A deterministic, unit-tested compounding model that applies a user-chosen annual Clinical Lab Fee Schedule cut across 2027-2029 to a lab's own test mix to show cumulative revenue at risk — genuinely arithmetic, not yet a CLFS rate engine, and not wired into the product a buyer actually clicks on.
Pillar 3 — the 2027 PAMA rate cliff — and the financial-defence layer under Pillar 2 (molecular and genomic diagnostics, including the kidney SKU and the 81xxx/NGS codes whose rate durability drives the menu-diversification argument) and Pillar 1 (precision medicine as the strategic answer to toxicology-code compression; toxicology is on the roster and is the segment most exposed to the cliff). Meridian does not address Pillar 4 (pulmonary and respiratory care) or Pillar 5 (elder care) in content — but it is bound to Pillar 4 mechanically, because the Studio's Meridian and PulmoIQ execute the identical engine path.
The single word overstates it unless read with this caveat, so read it. MEASURED applies to exactly one narrow surface: the deterministic compounding arithmetic in models/meridian/clfs.py and its unit tests. That result is COMPANY-REPORTED and was not independently executed in this review; the platform-wide "34/34 tests passing" figure is likewise company-reported. Unit tests establish internal consistency of a formula, not agreement with any published CMS payment amount. Everything a buyer would actually touch is weaker. The Studio's Meridian is LIVE DEMO — Google Gemini under TARA's general clinical-reasoning prompt, non-deterministic, no published accuracy, with the deterministic module not in the call path. Every quantified figure on the /pama marketing page — per-code 2027 rate targets, denial rates, cost-to-collect reduction, the monthly net-swing scenario — is MODELLED TARGET, hand-entered, with no Ardia code producing it. The honest zeros stand without softening: 0 customers, 0 pilots, 0 signed BAAs or DUAs, $0 revenue, $0 raised, no real patient data ever processed, no clinical outcomes, and no lab general ledger ever reconciled to a Meridian output. Founded December 2025, Dallas-Fort Worth.
Yes, and it must be stated in any Meridian material rather than discovered. Ten named Ardia models collapse to four engine paths. Meridian and PulmoIQ are literally the same path (TARA); ToxIQ is literally MolecuIQ's. Stated plainly: Meridian is a product framing and a UI label over a shared general engine. There is no Meridian-specific model weight, no fine-tune, no separately evaluated PAMA persona, and this review could not confirm that any Meridian-specific system prompt exists in the server — the company should be asked to produce the prompt text or confirm its absence in writing. Three consequences follow. First, a Studio Meridian session answers PAMA rate questions by general LLM reasoning, freehand, without touching clfs.py. Second, one prompt regression or one silent roll of the unpinned "-latest" alias degrades three products simultaneously, with no evaluation harness anywhere in the company to detect it — the strongest technical risk Ardia carries. Third, Meridian inherits TARA's retrieval profile, including the fact that pulmonary grounding comes from PubMed literature retrieval rather than a curated GOLD/GINA guideline corpus that has not been built. A diligence reviewer who reads models.html's "deterministic Python engine — no LLM" label next to studio.html's routing map will find this contradiction in about ten minutes, and it recolours every other claim on the site.
