A three-layer neuro-symbolic system that adjudicates a denied claim against the exact policy that governs it. Deterministic policy logic, patient-record-grounded clinical reasoning, and a compounding denial-pattern library — engineered so that every output is traceable and citation-checked to its source.
Neuro-symbolic: a language model reasons, a symbolic policy layer constrains it to real rules, and a guardrail layer must pass before any human sees the output.
Layer 1’s grounding corpus builds an EDI 835/837 denial-pattern library — after 60–90 days it enables pre-submission screening, flagging claims that match learned denial patterns before they ever go out. The data moat that compounds with every claim.
Every Ardia model runs on TARA at one of three reasoning tiers. The tier decides how much deliberation a task earns — a clean claim resolves in one pass; a complex molecular denial gets a full, tool-using investigation. Same governed core, same guardrails, different depth. You can watch all three run live on Watch TARA Reason →
Instant, low-deliberation answers with no tool calls. For clean claims, quick eligibility checks, and code lookups where the symbolic policy layer already holds the answer. Swift returns the cited rule and the routing decision — or escalates to Cortex when the case is ambiguous.
Shows its step-by-step reasoning before the answer. Cortex works the problem through — reconciling the denial code, the clinical record, and the governing policy — and exposes each step so a reviewer can follow the logic. For medical-necessity judgments and multi-factor denials.
The full investigation. Scholar plans, calls tools, searches policy & evidence, retrieves, reasons, and cites. It pulls the governing LCD/NCD/MolDX rule, checks NCCN / CPIC / PubMed where relevant, and assembles a fully-referenced appeal — every assertion traced to a named source, or it abstains.
TARA is governed by construction, not by policy memo. Before any model output reaches a human, it runs through the Crucible — a behavioral gate harness that blocks anything unsafe, ungrounded, or out of scope. When a gate fails the answer is withheld entirely, not returned with a warning. These gates are the safeguards that let a regulated buyer and a diligence-minded investor trust the system. They are enforced in code, tested, and non-optional.
Status: the Crucible guardrail harness (all six gates above) and Sentinel de-identification are now implemented and unit-tested — 34 tests green in-repo (Meridian, Cadence, Sentinel, Crucible), run on every push via CI. Integration into the live reasoning pipeline is in progress; full third-party audit (SOC 2, penetration testing) is planned pre-pilot.
Ardia is being built to handle Protected Health Information under HIPAA from day one. Sentinel is the compliance kernel every model inherits: it de-identifies before reasoning, encrypts at rest and in transit, logs every access, and enforces least-privilege. We are pre-revenue and process no production PHI today — the controls below are labelled by what is built now versus designed / planned before the first pilot under a signed BAA.
All 18 HIPAA Safe-Harbor identifiers (§164.514(b)(2)) — names, geography, dates, MRN, SSN, contact, device & biometric IDs — are stripped or generalized before any text reaches a model. Raw PHI never enters an LLM prompt.
PHI encrypted at rest with AES-256 and in transit with TLS 1.3, on HIPAA-eligible Google Cloud (Cloud Healthcare API, BigQuery). Keys managed in Cloud KMS with rotation.
Every PHI access and every reasoning step is written to an append-only, timestamped audit trail — reviewable end-to-end, built to a CLIA/CAP-grade standard for defensibility.
Minimum-necessary access by role — a reviewer sees only what their task requires. No standing broad access to PHI; every grant is scoped and logged.
A hard rule in the codebase: patient identifiers are never sent to a third-party model API and never written to console or error logs. De-identified, minimum-necessary data only.
No vendor touches PHI without a Business Associate Agreement. BAAs with the cloud and model providers are to be executed before any pilot processes real patient data.
TARA is a framework in build, but it already produces real, reproducible results — because parts of it are deterministic or trained from scratch on public data. We separate what is measured from what is modelled, on every page, on purpose.
TARA isn’t one product — it’s the governed reasoning core the whole constellation inherits. Each model does one job well, at the tier its job demands, behind the same guardrails and the same audit trail.
| Model | Job | Primary tier | Status |
|---|---|---|---|
| MolecuIQ | Molecular & genomic denial recovery (MolDX, NGS, NCCN) | Scholar | Modelled target |
| ToxIQ | Toxicology denial appeals (G0480–G0483, NCCI, PGx overlay) | Cortex | Modelled target |
| PulmoIQ | Respiratory intelligence — asthma / COPD (GOLD/GINA) | Cortex | Modelled target |
| Meridian | PAMA rate-cliff engine — deterministic, no PHI | Symbolic | Deterministic · unit-tested |
| Cadence | Movement / deterioration model (phone sensors) | Swift | Measured · 95.45% UCI HAR |
| Aria | Voice-first elder companion & safety spine (FAST/C-SSRS/911) | Cortex | Modelled target |
| Lumen | Plain-language report & scan explainer | Cortex | Modelled target |
| Sentinel | De-identification & audit compliance kernel | Symbolic | In build |
It executes rules against encoded policy, so it cannot invent a policy section. If a section governs the CPT under that MAC, the engine cites it; if none does, it returns nothing to fabricate.
Every clinical assertion cites a specific line in the patient record — not a plausible invention. If the record doesn't support an appeal, the engine says so rather than manufacturing support.
Every AI output requires human-in-the-loop practitioner attestation before it is submitted, satisfying Texas SB 1188.
Every reasoning step is designed to be captured in a full, timestamped, CLIA/CAP-grade audit trail — reviewable end to end.
TARA is engineered against the rules a healthcare AI actually has to satisfy — not as a compliance afterthought, but as design constraints baked into the guardrails and the audit trail. We claim alignment and readiness, not certifications we don’t yet hold.
Safe-Harbor de-identification, encryption, audit logging, minimum-necessary access. HIPAA-ready by design; no production PHI processed pre-BAA.
AI clinical outputs require review by a licensed professional before use. Enforced by the human-in-the-loop guardrail — nothing auto-files.
The Responsible AI Governance Act — transparency, documented intended use, and human oversight. TARA’s audit trail and non-diagnostic scope are built to it.
Appeals are grounded in the record and cited to policy — designed so nothing overstated is ever submitted. Human attestation is the backstop.
The reasoning trail is designed to the documentation standard lab accreditation expects — timestamped, attributable, reviewable end-to-end.
Not yet certified. SOC 2 readiness and audit are planned before enterprise deployment; we say so plainly rather than imply a badge we haven’t earned.
Ardia is pre-revenue and seed-stage. The framework is real and partly built; most models are designed and in build. Here is exactly where each piece stands — no rounding up.
| Property | Generic AI | TARA |
|---|---|---|
| Policy lookup | A single model may invent a plausible-sounding policy section. | Layer 2 executes deterministic rules over encoded LCD/NCD/MolDX policy — it cannot invent a section. |
| Clinical claims | May generate assertions not present in the record. | Layer 3 reasons only over the grounded record — every inference traces to a specific source. |
| Unsupported appeal | Tends to produce an answer regardless. | If documentation doesn't support an appeal, TARA says so. |
| Human review | Often none required by design. | Practitioner attestation required before submission (TX SB 1188). |
| Auditability | Reasoning is typically opaque. | Designed for a full timestamped, CLIA/CAP-grade audit trail. |
| Learning over time | No structured denial memory across claims. | Layer 1’s grounding corpus compounds an EDI 835/837 denial-pattern library, enabling pre-submission screening after 60–90 days. |
Enter a denial and watch TARA plan, pull the governing policy, reason against the record, and assemble a cited, review-ready appeal — at the Swift, Cortex, or Scholar tier. The trace is real, not a scripted animation.
