MolecuIQ is a guardrailed assistant running on Google Gemini (gemini-flash-lite-latest, an unpinned "-latest" alias) — not Claude — that reads a pasted molecular-lab denial in free text, classifies it, is instructed by its system prompt to refuse appeals the record does not support, and drafts an appeal citing a hand-curated table of eight CMS Local Coverage Determinations; its engine path is shared verbatim with ToxIQ and with the site's separately listed "precision medicine" model, and it is a live demo with zero customers, zero pilots, zero claims ever processed, zero appeals ever submitted, and no published accuracy of any kind.
MolecuIQ sits across pillars one and two — precision medicine, and molecular and genomic diagnostics. The site lists it twice, once under "Lab reimbursement (RCM)" and once under "Precision medicine," and both listings resolve to the same molec system prompt. Whether it is unique in straddling two pillars is not established here; other models plausibly straddle too. Pillar three, the 2027 PAMA rate cliff, is the commercial adjacency: Meridian's rate-cliff framing is the intended door-opener into the same molecular-lab CFO. Note now rather than later that a buyer meets two different Meridians — genuinely deterministic CLFS/PAMA 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, so a Meridian question asked in the Studio is answered by Gemini reasoning under TARA's prompt. Pillar four, pulmonary and respiratory care, and pillar five, elder care, are untouched by MolecuIQ; they belong to PulmoIQ, Aria and Cadence. Toxicology reaches MolecuIQ only because ToxIQ is routed to the same prompt, which is why three drug-testing LCDs sit in a molecular product's policy table. Molecular and genomic diagnostics is also where a kidney-genomics SKU would sit if one were built; none is. The honest summary of the positioning: one prompt, two pillar listings, a third pillar attached commercially through a calculator that does not run in the product a prospect is shown.
MolecuIQ runs in production today under guardrails, verified by direct probe of https://www.ardiahealthlabs.com/api/run on 2026-09-01, with no published accuracy, no customers, no pilots, no claims ever processed and no measured outcome. One thing is independently verified: the live engine identity. The test-suite figure is not. The company publishes "34/34 tests passing"; a reviewer run reported 75 across nine files; neither could be reproduced from the code available for this review, where the nearest checkouts show six test files and 46 test functions. Treat any test count as company-reported and unverified. More important than the count: whichever number is right, those tests cover the deterministic scaffolding only — zero tests assert that any MolecuIQ denial classification, appealability verdict or appeal draft is correct. The "precision medicine / NCCN + CPIC interpretation" capability described on the site is closer to a modelled target than a live demo: it produces answers, but nothing in the retrieval layer supplies NCCN or CPIC content, so those answers are ungrounded parametric generation. Retrieval does work where a corpus exists — molecular queries return real CMS Local Coverage Determinations with working cms.gov links, independently confirmed for L35025 and L38045.
Stated prominently, because the site obscures it. The Studio routes ten named models through four engine paths: ENGINE_MODEL = { molec:'molec', toxiq:'molec', pulmo:'tara', meridian:'tara', aria:'aria', lumen:'lumen' }. ToxIQ is MolecuIQ — toxiq maps to the molec key and is served by MolecuIQ's system prompt verbatim. The site's third entry, "MolecuIQ — genomics and pharmacogenomics" under Precision Medicine, is also the same molec prompt. Three of the ten advertised models are one artifact. That prompt opens "You are MolecuIQ, Ardia's molecular/toxicology denial-recovery assistant. This is ADMINISTRATIVE revenue-cycle work, not clinical care. You never diagnose," and contains zero instructions about NGS variant interpretation, star-allele calling, NCCN or CPIC. The precision-medicine product is therefore a revenue-cycle prompt answering genomics questions from parametric memory. One probe returned an answer consistent with the CPIC clopidogrel guideline; that is n=1, self-graded, against no reference set, from a pipeline that supplies the model no CPIC content, so it demonstrates fluency, not grounding. The routing is also load-bearing in a way a buyer would not guess: posting model:"pulmo" directly to the API returns {"error":"bad_model"}, and the UI works only because it rewrites pulmo to tara before sending. The structural consequence is the strongest technical risk in the company: one prompt regression, or one silent roll of the unpinned Gemini alias, degrades three products at once, and there is no eval harness anywhere that would detect it. A buyer told "we have a precision-medicine model, a toxicology model and a molecular RCM model" is being told there are three assets where there is one prompt.