PAMA section 216, codified at Social Security Act 1834A, replaced the 1984 charge-based Clinical Laboratory Fee Schedule with a market-rate method: CMS sets each CLFS amount at the volume-weighted median of private-payor rates reported by 'applicable laboratories.' Public commentary on the 2016 collection describes very low reporting participation concentrated in the largest national labs, whose national-scale contracted rates then became the Medicare baseline for every small independent lab; the 2018-2020 phase-in reduced payment for a large majority of CLFS codes under caps set by 1834A(b)(3). Each of those characterisations needs a named citation before it appears in a diligence document — none is sourced in Ardia's materials today. Congress has then repeatedly delayed the next phase through the LAB Act, the CARES Act, and successive appropriations acts and continuing resolutions; the correct way to state this is a citable list of statutes, not a round count of delays. The concrete pain is planning, not billing: a CFO cannot answer 'what is my 2029 Medicare run-rate?' because the answer depends on a per-code weighted median she has never seen, a reporting window whose date has moved repeatedly, and a phase-in cap that may not still be law when it bites — while she is being asked to sign a three-year payer contract, a lease, or a sale process against that line. Definitive drug testing G0480-G0483 is understood to dominate toxicology-lab Medicare revenue, and molecular 81xxx codes to carry elevated first-pass denial rates; both are industry-estimated, unsourced, and not measured by Ardia against any customer data. The compounding is what CFOs get wrong in their heads: three consecutive 15% cuts is 38.6%, not 45%, and PAMA also eliminated the CPI-U update, so a zero-cut year still erodes in real terms. Meridian's intended scope is that arithmetic and only that arithmetic.
HYPOTHESISED, not validated. Ardia believes the economic buyer is the CFO or owner-operator of an independent clinical laboratory in roughly the $5M-$80M net-revenue band, with the cheque signer shifting from owner to CFO somewhere near $30M. No part of that has been tested: 0 customers, 0 pilots, and no documented buyer-discovery interviews. The revenue bands, headcount ranges and signing threshold are assumptions to falsify in the first ten conversations, not findings. Structurally, the org chart in labs of that size is short — owner/CEO, CFO, VP or Director of Revenue Cycle, Billing Manager, analysts — and the CLIA Laboratory Director under 42 CFR 493.1443 sits on the operational side and is not the buyer. The end user is not the CFO either: it is an FP&A analyst or revenue-cycle director building a budget deck, plus the payer-contracting manager preparing renegotiation talking points. Sponsor ownership is common in independent toxicology and molecular labs, though Ardia has not sized this and cites no source; where it applies, the likely influencer is the sponsor's operating partner needing a defensible EBITDA bridge through 2029. Treat that segment as unquantified until a named data source is produced. A second buyer Ardia does not currently name may be better qualified: transaction diligence — quality-of-earnings teams, lab-focused bankers, PE deal groups who need a PAMA exposure model on a target inside a two-week window, pay project fees rather than software fees, and do not benchmark the deliverable against Excel. A third is the hospital outreach program billing referred specimens on TOB 14x. The uncomfortable truth about all three: none has a budget line called 'PAMA modelling software,' and each has a controller who can write the same formula in a spreadsheet.
Meridian sits entirely on the administrative side of the lab. The calculator never sees a result, never sees a patient, never sees a specimen — it consumes a code-level test mix with counts and a rate assumption. The clinically relevant surface is which code families carry the exposure: definitive drug testing G0480-G0483 and presumptive screening 80305-80307 in toxicology; 81225, 81226, 81401, 81445 and 81455 plus 87491-87591 in molecular and infectious disease; and high-volume chemistry such as 82306 (25-hydroxy vitamin D) and 82607 (vitamin B12), where per-test contribution margin is widely assumed to be thin. Ardia has no cost data from any lab and cannot presently say at what cut level any specific test flips contribution-negative; that claim would require a design partner's cost-per-test and is not modelled. Coverage for these families runs through MAC Local Coverage Determinations for controlled-substance monitoring and urine drug testing, ICD-10-CM medical-necessity linkage against LCD covered-diagnosis lists, and NCCI Medically Unlikely Edits as a frequency-denial driver. A citation note that matters in a product whose whole pitch is citation discipline: the only LCD retrievals independently verified on this platform are L35025 and L38045, both real with working cms.gov links. Any other LCD identifiers quoted in Meridian materials are company-reported probe outputs and are not re-verified here. Meridian-in-Studio is a different object from the calculator and must be described differently: it is an unconstrained free-text box on a general clinical-reasoning engine, so a user can paste clinical narrative into it, and Sentinel does not reliably strip plain personal names from what is pasted. Neither artefact is clinical decision support, neither is a device, and neither may be presented as diagnostic — but for the Studio path that is a policy commitment enforced by gates, not a structural property of the input.
