TARA is Ardia's governed reasoning path — a shared safety preamble plus a roughly 90-word persona prompt wrapped around a single Google Gemini call, with regex de-identification in front of it (structured identifiers redact; plain personal names verifiably do not) and six deterministic gates behind it that withhold the answer entirely on any failure — real, running and independently probe-verifiable today, but far thinner than the "neuro-symbolic policy engine" the marketing page describes, and carrying zero measured accuracy of any kind.
Cross-cutting core, not a pillar SKU. TARA is the shared reasoning path underneath all five pillars — precision medicine, molecular and genomic diagnostics, the 2027 PAMA rate cliff, pulmonary and respiratory care, and elder care — plus toxicology on the roster. At runtime this is literal, not metaphorical: PulmoIQ and Meridian execute on the TARA engine path, so two of the ten named models on the site are TARA wearing a different label.
Runs on production and returns six gate verdicts with a reason on every call, and a failed gate now withholds the answer entirely — enforced in code and verified on production by direct probe of https://www.ardiahealthlabs.com/api/run on 2026-09-01. Retrieval is real: molecular queries return CMS Local Coverage Determinations (L35025, L38045) with working cms.gov links, and a pulmonary GOLD/COPD query returned two real PubMed citations with working links (PMID 40050074, PMID 38032494). What does not exist is any measured performance figure for TARA: no published accuracy, no eval set, no golden set, no clinician or certified-coder adjudication, no clinical outcome. The company's only MEASURED artifacts belong to other models and are both company-reported, not independently reproduced: Cadence (95.45% held-out accuracy, macro-F1 0.9545, subject-independent, public UCI HAR — and not a fall detector) and Meridian's unit-tested CLFS/PAMA arithmetic. Crucible and Sentinel are MODELLED TARGET in the company's own labelling, with Sentinel de-identification specifically marked in-development because of a demonstrated name-detection gap. Everything the framework page says about a deterministic symbolic layer computing a coverage verdict the LLM cannot override is MODELLED TARGET — it is not built. Zeros that must never be softened: 0 customers, 0 pilots, 0 signed BAAs or DUAs, $0 revenue, $0 raised, no real patient data, no clinical outcomes. Pre-revenue, Dallas-Fort Worth, founded December 2025.
Yes, and it must be stated first, not buried. The Studio router is one line in studio.html: ENGINE_MODEL = { molec:'molec', toxiq:'molec', pulmo:'tara', meridian:'tara', aria:'aria', lumen:'lumen' }. Ten named models collapse to four engine paths, and the tara path serves three of the ten names. PulmoIQ and Meridian are not separate models — when a user runs either in the Studio, the request leaving the browser is {"model":"tara"} and the server loads the identical TARA system prompt. There is no pulmonary-specific prompt and no GOLD/GINA corpus in that path. This is verifiable from the outside: posting model:"pulmo" directly to /api/run returns {"error":"bad_model"}. The UI works only because it rewrites pulmo to tara before sending. ToxIQ likewise runs MolecuIQ's molec prompt. Meridian is the sharpest case of two products with one name. Genuinely deterministic CLFS/PAMA compounding arithmetic exists in models/meridian/clfs.py and is mirrored by a client-side calculator on model-pama.html, but it is NOT wired into the Studio answer path — a Meridian question in the Studio is answered by Gemini reasoning as TARA, not by the rate engine. A buyer meets two different Meridians depending on which page they land on. The engineering consequence is the single strongest technical risk in the company: one shared engine path means one prompt regression, or one silent "-latest" model roll, degrades three products simultaneously, with no eval harness anywhere that would detect it. Anyone diligencing "nine models" is really diligencing four prompts and one LLM.