Molecular and genomic testing is structurally the worst-reimbursed high-value service line in American medicine. A lab performs a $600-$3,000 NGS panel before it knows whether it will be paid, and the claim carries policy risk on five independent axes at once: CPT selection across Tier 1 (81105-81479), Tier 2 (81400-81408), Genomic Sequencing Procedures (81445, 81449, 81450, 81455, 81456, 81457-81459), MAAAs (81490-81599) and proprietary PLA 0xxxU codes, where picking a GSP when a PLA exists is an automatic denial; MolDX technical assessment and a registered DEX Z-code that must be on the claim or the front end rejects it; the governing LCD's medical-necessity criteria (L35025, L38045/L38158, L38294/L38335) plus NCD 90.2; ICD-10 linkage, where a benign or screening diagnosis silently kills coverage; and frequency, duplicate and NCCI MUE edits. Three statistics usually carry this argument and each needs a named source, a year and a scope note before it is usable in diligence: the first-pass denial rate on molecular claims (state whether it is molecular-specific or extrapolated from all-payer data), the per-denial rework cost (the commonly quoted $25-$118 range traces to all-specialty survey data, not molecular), and the share of denials never reworked (an all-payer figure not, to our knowledge, measured on molecular claims). Ardia has measured none of them and has processed zero claims. The expensive part is abandonment, not denial: an abandoned appealable claim forfeits revenue on a test whose reagents and sequencer time are sunk. It is not '100% margin' - expected recovery is the overturn rate times the allowed amount, less the coder time the appeal consumes, and Ardia cannot supply an overturn rate. If CLFS rates step down as scheduled, headroom for tolerating write-offs narrows; that is the thesis for 2026-2027 urgency, not a measurement.
The buyer hypothesis - and it is a hypothesis, not a finding - is that the economic buyer is financial rather than clinical. In an independent molecular laboratory the cheque would be signed by the CFO or VP of Revenue Cycle, with the Billing Manager as champion and daily user; in a smaller lab the CEO or COO signs personally, because the founder is usually the pathologist who started it. Where billing is outsourced, the true buyer moves to the lab-billing company and becomes its VP of Operations, who is looking for margin on a fixed-fee contract. A lab-billing company would plausibly be the better first customer - one contract touches many labs' volume, one integration, one BAA, and the counterparty already understands 835 remits. That reasoning is untested; no such conversation has happened. The quantified specifics usually attached to this map - a $10M-$150M net revenue band, 2-8 coder FTEs, the share of labs using outsourced billing, an eighteen-month enterprise sales cycle - are the founder's priors from prior healthcare-IT experience, not measured segmentation, and none has been tested against Ardia's pipeline, which is empty. The blocker to signature is the Compliance or Privacy Officer, who will stop this on the BAA question, correctly, given the current architecture. For an appeal that reaches redetermination, a Medical Director or pathologist signs the letter, so MolecuIQ's output is a draft into a clinician's hands, never a submission. Note the constraint on any recovery-linked pitch: Ardia has classified zero denials against a known outcome, submitted zero appeals and observed zero overturns, so recovery-share pricing would have to be priced on the customer's historical data, not Ardia's.
Lead with the boundary. None of the biomarker, germline or pharmacogenomic content below is encoded, retrieved or tested anywhere in MolecuIQ. There is no variant table, no star-allele-to-phenotype logic, no CPIC or PharmVar corpus, and no HGVS, VCF, LOINC or SNOMED handling. This is the buyer's clinical world, described so the reimbursement logic makes sense; it is not a capability inventory. The domain is somatic and germline molecular diagnostics plus pharmacogenomics, seen through a reimbursement lens. Somatic oncology: comprehensive genomic profiling in NSCLC (EGFR exon 19 del / L858R, ALK, ROS1, BRAF V600E, MET exon 14 skipping, RET, KRAS G12C, NTRK, with PD-L1 and TMB), colorectal (KRAS/NRAS/BRAF, MSI-H/dMMR), melanoma, breast, prostate, and pan-tumour MSI/TMB. Germline: hereditary panels (BRCA1/2 81162-81167, Lynch MMR genes) gated by NCCN high-risk assessment criteria. Pharmacogenomics: CYP2C19 (81225), CYP2D6 (81226), CYP2C9/VKORC1, TPMT/NUDT15, DPYD, UGT1A1, HLA-B57:01, HLA-B15:02, and panel code 81418. NCCN biomarker recommendations and NCD 90.2 generally support NGS in recurrent, relapsed, refractory, metastatic or advanced-stage cancer with therapeutic intent, which is the substantive reason a benign-nodule case should be refused. On one probe MolecuIQ did refuse such a case; neither NCD 90.2 nor NCCN is in the retrieval layer, so that refusal was an ungrounded generation that landed correctly on a single run, and refusal behaviour has never been measured across a set. MolecuIQ sits at the last step of the lab lifecycle: 835 remit returns with a CARC/RARC, the biller triages, and the claim is appealed or written off.