Two artefacts, and the difference between them is the whole story. (1) Meridian-calculator. models/meridian/clfs.py holds genuinely deterministic arithmetic: given a list of (code, rate, volume) rows and a cut fraction, it computes a baseline, compounds the cut across three years, and returns projected annual revenue, cumulative revenue at risk, and an ending run-rate as a percentage of baseline; a separate days-late penalty function multiplies out the 1834A(a)(9) civil monetary penalty. A JavaScript renderCalc() on model-pama.html is said to reproduce it. The two implementations have not been shown to agree, which is a classic source of silent divergence. Crucially, this is not a CLFS rate engine: it does not ingest CMS CLFS public use files, does not know any code's actual national payment amount, holds five hard-coded illustrative 2025 rates, and applies one uniform user-chosen cut to every code rather than the statutory per-code phase-in. It is compound-interest arithmetic on numbers the user types in. (2) Meridian-in-Studio. The verified routing map sends a Meridian selection to model 'tara', so the request runs the shared pipeline: Sentinel regex de-identification (in development, modelled-target maturity — SSN, phone, MRN, dates and ZIP redact, but plain personal names do not; 'John Smith' reached the model on probe) then PubMed / ClinicalTrials / CMS-LCD retrieval under cite-or-abstain and policy-override, which are enforced in retrieval and answer-binding and are deliberately NOT counted as gates, then Google Gemini under TARA's general prompt, then six deterministic Crucible gates — non_diagnostic, safety_escalation, scope_of_practice, de_identification, honesty, human_in_the_loop — each returning with a reason on every call, any failure withholding the answer entirely (verified in production). Crucible is also modelled-target maturity. The deterministic module is not in this path.
Two flows with almost nothing in common. The calculator: the company states the test mix is entered in the browser, computed in the browser, and never transmitted — no server receives it, nothing is logged or persisted. This review captured no network trace either way and no independent security review of model-pama.html exists; a buyer should confirm the no-egress claim with a browser network capture before relying on it in a compliance narrative. Even taken at face value, that is an absence of data flow rather than a compliance posture: Ardia's own HIPAA control matrix is self-graded 2 of 15, no control has been audited, and there are 0 signed BAAs or DUAs. The correct claim is least compliance surface, not best compliance story. Ardia's position is that a code-level test mix with counts is neither PHI under 45 CFR 164.514 nor a Limited Data Set; that position has had no outside counsel review, no privacy-officer sign-off, and no expert-determination memo, and the buyer's compliance function should confirm it rather than accept it. The Studio flow is materially different: free text goes through regex de-identification with a known name-redaction gap, then surviving text plus retrieved sources go to Google Gemini — a third party with which Ardia has no BAA, consistent with 0 signed BAAs company-wide. No PHI may lawfully traverse this path today, and the only control preventing it is regex that is known to be incomplete. The path additionally carries Texas SB 1188 US data-residency and licensed-professional-review obligations plus TRAIGA scope. One structural fact worth stating: file attachments do not work at all — the API returns {"error":"uploads_disabled"} and studio.html hardcodes attachments to empty, so a chosen file is read to base64 and discarded.