The concrete pain is that a licensed professional must reconstruct by hand the policy chain connecting a denied line item to the rule that governs it, then write an argument that survives a MAC reviewer. Take the case actually run on production: CPT 81445 (5-50 gene solid-tumor NGS panel), documented lung adenocarcinoma C34.11, denied CO-50 'not medically necessary.' Working that denial correctly requires knowing (a) which MAC jurisdiction applies and therefore whether MolDX governs, (b) that L38045/L38158 is the operative NGS solid-tumor LCD and L35025 the umbrella MolDX policy, (c) that NCD 90.2 sets the national floor for NGS in advanced cancer, (d) that MolDX requires an assigned DEX Z-code in the correct claim field (SV101-7 / Box 19), (e) that advanced, recurrent or refractory stage must be documented rather than implied by C34.11, (f) that the 120-day Level 1 redetermination clock runs from the remittance advice date, and (g) which pathology and staging documents to attach. TARA returned an answer referencing several of these elements and retrieved real LCDs with working cms.gov links; its correctness was not adjudicated by a certified coder, was not scored against a golden set, and no accuracy claim follows from a single unblinded run. We estimate 15-40 minutes of expert time per claim from the founders' prior operational experience; this has never been measured against a real worklist. Industry surveys commonly cited in RCM literature put the share of denials never reworked near two-thirds and rework cost at roughly $25 per claim in practice settings and $100+ in hospitals, and quote all-payer initial-denial averages near 11-12% with molecular and toxicology labs reported higher. These are third-party estimates, not Ardia data — Ardia has processed zero claims — and each must carry its source and vintage before use.
Hypothesised buyer map, derived from the founders' prior payer and health-IT experience and NOT yet validated by customer discovery: Ardia has 0 customers, 0 pilots, and no completed buyer-interview programme. On that hypothesis, for the lab lane where TARA's grounded retrieval is strongest, the economic buyer at an independent molecular or toxicology lab is the CFO, or in owner-operated labs the pathologist-owner or COO holding the P&L. The end user is the denials/AR specialist working a several-hundred-line worklist in XiFin RPM, Telcor or Quadax. The champion is the Director of Billing or VP Revenue Cycle, measured on days-in-AR, net collection rate and first-pass yield. The most underrated gatekeeper is the outsourced billing vendor: it controls the 835 files TARA needs and may reasonably view Ardia as margin competition, so TARA either sells through it or must win the lab's willingness to force a data feed. Several of those platforms already ship denial-management and appeal-letter modules, and no competitive teardown of them exists yet, so no differentiation claim against them can currently be made. At a health-system reference lab the purchase additionally traverses an AI governance committee, CISO third-party risk, privacy for the BAA and often clinical informatics — a long cycle with a real chance of dying on 'no SOC 2, one engineer.' That asymmetry is the argument for starting with independent DFW labs. Working assumptions that are unsourced and untested and must be replaced with discovery evidence: signing authority around $250K ACV, outsourced-RCM take rates of 4-7% of collections, health-system AI procurement cycles of 9-15 months, target labs of 50-250 staff. For the pulmonary and elder-care personas riding the same TARA path, the buyer is a different org chart entirely — which is precisely why a core serving everyone serves no single procurement process.
Lead with the boundary. TARA has curated policy-retrieval coverage for exactly eight hand-verified CMS coverage instruments in the molecular, pharmacogenomic and toxicology lanes. Everything else named below is parametric pretraining knowledge supplemented by live PubMed and ClinicalTrials.gov literature retrieval — there is no Ardia-controlled guideline corpus for NCCN, CPIC/PharmGKB, GOLD or GINA, and the GOLD/GINA corpus is explicitly not built. The domains below therefore describe the intended surface, not verified competence, and TARA has no accuracy figure on any of them. Molecular oncology: NSCLC biomarker testing (EGFR, ALK, ROS1, BRAF V600E, KRAS G12C, MET exon 14, RET, NTRK), where MolDX LCDs gate payment on documented advanced, recurrent or refractory stage plus intent to guide targeted therapy; the billing surface is CPT 81445, 81449, 81455, 81457-81459, 81479 and a long PLA tail. Pharmacogenomics: CPIC level-A pairs such as CYP2C19/clopidogrel, DPYD/fluoropyrimidines, TPMT and NUDT15/thiopurines, SLCO1B1/simvastatin, billed 81225, 81226, 81418 under the restrictive MolDX PGx policies. Toxicology: presumptive UDT 80305-80307 and definitive G0480-G0483/G0659, where the denial drivers are frequency, medical necessity and the presumptive-then-definitive sequence. Pulmonary: GOLD group A/B/E escalation logic, GINA steps and T2-high biologic eligibility, spirometry and FeNO coding. A GOLD/COPD probe returned two real PubMed citations with working links (PMID 40050074 on asthma guidelines, PMID 38032494 on LABA/ICS overuse in COPD), so the pulmonary lane is grounded in retrieved literature — but literature is not a guideline corpus, and no clinician has reviewed any pulmonary output. Elder care: functional-decline trending, explicitly non-diagnostic and explicitly not fall detection. Imaging is out of scope entirely: TARA explains an imaging report's text and never the image.