As it executes in production, verified by probe and by reading deployed code. (1) INPUT: a client POSTs {model:'molec', text, ground:true} to /api/run, capped at 6,000 characters. There is no structured claim object; it is a text box. Attachments do not work: the API returns {"error":"uploads_disabled"} for any attachment, and studio.html hardcodes attachments:[] so a chosen file is read to base64 and discarded. A scanned denial letter or EOB PDF - the form most denials actually arrive in - cannot be ingested at all. (2) DE-IDENTIFICATION: Sentinel runs over the text before anything leaves the process, as pure regex across the pattern-detectable subset of the 18 HIPAA Safe Harbor identifiers. Sentinel does not reliably detect plain personal names. This is a known gap, marked in-development, not a design choice: names are redacted only from a caller-supplied roster, no roster is wired in production, and on a live probe 'John Smith' passed through unredacted to Google. Structured identifiers redact reliably. call_model is coded to fail closed if the guardrail module fails to import; that path has not been exercised in production in any evidence available here. (3) RETRIEVAL: match_policies() is a pure function with company-reported unit-test coverage this review could not reproduce; it extracts CPT/HCPCS/PLA codes and keywords and scores them against eight LCDs, returning up to three citations with cms.gov links. gather_sources() then hits PubMed E-utilities and ClinicalTrials.gov v2. (4) REASONING: Gemini is called with the guardrail preamble plus the MolecuIQ prompt. (5) GATES: all six run, with administrative=True, which relaxes the non_diagnostic gate - five gates at full strength plus one deliberately relaxed, with no measured false-negative rate. (6) WITHHOLDING: a failed gate withholds the answer entirely, enforced in code and verified on production. (7) AUDIT: a PHI-free event is written.
What enters: free text pasted by a biller - realistically a denial letter body or 835 remit line carrying a CPT or PLA code, a CARC/RARC, a payer, an ICD-10, and, unless the biller is careful, patient identifiers. Cap 6,000 characters; attachments are disabled in two independent places. What is redacted before egress, verified live: sending 'Patient John Smith, MRN 4482910, SSN 123-45-6789, DOB 03/12/1958, phone 214-555-0134' returned {'removed': 4, 'categories': ['date','mrn','phone_or_fax','ssn']}. Four structured identifiers stripped. What is not redacted and does reach Google: the patient's plain personal name, the free-text clinical narrative, the ICD-10 and CPT codes, the payer, the ordering physician's name and the assay description. Public citation domains are exempted from URL redaction so citations survive. What reaches the model: de-identified text plus an appended grounding block of retrieved LCD and PubMed citations, sent over TLS to generativelanguage.googleapis.com - Google AI Studio, a third party with no signed BAA and no BAA availability on that API surface. The same de-identified text also goes outbound as a keyword string to NCBI E-utilities and ClinicalTrials.gov, a second, less obvious egress path nobody has written down. Because Sentinel does not redact plain names, a patient's name can reach all three. What returns: markdown prose - denial classification, root cause, appealability verdict, appeal strategy, LCD citations, and a mandatory 'Not a diagnosis' closing line - plus machine-readable sentinel and crucible blocks. Upstream provider error bodies are drained server-side and never returned, because an error body can echo the prompt. Ardia persists nothing patient-identifiable, by design and unaudited: a PHI-free audit event only. There is no database, no case file and no outcome record, which is why the product cannot learn from, or report on, anything it has done.