Load-bearing statute and regulation: Social Security Act 1834A, added by PAMA section 216 — subsection (b)(3) is the phase-in reduction cap the model encodes, (a)(9) is the $10,000-per-day civil monetary penalty for failure to report or misreporting, and (d) is the Advanced Diagnostic Laboratory Test carve-out Meridian does not model. 42 CFR Part 414 Subpart G (414.500-414.523) is operative: 414.502 defines 'applicable laboratory' and 'applicable information', 414.504 sets reporting requirements, 414.507 sets the payment methodology and the phase-in caps. The CMS Clinical Laboratory Fee Schedule annual and quarterly public use files are the rate source of record. Meridian holds five hard-coded illustrative 2025 payment amounts. This is not a data-feed gap on a working engine — it is the absence of the engine: no ingestion of the CLFS public use files, no effective-date or version stamping, no coverage of the remaining CLFS code set, and therefore no ability to produce a rate for any code a real lab actually bills. Code sets in scope: HCPCS Level II G0480-G0483 and G0659; CPT 80305-80307, 82306, 82607, 81225, 81226, 81401, 81445, 81455, 87491-87591; ICD-10-CM for medical-necessity linkage; NCCI MUEs for frequency exposure; LOINC for mapping an internal test menu to billable codes. Coverage instruments: MAC Local Coverage Determinations (verified retrievals on this platform: L35025, L38045) and Palmetto GBA MolDX with DEX Z-codes as a roadmap overlay, not built. Transaction standards for the integration roadmap: X12 837P loop 2400 SV1 for code and units, X12 835 ERA CLP and SVC segments for actually-allowed and paid amounts, and TOB 14x for hospital outreach. FHIR R4 is the wrong standard here and any roadmap mention should be corrected — the money is in X12. NCCN and CPIC are not Meridian standards. Texas SB 1188 and TRAIGA apply to the Gemini-backed Studio path.
Today there is none, and that is the accurate answer. Meridian-calculator is an HTML page with a table you type into. Meridian-in-Studio is a text box — and not even a text box that accepts files: the API returns {"error":"uploads_disabled"} for any attachment and studio.html hardcodes the attachment list to empty, so a lab cannot upload a test-mix export today even in principle. The realistic landing sequence, ordered by value per unit of effort: step one is CSV or XLSX upload of a code-level test mix exported from whatever billing system the lab already runs — XIFIN RPM, Telcor, Ligolab, Quadax, or a homegrown database — because a billing manager can produce that report in an afternoon and it carries no PHI. Step two is where the product would become genuinely differentiated: X12 835 ERA ingestion, whose CLP and SVC segments carry the amount actually allowed and paid per service line, letting Ardia compute a realized rate per code rather than a billed rate. That is the single most valuable input to a PAMA forecast and something a CFO usually cannot extract cleanly from her own system. Step three is 837P claims ingestion for volume and payer mix. Both arrive by SFTP from clearinghouses or the billing vendor, and both require a BAA — so both are gated on the HIPAA matrix moving off 2 of 15. LIS-side integration is largely a distraction here: the LIS knows tests, not adjudicated dollars. The least glamorous integration matters most: the output must land as an Excel file and a board-ready PDF, because the buyer's real workflow is a budget model and a lender or sponsor deck, not a dashboard she will log into.
COMPANY-REPORTED, not independently reproduced in this review. The company states that its Meridian unit-test suite passes; a clfs.py module and a small pytest file were located, but only inside archived website snapshots dated Aug 02 and Aug 03 and a folder labelled 'PREVIOUS-before-2026-08-05-changes'. This review did not execute the suite and did not confirm the module is present in the currently deployed tree. The platform-wide '34/34 tests passing' figure is likewise company-reported and unverified. Treat both as company-reported until a diligence party runs them against the production commit. The company publishes the following outputs for an illustrative seed mix it constructed — 80307 at $62.14 x 4,200; G0483 at $246.92 x 1,100; 82306 at $36.55 x 5,200; 81225 at $128.00 x 900; 82607 at $18.61 x 3,100 — baseline $895,551.00; projected $761,218.35 / $647,035.60 / $549,980.26; cumulative revenue at risk $728,418.79; ending run-rate 61.41% of baseline. No real lab's rates or volumes are involved, and this review neither executed the module nor reconciled those figures against the live site. The claimed measured surface is therefore narrow and unconfirmed: unit tests of a compounding formula establish internal consistency, not agreement with published CMS CLFS amounts. The house standard for the word 'measured' should be stated openly, because only two things in the entire company qualify and neither has been independently reproduced: this arithmetic, and Cadence at 95.45% held-out accuracy with macro-F1 0.9545, subject-independent, on the public UCI HAR dataset — which is an activity classifier and explicitly not a fall detector. Everything else in the portfolio is live demo or modelled target. No Ardia measurement has ever been reproduced by a third party.
