a fragile LLM cannot own the PA gate

Prompt chains that read referral notes, retrieve policy PDFs, and draft approve/deny letters look autonomous on a slide. In production they ask one model to be guideline book, judge, and narrator. Token-by-token planning burns seconds and cents; worse, nothing in the loop binds the disposition to CMS criteria or a typed ICD/CPT relation. Latency and unit cost matter at PA volume — blueprint baselines put a typical agent loop near ~8.4 s and ~$0.12 per turn — but the decisive defect is architectural: eligibility remains probabilistic when the statute demands certainty.

Executable invariants: the OBDA path

ƐMinds deploys the Ontologically Bound Decision Architecture — the Tripartite Cognitive Engine — so the same request takes a different path. The HealthCare proof point from the blueprint is exact: automated prior authorization strictly bound to CMS guidelines and clinical coding ontologies. The figure below is that path: Brain hydrates typed state, Reflex decides, Voice stays on the edge, LEAP keeps the rules auditable.

1. Brain — hydrate CMS and coding truth

Request ingress hits the Ontological Core. Parameterized subgraph hydration pulls the typed slice that matters: member eligibility, plan coverage nodes, CMS guideline invariants, ICD/CPT/HCPCS relations, site-of-service constraints, and documentation requirements already encoded as deductive rules. This is domain truth — not a vector neighborhood of similar denial letters.

2. Reflex — Choice, Score, Noul on coverage

The Reflex Layer evaluates a serialized state payload against a structured dictionary of typed questions in a single forward pass — typically on the order of ~80 ms for the decisional block:

• Noul — does this request satisfy the CMS / plan medical-necessity assertion? (calibrated boolean)
• Score — clinical complexity / pend urgency on an ordinal scale
• Choice — disposition from a closed, schema-valid set (e.g. approve · pend · deny · human clinical review)

High confidence (≥ 0.95) mutates the case graph and routes the workflow immediately. Low confidence escalates — the loop does not invent coverage certainty.

3. Voice — LLM only for edge and exception letters

When confidence dips, documentation is incomplete, or the task is generative (a nuanced denial rationale letter, an appeal response draft, a clinician-facing clarification), the LLM Voice synthesizes prose. It stays off the inner eligibility loop: no coverage folklore buried in a prompt, no boolean CMS checks performed by token prediction. System 2 is reserved for exceptions.

4. Canvas — LEAP so clinical and compliance can audit

Coverage criteria, CMS guideline mappings, coding invariants, and escalation thresholds live in the LEAP literate canvas. Medical directors, UM nurses, and compliance officers verify rules in markdown; agents propose ontology expansions and run invariant tests in the same medium. What the client gets is zero IP drift — PA logic is not trapped in tribal prompt chains that walk out with a contractor.

What the client gets — crystal clear

• Strictly bound to CMS guidelines / coding ontologies — deterministic compliance, not free-form model opinion; the Zero-Hallucination Compliance Gateway shape for Healthcare PA.
• Sub-second / sub-300 ms triage — OBDA loop ~180 ms vs ~8.4 s agent loops, so UM portals and synchronous workflows hold.
• Fixed unit economics — ~$0.0002 per turn vs $0.12 LLM loops; judgment priced on typed evaluation, not open-ended burn.
• Auditability for regulators — every disposition reconstructible from ontology state, calibrated Reflex outputs, and LEAP-governed rule versions.
• Zero IP drift — rules live in ontology + LEAP, not prompt folklore.

Phase 1 → Phase 2 path

The commercial on-ramp matches the blueprint — no mystery packaging:

• Phase 1: Sub-Second Latency & Logic Audit — 2 weeks, $25,000–$45,000. OpenTelemetry profiling of existing PA / UM agent loops; autoregressive waste map; complete LEAP specification blueprint mapping CMS coverage and clinical coding logic onto TypeDB and the Reflex Layer.
• Phase 2: OBDA Core Implementation — 6–10 weeks, $120,000–$250,000. Production ontology, fast-hydrate pipeline, Reflex decision paths, LEAP operational manual — live PA triage with compliance audit logging.

Phase 3 retainers and licensing exist for governed autonomy thereafter; the point of essay 03 is what production feels like after Phase 2: CMS-bound dispositions with a reconstructible audit trail instead of probabilistic eligibility.

Sibling proof points

Adjacent industries reuse the same engine shape from the applications map — for example the SaaS renewal-reflex scenario and FinTech in-line fraud — different statutes, same architecture:

• SaaS Operations — account renewal triage, churn defense, and expansion routing with zero human delay.
• FinTech — in-line card fraud and transaction blocking under 100 ms without breaking payment-processor SLAs.

Different statutes. Same architecture: Brain → Reflex → Voice on the edge → LEAP Canvas.