Genuinely wired in code (ardia-studio-app/policies.py): eight CMS Local Coverage Determinations - five MolDX molecular and pharmacogenomics policies (L35025 the MDT umbrella, L38045 and L38158 NGS for solid tumours, L38294 and L38335 pharmacogenomics) plus three drug-testing policies (L36393 controlled substance monitoring, L34645 and L36029 urine drug testing) that are present only because ToxIQ shares this engine path. The eight identifiers are confirmed present in the file; the author reports hand-verifying all eight against the live CMS MCD on 2026-08-07, and independent confirmation of working cms.gov links exists for L35025 and L38045 only. There is no automated re-validation against CMS, so the table decays silently as LCDs are revised or retired; a monitoring job is an unbuilt requirement, not a nice-to-have, before any customer relies on a citation. Code comments decline to assert 81455 into a list the author could not verify, so it matches by keyword instead. CPT recognised by regex and reasoned about but held in no table: Tier 1, Tier 2, GSP, MAAA, hereditary 81162-81167, PLA 0xxxU. ICD-10-CM is parsed and cited by the model; there is no code table and no LCD-to-covered-diagnosis crosswalk. Claimed on the site but not implemented anywhere: MolDX/DEX Z-codes (no registry, no mapping, no integration), X12 837P/835/277CA (no parser; the input is a text box), NCCI PTP edits and MUEs, NCD 90.2, a CARC/RARC taxonomy, and appeal-level clocks. NCCN is copyright-licensed and unlicensed here; CPIC, PharmGKB and PharmVar are freely redistributable and are the obvious first corpus to actually build. LOINC, SNOMED CT, HGVS, VCF, HL7 v2 ORU and the FHIR Genomics Reporting IG are absent, and all are prerequisites for the precision-medicine framing to be real.
Today: none. Zero integrations exist. The only interface is a public JSON endpoint and a copy-paste text box, and attachments are disabled in two independent places, so the product cannot even ingest the scanned denial letter or EOB PDF in which most denials physically arrive. That is the correct description to give an investor. What real landing looks like, ordered by value per unit of engineering. (1) Batch 835/837 over SFTP - the lowest-friction, highest-signal path. Labs and billers already drop X12 837P outbound claims and 835 remittance advice into SFTP directories daily; a parser reading 835 CLP and CAS segments joined to 837 service lines gives MolecuIQ a structured claim object instead of pasted prose and needs no vendor cooperation beyond a directory and a DUA. Add 277CA for front-end rejects, which is where missing-Z-code failures actually surface. This is entirely deterministic work and should be built first. (2) The billing system of record - XiFin, Telcor RCM, Quadax, Lifepoint, Health Systems Concepts - where the denial worklist already lives; realistically Ardia writes a draft back as a note against the claim rather than replacing the worklist. XiFin is not a distribution partner here, it is the incumbent. (3) Clearinghouses - Availity, Waystar, Optum/Change - as transport for the same X12. (4) LIS via HL7 v2.5.1 ORM/ORU or FHIR R4 ServiceRequest, DiagnosticReport and Observation with the Genomics Reporting IG; this is what the precision-medicine framing requires and it is far heavier than the RCM path. (5) EHR pre-order: a CDS Hooks order-select or order-sign hook that warns, before the specimen is drawn, that the linked ICD-10 will not satisfy the governing LCD. Preventing a denial beats recovering one. (6) Prerequisite for all of it: a BAA-covered US-region deployment, customer-scoped audit logging, SSO, role-based access, and a pinned model id.
Blunt version. Only two things across the whole platform are measured, and both are company-reported rather than independently reproduced: Cadence's 95.45% held-out accuracy and 0.9545 macro-F1, subject-independent, on the public UCI HAR dataset from a scikit-learn logistic regression - it is not a fall detector, it belongs to a different model, and it says nothing about MolecuIQ - and Meridian's unit-tested CLFS arithmetic, whose '34/34 tests passing' figure is likewise company-reported. Test counts generally are company-reported and unverified: the site publishes 34/34, a reviewer run reported 75 across nine files, and this review could not reproduce either from available code. Do not use the discrepancy to claim the published figure is stale until one run is reproducible on a named commit. What survives regardless: whatever the count, those tests cover the deterministic scaffolding - the regex de-identifier, the LCD matcher, the gate functions, the upload egress controls - and not one line of MolecuIQ's reasoning quality. Zero tests assert that a denial classification is correct. LIVE DEMO, probed 2026-09-01: MolecuIQ answers molecular-denial questions in production and returns real CMS LCD citations with working cms.gov links, independently confirmed for L35025 and L38045. On one probe it refused to appeal an unsupportable benign-nodule claim; 'demonstrably' is too strong for n=1. Bound the retrieval claim honestly: it grounds answers inside the eight-LCD table and retrieved PubMed literature; outside those, MolecuIQ generates from parametric memory. MODELLED TARGET: Z-code mapping, pre-submission scrubbing, 837/835 ingestion, NCCN/CPIC-grounded interpretation, appeal packet generation, outcome tracking - and also Crucible and Sentinel, both of which run in production at that maturity, meaning every safety claim here rests on two unvalidated components. HONEST ZEROS: 0 customers, 0 pilots, 0 signed BAAs or DUAs, $0 revenue, $0 raised, no real patient data, no clinical outcomes, no published accuracy, no third-party audit.