Verified against source and the live endpoint. (1) The browser posts JSON to same-origin /api/run: {model, text, engine, code, attest_synthetic, attachments}. The Studio maps the chosen persona through ENGINE_MODEL, so PulmoIQ and Meridian arrive as model:'tara', and maps the tier selector to engine: fast, '' or deep. (2) api/run.py caps input at 6,000 characters and optionally checks a shared access code, currently unset in production (gated:false). (3) Attachments do not work and are disabled in two independent places: the API returns {"error":"uploads_disabled"} for any attachment, and studio.html hardcodes attachments:[] so a chosen file is read to base64 and discarded. Image, X-ray and MRI analysis is not a capability. (4) Sentinel (models/sentinel/deidentify.py) runs regex Safe-Harbor redaction before anything leaves the process. IMPORTANT AND UNRESOLVED: it does not reliably detect plain personal names — on a live probe 'John Smith' passed through to Google untouched, as did an MRN whose internal space defeated the token pattern. Structured identifiers (SSN, phone, MRN, dates, ZIP) do redact. Sentinel is marked in-development and its own docstring declines to claim name detection. (5) Retrieval runs on de-identified text only: policies.match_policies() keyword- and CPT-scores eight hand-verified CMS LCDs, and research.gather_sources() queries PubMed and ClinicalTrials.gov, both best-effort. (6) Matched sources are appended as a numbered grounding block. (7) One generateContent call to Gemini with systemInstruction = guardrail preamble plus the TARA persona prompt. (8) Crucible runs six deterministic gates on the output — de_identification, non_diagnostic, scope_of_practice, safety_escalation, honesty, human_in_the_loop — each returning a verdict with a reason on every call. On ANY gate failure production withholds the answer entirely, enforced in code and verified on production. Two further controls, cite-or-abstain and policy-override, are enforced in the retrieval and answer-binding layer and are deliberately NOT counted as gates. The gate count is six and only six; earlier '8 gates' claims were wrong and have been removed. Crucible carries modelled-target maturity: six gates run, but the logic is string and regex matching, not a validated classifier. (9) A PHI-free audit event is recorded. Per source reading and not exercised by probe, call_model is written to fail closed if the guardrail module cannot import; that failure path has not been tested. Honest summary: retrieval-augmented generation with a pre-filter and a post-filter — not a system in which a symbolic layer computes a coverage verdict the LLM cannot override.
What enters: free text up to 6,000 characters — a denial letter, an EOB excerpt, a coding question, a clinical scenario. Attachments are rejected outright; the API returns uploads_disabled and the Studio discards any chosen file client-side, so no image or PDF ever reaches the model. What is redacted before the model: email, URLs (public citation hosts whitelisted), IPv4, SSN in NNN-NN-NNNN form, labelled MRN, member-ID, account, licence and device-serial tokens containing a digit, phone and fax, numeric and Month-D-YYYY dates, and ages 90+. What is NOT redacted, verified live: personal names. A probe sending 'Patient John Smith, MRN 44 82910, seen 03/14/2025, phone 214-555-0148' returned sentinel {removed: 2, categories: ['date','phone_or_fax']} — the name reached Google, and so did the MRN, because the pattern requires a four-character alphanumeric run after the label and the space split it. What reaches the model: de-identified text, the numbered grounding block (LCD titles, IDs and cms.gov URLs; PubMed titles, journals, years and URLs; NCT IDs), and the concatenated guardrail plus persona instruction. What is logged BY ARDIA: nothing per-request — log_message is overridden to silence client IP and path — plus a PHI-free audit event carrying model key, redaction counts and gate outcomes. What third parties log is outside Ardia's control: Google receives the full prompt under its own retention and abuse-monitoring policy and Ardia holds NO BAA with Google; NCBI and ClinicalTrials.gov receive the extracted keyword query under their own terms. Under this architecture TARA legally cannot process PHI at all. Gate verdicts and redaction counts are returned to the caller, which makes behaviour externally observable — but observability is not correctness: the same response reported removed:2 while a name and an MRN reached Google. The counters verify the filter ran, not that it worked.