Meridian can produce its first non-company-reported number without a customer, a BAA, an IRB, or a single labelled example — which is why not having done so at nine months old is itself a finding. The design is a retrospective backtest against a phase-in that already happened. Gold set: the CMS CLFS public use files for 2017 through 2020, which publish the pre-PAMA baseline and each realised annual national payment amount. Nobody labels anything; CMS is the labeller. Denominator: every code present in all four files, an N fixed by the data rather than chosen — on the order of a thousand-plus codes — with results reported both unweighted and weighted by 2017 Medicare allowed charges, because dollars decide budgets, not code counts. System under test: Meridian's current model, a uniform statutory-cap cut applied to every code. Comparators: a null model that projects no change, and a per-code oracle that uses each code's own realised trajectory, which bounds how much accuracy per-code modelling could ever buy. Pre-registered primary metric: dollar-weighted median absolute error in projected three-year run-rate as a percentage of baseline. Secondary: the share of a lab's projected dollars landing within plus or minus five percentage points. Kill criterion, declared before running: if the uniform-cut model's dollar-weighted median absolute error on the 2018-2020 backtest exceeds ten percentage points of run-rate, or fails to beat the null by a clear margin, the single-slider product does not work as a forecasting tool and must either become per-code or be withdrawn and sold honestly as a scenario calculator. A second gate applies at design-partner stage: if the baseline year cannot be reconciled to a lab's own general ledger within three percent, the model is not usable for budgeting regardless of its projection accuracy.
What a sceptic can check right now, stated as checks rather than as results this review obtained. Engine: GET https://www.ardiahealthlabs.com/api/run returns {"ok":true,"provider":"gemini","gated":false} — verified. A fast-tier POST returns model_id gemini-flash-lite-latest in about three seconds; scholar tier returns gemini-flash-latest in about fifty-four. It is Gemini, not Claude, and the id is an unpinned alias. Routing: read ENGINE_MODEL in studio.html and confirm the ten-to-four collapse, with meridian mapped to tara and pulmo mapped to tara. Confirm the API rejects the label directly — posting model:"pulmo" returns {"error":"bad_model"}, verified — so the UI works only because it rewrites first. Gates: any completed call returns all six gates with a reason, and a failed gate withholds the answer entirely, enforced in code and verified in production. Retrieval: a molecular query returns real CMS Local Coverage Determinations with working cms.gov links, L35025 and L38045 among them; a pulmonary GOLD/COPD query returns real PubMed citations, PMID 40050074 and PMID 38032494 — the honest gap is that this is PubMed literature retrieval, not a curated GOLD/GINA guideline corpus, which is not built. De-identification: send a plain personal name and observe that it reaches the model, while SSN, phone, MRN, dates and ZIP redact. Attachments: send any file and observe {"error":"uploads_disabled"}. Meridian specifically: read clfs.py — it fits on one screen and there is no hidden logic — then run its unit tests yourself, because the company's passing result has not been independently reproduced. What cannot be checked because it does not exist: a customer, a pilot, a signed BAA or DUA, a dollar of revenue, a dollar raised, or a general ledger a Meridian output has ever reconciled against.
Read this before anything else. (1) Meridian in the Studio is a prompt with a label. Selecting it routes to TARA on Gemini; there is no Meridian persona confirmed anywhere in the server, no fine-tune, no evaluated PAMA prompt, and the deterministic module is not in the call path. Meanwhile models.html labels Meridian a deterministic Python engine with no LLM. Both statements are individually defensible and jointly misleading. (2) The deterministic module is compound interest. It holds no fee schedule, no code table, no geography or MAC adjustment, no MolDX overlay — five hard-coded rates and one exponent. (3) It applies one uniform cut to every code. Real PAMA rates are set per code from weighted medians: some fall to the cap, some barely move, some rise with no statutory cap on increases. A single-slider model is structurally incapable of the per-code forecast that is the only forecast a CFO can act on. (4) ADLTs and newly priced codes are exempt from the phase-in cap under 1834A(d); a molecular lab with an ADLT gets a materially wrong answer with no warning. (5) The cut years and rates are module constants with no effective date, version, or citation, and unit tests that pin a frozen total would, after any CMS revision, actively enforce a stale number — a suite defending the wrong invariant. (6) The eliminated CPI-U update is not modelled. (7) The engine is unpinned: '-latest' aliases can roll silently, and there is no eval harness anywhere to detect a regression across the three products sharing this path. (8) Sentinel does not reliably redact plain personal names. (9) No moat: no proprietary data, no network effect, no switching cost. (10) Zero design partners means no output has ever been reconciled to a real general ledger.