The concrete plan to produce MolecuIQ's first non-zero number. GOLD SET: 1,000 adjudicated molecular denials drawn from a Limited Data Set under a DUA with one lab or lab-billing company - paired 837P and 835 files, at least three payers, at least two MAC jurisdictions so Palmetto bias is testable, with a 250-claim subset carrying the actual appeal outcome (overturned, partially overturned, upheld) from the remit or redetermination letter. LABELLERS: two independent labellers - a CPC/CPMA-credentialed molecular coder and the counterparty's appeals lead - labelling each claim appealable or not-appealable on the documentation, with a third adjudicator (an external molecular pathologist or Dr. Mamillapalli) resolving disagreements, and Cohen's kappa reported. The founder must not label; a founder grading their own eval is disqualifying. DENOMINATOR: every denial in the sample carrying an in-scope CARC (CO-50, CO-16, CO-97, N115), not only the ones the model handles - excluding hard cases is the standard way this number gets inflated. COMPARATORS, three: the lab's own historical triage decision; a trivial rule ('appeal every CO-50'); and the identical Gemini prompt with the LCD grounding block removed, which isolates whether the eight-LCD corpus contributes anything at all. PRE-REGISTERED PRIMARY METRIC, fixed before unblinding the held-out half: precision on the appealable class at a fixed recall of 0.80. Secondary, and measurable today on synthetic claims with no DUA: citation validity - the fraction of LCD identifiers in the output that are in the retrieved set, govern that claim's MAC jurisdiction, and were effective on the date of service. KILL CRITERION: if precision at 0.80 recall fails to beat the 'appeal every CO-50' rule by at least 10 points, or the no-retrieval baseline matches the grounded system within confidence intervals, or citation validity falls below 98%, MolecuIQ does not work as conceived and should be reduced to a citation-lookup utility.
Reproducible by a skeptic in under ten minutes. (1) ENGINE IDENTITY: curl -s https://www.ardiahealthlabs.com/api/run returns {"ok":true,"provider":"gemini","gated":false}; a fast-tier POST returns model_id 'gemini-flash-lite-latest' in about three seconds, scholar tier 'gemini-flash-latest' in about fifty-four. That alone falsifies any 'Claude (Opus/Sonnet)' engine label. (2) LIVE MOLECULAR DENIAL: POST {"model":"molec","text":"Denial: CPT 81445 NGS solid tumor panel denied CO-50 not medically necessary, Palmetto MolDX jurisdiction, documented dx C34.11 lung adenocarcinoma. Is this appealable?","ground":true}. Returns six of six Crucible gates passed and sources L38045 and L35025 with working cms.gov links. (3) THE REFUSAL: post the CPT 81455 comprehensive tumour panel on a 4mm benign pulmonary nodule with ICD-10 R91.1 and ask for a strong appeal letter. The engine refuses, cites L38045 and L35025, and suggests checking whether the test was ordered on the wrong specimen. That is n=1 and self-graded; it is a demo, not a measurement. (4) THE DE-IDENTIFICATION GAP: post text containing a name, MRN, SSN, DOB and phone. The sentinel block returns removed:4 with categories date, mrn, phone_or_fax, ssn - 'name' conspicuously absent, and the name reached Google. (5) THE Z-CODE ERROR: 'Z-code not on file' in a PGx denial produces a root-cause analysis about an ICD-10 encounter code; ask the definition cleanly and it answers correctly. Both reproduce. (6) THE ROUTING: POST model:"pulmo" and the API returns {"error":"bad_model"} - the UI works only because it rewrites pulmo to tara, and toxiq and the precision-medicine listing both resolve to the molec prompt. (7) UPLOADS: attach anything and the API returns {"error":"uploads_disabled"}. (8) WHAT CANNOT BE VERIFIED BECAUSE IT DOES NOT EXIST: any customer, pilot, signed BAA or DUA, revenue, funding, real claim processed, denial classified against a known outcome, appeal submitted, overturn recorded, or published accuracy figure.