Partly implemented, and the line between implemented and merely named matters more than the list itself. Hard-coded in policies.py, each with a cms.gov MCD deep link: LCD L35025 (MolDX umbrella), L38045 and L38158 (MolDX NGS for solid tumors), L38294 and L38335 (MolDX pharmacogenomics), L36393 (controlled substance monitoring and drugs of abuse), L34645 and L36029 (urine drug testing). Two of the eight — L35025 and L38045 — were independently confirmed to return through the live API with working links. The company reports that all eight were hand-verified against the live CMS MCD on 2026-08-07 because CMS publishes no MCD JSON API; that verification is company-reported and has not been reproduced. Hand-curation also means the set goes stale silently whenever CMS revises a policy: there is no refresh mechanism, no revision-date check and no staleness alarm. The matcher keys on CPT 81445/81449/81455/81457-81459/81479, PLA 0244U/0250U/0329U/0334U/0379U/0391U/0543U, PGx 81225/81226/81418, presumptive tox 80305-80307 and HCPCS G0480-G0483/G0659. Referenced by the model from parametric memory but implemented nowhere in code: NCD 90.2, MolDX DEX Z-codes, the CARC/RARC denial taxonomy, NCCI PTP edits and MUEs, ICD-10-CM, X12 837P/835/277CA/275, HL7 v2 and FHIR R4, LOINC, SNOMED CT, NCCN, CPIC/PharmGKB, GOLD and GINA. Two licensing facts belong in any plan and are routinely missed: CPT descriptors are AMA-copyrighted and programmatic use requires a paid AMA CPT data licence, and GOLD, GINA and NCCN content is copyrighted and cannot be scraped into a corpus. SNOMED CT is free in the US via an NLM UMLS licence, LOINC free with registration, CPIC openly licensed, and CMS NCCI PTP/MUE and CLFS quarterly files are free downloads — which is exactly why any real symbolic layer should be built from the CMS files first.
MODELLED TARGET — none of this is built. The intended v1 integration is deliberately boring: a nightly SFTP drop of the lab's X12 835 remittance and matching 837P claim files into US-resident infrastructure, parse denials by CARC/RARC, match the governing LCD, draft an appeal packet per claim, and write back a prioritised worklist CSV and draft PDFs over the same SFTP. Current state: no X12 835/837 parser exists, no CARC/RARC table exists, no PDF generation exists, no SFTP surface exists, and no VPC exists — production today is a Vercel serverless function calling the Google Gemini API with no BAA, no SOC 2 and no US-data-residency attestation. Zero labs have sent a file because there are zero customers. The virtue of the design is that it needs no EHR, no FHIR and no clearinghouse contract, which is why it is the only integration a pre-BAA, single-engineer, $0-raised company can credibly attempt — and even it is unbuilt. The next ring out is the lab's billing platform (XiFin RPM, Telcor, Quadax, LigoLab) and the LIS on the specimen side; each is simultaneously a channel and a blocker. Beyond that sits the clearinghouse layer carrying 837P out and 277CA and 835 back. The piece most decks miss: filing a Medicare appeal electronically requires either the MAC's own portal (Palmetto eServices, Noridian Endeavor, NGSConnex) or the esMD gateway through a certified Health Information Handler. Ardia is neither, so 'denied claim to filed appeal' honestly ends at a human downloading a PDF and uploading it. Four operational questions remain unanswered and must be before a first file moves: who runs the SFTP endpoint, how US residency is enforced under Texas SB 1188, how the LDS/DUA is papered, and what resolves the Google-no-BAA problem once real remittance data is in scope.
Blunt version. 0 customers, 0 pilots, 0 signed BAAs or DUAs, $0 revenue, $0 raised, no real patient data processed, no clinical outcome of any kind. TARA specifically has no published accuracy figure, no held-out benchmark, no golden set, no inter-rater agreement study against a certified coder, and no measurement of the only thing that matters commercially — whether a TARA-drafted appeal is overturned more often than the status quo. The company's two MEASURED artifacts belong to other models and both are company-reported, not independently reproduced here: Cadence's 95.45% held-out accuracy and macro-F1 0.9545, subject-independent on the public UCI HAR dataset with a scikit-learn logistic regression (which proves the pipeline, not elder-cohort performance, and which is not a fall detector); and Meridian's unit-tested CLFS/PAMA compounding arithmetic, which is real deterministic code but is not wired into the Studio answer path. '34/34 tests passing' is likewise company-reported and has not been independently executed. From reading the test files, their scope appears to be regex behaviour, PAMA arithmetic and gate string-matching — zero tests assert that any answer is clinically or coding-correct. So even a green suite is evidence the plumbing behaves, not that the product is right. What was independently verified on 2026-09-01 and any skeptic can re-verify: the endpoint is live, unauthenticated and Gemini-backed; the Fast tier serves gemini-flash-lite-latest in about 3 seconds and Scholar serves gemini-flash-latest in about 54 seconds; six gates are computed and returned with reasons on every call and a failure withholds the answer; molecular queries return real CMS LCDs with working links; a pulmonary query returns real PubMed citations with working links; attachments are refused; and Sentinel removes structured identifiers while missing plain names. TARA's correct label is LIVE DEMO.