Read this section first. Stripped of framing, MolecuIQ today is a system prompt with a label on it, wrapped in regex scaffolding and pointed at Gemini. (1) The Z-code claim is not real. The site says MolecuIQ maps orders to the right MolDX Z-code; DEX Z-codes are lab- and assay-specific identifiers from Palmetto's registry, no public mapping exists, none exists in the repository, and there is no DEX integration. The clearest overclaim on the site sits on the code set most central to the product. (2) Verified reasoning failure on that same code set: on a live probe the phrase 'Z-code not on file' was read as an ICD-10 Z-code, and the entire root-cause analysis and appeal recommendation were built on that misreading, confidently and well-formatted. Asked the definition in a clean context, the same engine answers correctly. Context-dependent semantic collapse on the domain's key term is not fixable by prompt tuning alone. (3) The policy corpus is eight hand-curated LCDs frozen on 2026-08-07, with no revision tracking, no effective-date logic and no MAC jurisdiction routing, so a Novitas or NGS-MAC lab receives Palmetto citations that do not govern its claims. (4) No NCCN or CPIC corpus exists; guideline claims are unbound generation. (5) Sentinel misses plain patient names; 'John Smith' reached Google. (6) No BAA is possible on the current stack. (7) The model is not pinned, so identical denials can produce different letters and different verdicts with no version record - disqualifying for an audited workflow. (8) No claim data model: a 6,000-character text box, no 837 or 835 parser, no CARC/RARC taxonomy, no NCCI edits, no timely-filing clock, no appeal-level state machine, and no attachment ingestion at all. (9) It produces prose, not a submittable packet. (10) The non_diagnostic gate is deliberately relaxed for this model, and the precision-medicine listing is this model. (11) No post-hoc check that every L-number in the prose is in the retrieved set. (12) The evidence base is empty.
The RCM function has a strong statutory argument for sitting outside FDA jurisdiction: administrative support of a health care facility - billing, claims processing, coverage-determination support - is expressly outside the device definition under 21st Century Cures Act 3060 / FD&C Act 520(o)(1)(A), independently of the CDS carve-out. That is Ardia's own reading of the statute. No outside regulatory counsel opinion has been obtained, no Q-Submission filed, and no FDA feedback sought; obtaining a written opinion is an open work item, not a completed one. The precision-medicine framing is what drags toward device territory: software ingesting variant data and outputting therapy-relevant recommendations must clear all four 520(o)(1)(E) CDS criteria and can fail on criterion 1 (analysing a pattern from an in-vitro diagnostic) and criterion 3 (a directive rather than a weighable recommendation). The internal rule should be explicit: MolecuIQ answers reimbursement questions and summarises published guidance with citation; it does not select therapy. CLIA is inapplicable - Ardia holds no certificate, examines no specimen and issues no result. LDT context for customer conversations only: FDA's May 2024 LDT final rule was vacated by the Eastern District of Texas in March 2025, so MolDX technical assessment and Z-code registration remain the practical gatekeeper. HIPAA: production calls Google AI Studio, which is not BAA-covered; Google's BAA covers Vertex AI. Until the engine moves under an executed BAA in a US region, PHI is legally impossible here. Ardia's HIPAA control matrix is self-graded 2 of 15, with no SOC 2, no HITRUST and no penetration test. State law binding Ardia today: Texas SB 1188 US data residency and TRAIGA AI-governance obligations, which bear directly on the three egress paths described in dataFlow. No data-residency assessment, TRAIGA governance record or vendor data-processing review has been documented for any of them. This is unaddressed, not addressed-and-cleared.
'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 pathway. MolecuIQ drafts a redetermination citing an LCD. A human - a biller, then a Medical Director or pathologist - signs it, and a Medicare appeal carries an attestation that the statements in it are true and complete. If the model cites an LCD that was revised or retired since 2026-08-07, cites a Palmetto MolDX policy at a lab adjudicated by Novitas, or asserts a coverage criterion the record does not support, that false statement is now in a submission to a federal health care program over a signature. All six Crucible gates can pass while this happens: no gate checks whether a citation governs this claim, and none can, because the system holds no jurisdiction routing and no effective-date logic. Exposure sits with the customer first - 18 U.S.C. 1035 for false statements, and False Claims Act liability where a pattern of unsupported appeals looks like reverse false claims - and urine drug testing, which reaches this same prompt through ToxIQ, is a named OIG enforcement priority, so the shared engine path imports toxicology enforcement risk into a molecular product. Payers have begun flagging AI-drafted appeals, and a lab that cannot show human review of every assertion invites pattern scrutiny across its whole appeal book. Practically, a lab compliance officer running a seven-element program cannot permit an unvalidated tool into the appeal workflow: element three requires effective training, element five auditing and monitoring, and there is nothing to audit against. Ardia's own exposure runs through contract, and Ardia has never written one - no indemnification clause, no limitation of liability, no allocation of responsibility for a bad citation, no professional liability or tech E&O cover, and no BAA. That contract, not the FDA analysis, is the document that decides who holds the bag.