The first non-zero number should be TARA's governing-policy selection accuracy, because it needs no PHI, no BAA and no customer, and it is the claim every downstream promise rests on. Gold set: 400 denial cases in the molecular, PGx and toxicology lanes, each a denied line with CPT/HCPCS, ICD-10, MAC jurisdiction, CARC/RARC and the payer's stated rationale — bootstrapped from public CMS MCD text plus synthetic denial letters, with an explicit split across at least two MAC jurisdictions and a deliberate share of cases where the correct answer is 'no LCD applies.' Labelling: two independent certified coders (CPC or CCS-P with molecular and toxicology experience), each assigning the governing policy and the dispositive criterion, with a third adjudicating disagreements and Cohen's kappa reported. Coder labelling is the critical path and Ardia has not contracted it. Denominator: all 400 cases, with abstentions counted as errors in the primary and reported separately in the secondary. Comparators, both required: bare Gemini with no retrieval and no grounding block, and a second arm of the current human process where an appeal record exists. Pre-registered primary metric: top-1 governing-policy accuracy with 95% Wilson intervals, powered at roughly plus-or-minus 5%. Pre-registered secondary: citation faithfulness — the share of asserted criteria that a coder confirms actually appear in the cited policy text — and an abstention-appropriateness rate. Kill criteria, fixed in advance: TARA does not work if top-1 accuracy fails to exceed the ungrounded Gemini baseline by a margin whose confidence interval excludes zero; if citation faithfulness falls below 90%, since an unfaithful citation inside a signed attestation is the liability pathway above; or if abstention on out-of-scope cases falls below 80%. A fourth kill criterion applies at pilot: if the overturn-rate lift interval includes zero, the product does not pay for itself and the price mechanic collapses.
A skeptic can verify all of this in under ten minutes and should. (1) Which engine runs: curl https://www.ardiahealthlabs.com/api/run returns {"ok":true,"provider":"gemini","gated":false} — Google, not Anthropic, and the endpoint is unauthenticated. (2) A governed run with real CMS citations: POST {"model":"tara","text":"Medicare denied CPT 81445 NGS solid tumor panel for documented lung adenocarcinoma ICD-10 C34.11, denial CO-50 not medically necessary. Which LCD governs and is this appealable?","engine":"deep","attest_synthetic":true}. It returns six gate verdicts each with a reason, and sources containing L35025 and L38045 with working cms.gov deep links — open them and confirm the policies are real and active. Its coding correctness has not been adjudicated by any coder. (3) The tier and model facts: the fast tier returns model_id gemini-flash-lite-latest in about 3 seconds, the deep tier gemini-flash-latest in about 54 seconds — and note that '-latest' means the served model can change without a deploy. (4) The pulmonary lane: POST a GOLD/COPD escalation question. It returns real PubMed citations (PMID 40050074, PMID 38032494) with working links — grounded in literature, not in a GOLD or GINA corpus, because no such corpus exists. (5) The de-identification gap: POST text containing 'Patient John Smith, MRN 44 82910, seen 03/14/2025, phone 214-555-0148' and read sentinel {removed:2, categories:['date','phone_or_fax']} — the name and the spaced MRN reached Google. (6) Imaging: attach any file; the API returns {"error":"uploads_disabled"}, and grep studio.html for attachments:[] to see the client discard it too. (7) The shared engine: grep the Studio page for ENGINE_MODEL and see ten names map to four engines, then POST model:"pulmo" directly and get {"error":"bad_model"} — proof the UI rewrites it to tara. (8) The code: clone the public repo and read api/run.py, the call_model pipeline, policies.py (the entire symbolic layer), models/crucible/gates.py and models/sentinel/deidentify.py, whose own docstring declines to claim name detection.
Read this section first; it is worse than the marketing implies. (1) TARA is a prompt with a label — roughly 90 words of persona text on a roughly 450-word shared preamble, sent to Google Gemini. No owned weights, no fine-tuning, no distillation, no eval harness, no accuracy figure of any kind. The defensible artifacts are a regex de-identifier, six regex gates and a small hand-curated table of eight LCDs. (2) The symbolic layer on the framework page does not exist. match_policies() scores keywords and CPT substrings and returns titles and URLs; it does not read policy text, does not evaluate coverage criteria and emits no verdict. NCCI, MUEs and the CARC/RARC taxonomy named on that diagram appear nowhere in the repo. (3) The tiers are not depth — they set a model-id preference order and nothing else. Scholar took about 54 seconds and still landed on a Flash model. (4) The engine is not reproducible: '-latest' aliases resolve at call time, so the served model can roll forward silently with no version pin and no output hash. A system claiming a CLIA/CAP-grade audit trail cannot have that. (5) Sentinel misses plain names — 'John Smith' reached Google, as did a spaced MRN. Name is Safe-Harbor identifier #1, and there is no BAA behind the regex. (6) The gates check surface form, not truth. They cannot detect a fabricated LCD criterion, a wrong CPT descriptor, a superseded policy revision or a miscited Z-code requirement. (7) Even returned sources are titles and URLs, not policy text, so the model cites links it has not read. (8) Pulmonary grounding is PubMed literature, not a GOLD/GINA corpus, and that corpus is not built. (9) Imaging does not work at all. (10) The production endpoint is unauthenticated with no visible rate limit on a paid third-party API. (11) No prompt-injection defence on a system whose primary input is payer-authored correspondence. (12) One shared engine path serves three products with no eval harness, so one regression degrades all three invisibly. (13) Bus factor one, on a product whose entire value proposition is institutional trustworthiness.
TARA's defensible position is that it is administrative and decision-support software, non-diagnostic always, with a licensed human between the draft and any action. Our own assessment — NOT an opinion of regulatory counsel, never tested with FDA, with no pre-submission, no correspondence and no legal memo of record — is that revenue-cycle and coverage-appeal work falls under the 'administrative support of a health care facility' exemption at FD&C Act 520(o)(1)(A) as amended by section 3060 of the 21st Century Cures Act, and that CLIA (42 CFR 493) does not apply because TARA performs no testing and reports no result. The clinical-reasoning lane needs more care: to stay inside the 520(o)(1)(E) clinical decision support exemption a function must not process a signal from a device or IVD, must display or analyse medical information, must provide recommendations rather than a specific directive, and must let the clinician independently review the basis — a criterion FDA's September 2022 CDS guidance reads narrowly. The '## How I got there' section, the citation block and the human-in-the-loop and non-diagnostic gates are the artifacts that argument depends on, and they are not yet reliable: a pulmonary therapy-escalation answer is grounded in retrieved PubMed literature rather than in the governing GOLD or GINA text, so the basis a clinician would need to review is the wrong instrument, and it still passed all six gates. That is an open regulatory gap, not a design strength. HIPAA: Ardia would be a Business Associate once it touches PHI; the self-graded control matrix stands at 2 of 15; there is no SOC 2 and no third-party audit. Texas SB 1188 imposes US data residency and physician-review requirements, and TRAIGA adds disclosure, documented intended use, human oversight and AG enforcement. The hardest fact is contractual, not statutory: no BAA with Google means TARA cannot lawfully see PHI today.
Non-diagnostic does not dissolve liability; it moves it from FDA to CMS, OIG and contract law, which for a lab billing product is the harder surface. Trace the harm pathway concretely. TARA drafts a redetermination request. A human at the lab signs it, and that signature is an attestation that the submission is true, accurate and complete. Suppose TARA cites L38045 for a claim actually governed by a different jurisdiction's policy, or asserts a coverage criterion the cited LCD does not contain, or reproduces a superseded revision from the hand-curated table that CMS has since changed. All six gates pass, because the gates are regexes over the final string and cannot read the policy. The appeal is filed. If it succeeds, the lab has been paid on a representation that was not true. That is the fact pattern the False Claims Act reaches under 31 USC 3729, including the reverse-false-claims theory for retaining an overpayment, with treble damages and per-claim penalties, and it is the fact pattern OIG already watches: urine drug testing is a named OIG enforcement priority and MolDX-governed molecular testing sits close behind. Payers and MACs are beginning to flag AI-drafted appeals, which raises both scrutiny and the odds that a systematic error is found across a whole batch rather than one claim. Ask the harder question first: can a lab compliance officer operating a seven-element program under OIG compliance guidance permit an unvalidated AI tool into the appeal workflow at all? Without an accuracy figure, a validation record, a change-control process or a documented human-review step, the honest answer is probably not — which makes evaluation a sales prerequisite, not a research luxury. And Ardia has never written the contract where this is allocated: no MSA, no indemnification clause, no limitation of liability, no tech E&O or professional liability coverage, no defined error-reporting or correction obligation. Today, the lab holds all of it.