The cheapest credible validation path avoids PHI entirely, and Ardia should sequence it that way. TIER 0, no agreement needed, available today: CPIC allele-definition and allele-functionality tables, PharmVar star-allele nomenclature, and the CDC GeT-RM pharmacogenetic reference materials - roughly 137 well-characterised DNA samples with consensus genotypes across CYP2D6, CYP2C19, CYP2C9, VKORC1, TPMT and DPYD - are free, public, redistributable, and contain no human subjects. A deterministic star-allele to diplotype to phenotype to CPIC-recommendation mapper validated against that panel would produce MolecuIQ's first real accuracy number at zero cost and with zero permissions. It has not been done. TIER 1, the actual unlock: a Limited Data Set under a Data Use Agreement per 45 CFR 164.514(e) with one molecular lab or, better, one lab-billing company. An LDS may retain service dates and ZIP, which matters because timely-filing clocks and MAC jurisdiction are the whole game, and it needs a DUA rather than a BAA. Retrospective 837P claim files with paired 835 remits. Defensible volume: 3,000-5,000 adjudicated molecular claims, of which at least 600 denied, spanning at least three payers and at least two MAC jurisdictions so the Palmetto-MolDX bias gets tested against Novitas or NGS-MAC, plus a labelled subset of at least 200 appealed claims with known outcomes. TIER 2, full BAA and prospective live claims, only after the engine sits on Vertex under a Google BAA and a compliance officer has reviewed the egress path including the PubMed keyword call. IRB: retrospective claims analysis for a covered entity's own operations is not human-subjects research; a commercial IRB not-human-subjects determination costs low four figures if Ardia wants to publish. The binding constraint: because Sentinel cannot redact patient names, any file Ardia receives must arrive already de-identified by the counterparty, and the DUA must say so rather than implying Ardia's de-identifier suffices.
Sizing, with arithmetic shown and every input labelled as an estimate. Universe: roughly 320,000 CLIA-certified laboratories exist, but the addressable set - labs performing high-complexity molecular and genomic testing - numbers in the low thousands, and the population registered with Palmetto's DEX Diagnostics Exchange is on the order of several hundred to about a thousand labs. Call the serviceable set 1,500 molecular and genomic labs plus perhaps 40-60 lab-billing companies that aggregate them; that figure is an estimate, not a count Ardia has performed. Spend: Medicare's Clinical Laboratory Fee Schedule pays roughly $8-9B annually, with molecular pathology and genomic sequencing the fastest-growing component; all-payer US molecular diagnostics is commonly put in the $25-30B range. Each of those figures needs a named source and year before it appears in a deck. Applying the standard denial-management heuristic of 2-4% of touched revenue gives a nominal $500M-$1.2B ceiling, most of which Ardia will never see because incumbents own the billing system of record. SAM, the labs that would buy a bolt-on: 1,500 labs at $60k-$150k ACV is $90M-$225M. SOM for a Dallas-Fort Worth seed company with one engineer over three years: 15-40 labs, $2M-$5M ARR, and only if a first reference customer produces a published overturn-lift number. Worked buyer volume for a mid-size molecular lab, used consistently throughout this dossier: 6,000 molecular claims per month at a blended $600 allowable is $3.6M billed monthly and $43.2M annually; at a 22% first-pass denial rate that is 1,320 denials per month; assume 40% are genuinely appealable on documentation (528) and the team can work 45% of those (238), leaving roughly 290 appealable claims per month abandoned for lack of hands - $174,000 monthly, $2.09M annually. Every input in that chain is an assumption Ardia has never measured.
One model, chosen for a company with no proof: per-drafted-appeal transaction pricing at $20 per appeal draft, with a $2,000 monthly platform minimum. It benchmarks against the commonly quoted $25-$118 industry rework cost, it scales with delivered work rather than seats, and unlike contingency it needs no outcome-tracking infrastructure - which does not exist - and no revenue recognition against overturns Ardia cannot observe. Buyer arithmetic, using this dossier's mid-size lab: 1,320 denials per month, roughly 528 appealable, roughly 238 worked today, leaving about 290 abandoned for lack of hands. Pricing those 290 at $20 is $5,800 per month, $69,600 per year. Those same claims at $600 blended allowable represent $2.09M of annually abandoned revenue, so the break-even overturn rate on newly worked claims is $69,600 / $2,088,000 = 3.3%. That is the honest ROI statement: the buyer needs only a 3.3% overturn rate on claims currently written off to break even, and Ardia does not need to assert a lift it has never measured. The customer's own historical overturn rate supplies the upside case. COGS: Gemini Flash Lite list pricing is roughly $0.10 per million input tokens and $0.40 per million output tokens (verify current list before quoting). A MolecuIQ call runs roughly 5,000-7,000 input tokens including the grounding block and about 1,200 output tokens, so about $0.001 per call; at four calls per appeal and 290 appeals, inference costs about $1.20 per month against $5,800 of revenue. Inference is not the cost driver. Loaded gross margin after hosting, Vertex, audit logging and support lands realistically at 80-88%, and the true costs are LCD corpus maintenance and compliance, not tokens. Caveat that must travel with all of it: this price has never been quoted to anyone, and zero customers have tested it.