The cheapest legally clean instrument is a HIPAA Limited Data Set under a Data Use Agreement (45 CFR 164.514(e)), which may retain dates and ZIP3 and strips the sixteen direct identifiers — sufficient for 835/837 denial analytics and for scoring LCD selection. A full BAA is required only once an appeal packet must carry the patient's actual record, that is at pilot, not at validation. Show the statistical derivation rather than asserting numbers: n approximately 384 gives a plus-or-minus 5% half-width at 95% confidence for a proportion near 0.5, so roughly 400 independently adjudicated denial cases is the rounded requirement for a single-arm accuracy estimate — and that assumes independent cases and certified-coder adjudication with inter-rater agreement still unmeasured. The 5,000-denied-line figure previously floated for an overturn-rate comparison is an unpowered guess: no effect size was assumed and no power calculation was run. It must be rebuilt from an explicit assumed baseline overturn rate and a target lift before it appears in any plan. A partial golden set can be bootstrapped from public MCD text and synthetic denial letters, but no timeline should be committed: Ardia has one engineer, $0 raised, and no contracted certified coder, and coder labelling — not text assembly — is the critical path. For Sentinel the validation set is public and free: the i2b2/UTHealth 2014 de-identification corpus, scored on per-category recall, published including the name recall that is currently near zero. Our reading is that retrospective denial analytics for operations and product validation would not constitute human-subjects research under 45 CFR 46.102, but Ardia has obtained no not-human-subjects determination and no IRB or counsel has reviewed this; a written determination from a commercial IRB should be obtained before any publication relies on the position. For a pulmonary corpus the binding constraint is licensing, not volume.
TARA has no standalone TAM — it is a cost centre that makes the persona SKUs possible, and any deck giving TARA its own market number is double-counting. Bottom-up sizing of the market it enables, from published industry counts that should each be sourced before use: on the order of 7,000-8,000 CLIA-certified high-complexity independent labs in the US, of which perhaps 1,200-1,500 do meaningful molecular, genetic or toxicology volume and therefore carry MolDX exposure. At the ACV derived in the pricing dimension, that is a serviceable market in the low hundreds of millions in the lab lane alone. The PAMA overlay supplies the timing argument and is the one piece of arithmetic here that is deterministic rather than assumed: the statutory cut is capped at 15% per year for 2027, 2028 and 2029 and compounds on the prior year's already-reduced rate, so a lab's final-year run-rate is 0.85 cubed, or 61.4% of baseline. For a lab with $10M of CLFS revenue that is $1.5M plus $2.775M plus $3.859M — $8.13M of cumulative shortfall across three years and a permanent hole of roughly $3.9M per year thereafter. That is why a CFO takes the meeting in 2026, and it is the honest bridge between TARA and the PAMA pillar. Everything downstream of that arithmetic is modelled: denial rate, current appeal coverage, current overturn rate, and any lift TARA produces. Ardia has measured none of them on a single real claim, holds no 835 files and has no customer. The first pilot's only real job is to replace those four assumptions with observed numbers; until it does, the market model is a hypothesis with a spreadsheet attached.
Pick one mechanic and hold it: $15 per drafted appeal packet, with a $3,000 monthly platform minimum credited against usage. The drafts contained three incompatible models — flat SaaS, per-appeal, and 15-25% contingency — and only the per-appeal fee matches how the product actually creates value, is forecastable by the buyer, and starts small enough for a CFO to approve without a committee. Inference economics are not the constraint. A TARA call carries roughly 3,000 input tokens (guardrail preamble, persona prompt, grounding block, up to 6,000 characters of user text) and returns on the order of 1,200 output tokens. On Flash-Lite that is a fraction of a cent per call; on the Scholar-tier Flash model it is roughly half a cent. Even at three calls per packet and generous headroom, inference costs under two cents per $15 packet — better than 99% gross margin on compute. Real COGS sits elsewhere: implementation and support, and the certified-coder QA sample that the liability and evaluation dimensions both require. Budget a quarter-FTE coder at roughly $95K loaded reviewing a 10% sample, or about $24K per account per year, which puts realistic blended gross margin near 70% at a single account and improving with scale. Buyer arithmetic, explicitly modelled and not observed: in the scenario where TARA lifts appeal coverage from 35% to 80% on 13,200 denied lines at $420 average allowed with overturn moving 45% to 50%, incremental recovery is roughly $1.35M against 5,280 packets priced at $79,200 — about 17x. Even at one-fifth of the modelled lift it clears 3x. That headroom is why the CFO will counter with contingency at 15-25% of recovered dollars, which on this scenario is $200-340K — worse for the buyer but a fair trade for Ardia's absent evidence. Ardia should accept contingency on the first three accounts precisely because it cannot yet prove the lift.
On the reasoning-core-plus-governance claim, the incumbents are formidable and mostly free. General guardrail frameworks: NVIDIA NeMo Guardrails, Guardrails AI, Microsoft Presidio, Azure AI Content Safety with groundedness detection, Patronus, Galileo, Lakera. The most direct conceptual competitor is Amazon Bedrock Automated Reasoning Checks — literally a neuro-symbolic policy layer, backed by SMT solvers from the same formal-methods lineage as AWS Zelkova, shipped with AWS distribution and a HIPAA-eligible BAA. Against that, Ardia's symbolic layer is a short Python file that keyword-scores eight LCDs; the gap is enormous and pretending otherwise in a diligence room is fatal. On PHI de-identification: Private AI, John Snow Labs Healthcare NLP, AWS Comprehend Medical PHId — all with trained NER, published benchmarks and BAAs, all of which currently beat Sentinel outright. On the actual business, lab and provider denial management: Waystar, Experian Health, Optum, R1, Availity, Adonis, Anomaly, Janus, Rivet, and lab-native XiFin, Telcor, Quadax, hc1 and LigoLab, several of which already sit on the 835 files and already market AI denial insights and appeal-letter generation. No teardown of what those modules actually ship has been completed, so no differentiation claim against them can currently be defended. The cheapest competitor of all is a biller pasting an LCD PDF into ChatGPT for free. The honest differentiation is narrow, and narrow is why it is worth defending: gate verdicts are returned in the API response on every call, so a buyer watches governance run rather than reading about it; retrieval is grounded in hand-verified, date-stamped CMS policies with working MCD deep links; and a failed gate withholds the answer rather than annotating it. What is NOT differentiation: the LLM (it is Google's), the prompt scaffolding, the tier selector, or the phrase neuro-symbolic.
Sequenced, cheap credibility fixes first because they cost days and change the diligence conversation. Days 0-30, stop the bleeding: pin an explicit Gemini model id and version rather than resolving '-latest', version the system prompt, and log model id, prompt hash and output hash into the audit trail so a run is reproducible; set the access code and add rate limiting on the public endpoint; reconcile every remaining eight-gate assertion in docstrings and tests to the six-gate reality; add streaming so the deep tier stops feeling broken; build a minimal regression harness over the shared engine path so a prompt change to TARA cannot silently degrade PulmoIQ and Meridian. Days 30-90, earn the first number: replace regex name redaction with real NER and publish per-category recall on the public i2b2 2014 corpus including the honest name figure — without this no BAA can be signed in good conscience, and no BAA means no pilot and no revenue; and stand up the golden set and evaluation described in evaluationDesign. Days 90-180, build the layer that is currently fiction: ingest full LCD text and coverage criteria for the top twelve MolDX and UDT policies into a structured rules table with revision-date tracking, load the free quarterly CMS NCCI PTP, MUE and CLFS files, wire Meridian's deterministic clfs.py into the Studio answer path so one Meridian exists instead of two, and bind every assertion to a retrieved policy passage rather than a URL. Days 90-270 in parallel: an X12 835/837P parser and an SFTP ingest and return surface. Days 180-365: LDS/DUA with one DFW lab, measured overturn lift against their own baseline, then BAA and SOC 2 Type I, then paid pilot. The single unlock is one signed Limited Data Set DUA giving Ardia twelve months of 835 files; everything else is downstream of that signature.