Named, real, and the honest answer is that MolecuIQ is not currently differentiated against any of them. Lab-specific RCM incumbents matter most: XiFin is the dominant molecular and anatomic-pathology lab RCM platform, already encodes MolDX rules and DEX Z-code workflow inside the billing system of record, already produces appeals, and already holds the 837/835 data Ardia is asking for a DUA to see; Telcor RCM, Quadax, Lifepoint Informatics, Health Systems Concepts, Orchard and Advanced Data Systems occupy the rest. This is the hardest fact in the dossier: the incumbent sits inside the system of record, owns the data and the customer relationship, and can bolt an LLM onto a workflow it already owns faster than Ardia can build a workflow around an LLM it already has. General denial AI: Waystar (which acquired Iodine Software in 2025), Optum/Change, Experian Health AI Advantage, FinThrive, R1, Ensemble, Availity, Rivet, Adonis, Janus, Anomaly, and Cohere Health on prior authorisation - several already ship LLM-drafted appeals at national scale with real overturn data Ardia does not have. MolDX and market-access specialists: Palmetto's own DEX exchange, plus Diaceutics, Boston Healthcare, Health Advances and Avalere, who sell the coverage-strategy expertise MolecuIQ automates, to the same labs. Precision-medicine interpretation, if that framing is pursued: Tempus, Foundation Medicine, Qiagen QCI Interpret, SOPHiA GENETICS, PierianDx/Velsera, Fabric, Congenica, Genomenon, and for PGx specifically Translational Software, OneOme, Coriell and GenXys - all with licensed content, curated variant knowledge bases and EHR integrations Ardia lacks; that assessment is from public positioning, not hands-on evaluation. Honest differentiation, and only three items, none defensible for long: refusal as a first-class output, LCD citation with working links in a category where fabricated citations are the default failure, and the PAMA cliff as an opener. The obvious objection stands unanswered: nothing stops XiFin or Waystar building an eight-LCD matcher and a Gemini prompt in a fortnight.
Sequenced, cheap deterministic work first, sized against one engineer's capacity - The founder is the sole engineer, which is the binding constraint on every item below. P0, 0-60 days, no money, no customer, no permission required. Add an output-side citation validator that flags any LCD or NCD identifier in the prose that is not in the retrieved source set and does not govern the claim's MAC jurisdiction and date of service; this is the highest severity-to-effort ratio in the product. Pin the Gemini model id and record it on every response instead of resolving a moving '-latest' alias. Ship free-text name detection in Sentinel, or at minimum have the de_identification gate report 'not assessed' rather than 'passed' when names cannot be checked - a green gate over a leaked name is a control failure, not a gap. Build the deterministic CPIC/PharmVar star-allele to diplotype to phenotype to recommendation mapper and validate it against the CDC GeT-RM panel: public data, no BAA, no DUA, no IRB, and it yields MolecuIQ's first measured accuracy number. Build an 835 parser and a CARC/RARC to root-cause taxonomy as pure unit-tested code. Stand up a regression eval harness across the shared molec prompt, because one prompt change currently degrades MolecuIQ, ToxIQ and the precision-medicine listing simultaneously with nothing to detect it. Correct the /models page to say Gemini and drop the Z-code and EDI claims until built. P1, 60-180 days: sign one Limited Data Set DUA; migrate to Vertex AI under a Google Cloud BAA; add LCD revision diffing and MAC-jurisdiction routing; add appeal-deadline tracking. P2, 180-360 days: appeal-packet generation, outcome capture, a CDS Hooks pre-order check, SOC 2 Type I. The single unlock is the DUA - a lower bar than a BAA, signable by a billing company's VP of Operations - but the entire P0 list must land before that conversation, because a compliance officer asking whether the de-identifier catches names currently gets a no.