Ardia's position — not a legal opinion, and not reviewed by regulatory or privacy counsel — is that Meridian-calculator falls outside FDA device jurisdiction under 21 U.S.C. 321(h) because it makes no claim about diagnosis, cure, mitigation, treatment or prevention of disease; outside CLIA and 42 CFR Part 493 because it examines no material derived from the human body; and outside HIPAA because it processes no PHI. No 513(g), no outside-counsel memo, and no privacy assessment supports these conclusions today, and a buyer should obtain its own read rather than rely on these characterisations. The company-wide posture is administrative, non-diagnostic, non-device decision-support software, always, with 0 signed BAAs or DUAs and a HIPAA control matrix self-graded 2 of 15 — nowhere near ready to receive a claims file. Aria, elsewhere in the portfolio, escalates suspected emergencies to 911; Meridian has no such surface. State law bites on the Studio path specifically: Texas SB 1188 imposes US data-residency obligations and requires licensed-professional review of AI output, and TRAIGA applies to AI systems used in healthcare and offers a 36-month regulatory sandbox. Meridian-in-Studio, running on Gemini, sits inside both. Two non-FDA exposures deserve naming. First, if Meridian's output shapes what a lab reports to CMS during the private-payor data-reporting window, the /pama page's language about submitting 'the most favorable defensible rates' sits close to 42 CFR 414.504, which requires reporting all applicable information rather than a selected subset, and to the 1834A(a)(9) penalty. That is copy risk, fixable with counsel review, and it has not been reviewed. Second, contingency pricing on Medicare recoveries invites percentage-of-collections scrutiny; that belongs to MolecuIQ and ToxIQ but appears on the same page.
'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. Meridian's six gates test posture — non-diagnostic framing, scope of practice, safety escalation, de-identification, honesty, human-in-the-loop — not arithmetic truth or citation applicability. A confidently wrong rate projection, or a citation that is real but does not support the sentence it is attached to, passes all six and is withheld by none. Downstream, the person holding the bag is the human who signs. If a Meridian-adjacent output shapes what a lab reports in the private-payor data-reporting window, 42 CFR 414.504 requires all applicable information rather than a favourable subset, 1834A(a)(9) carries a $10,000-per-day penalty, and knowingly retaining an overpayment is a reverse false claim under 31 U.S.C. 3729(a)(1)(G). If the shared TARA path emits coverage or appeal language that a biller pastes into a CMS redetermination, the signer attests the submission is true and complete — and urine drug testing is a named OIG enforcement priority, so the toxicology segment Meridian targets is exactly where scrutiny is highest. Payers are beginning to flag AI-drafted appeals, which converts an efficiency story into an audit trigger. A compliance officer running a seven-element program must document validation of any tool entering the appeal workflow; an unvalidated model served through an unpinned '-latest' alias that can roll silently cannot be documented, which is a defensible reason to refuse it entry. And the contractual layer does not exist: Ardia has written no MSA, has no limitation of liability, no indemnity language, no tech E&O coverage on record, and no balance sheet to stand behind the uncapped regulatory-penalty indemnity a first customer's counsel will ask for.
The data need to move Meridian from correct arithmetic to validated forecast is unusually small, which is why its absence is a go-to-market failure rather than a data-access problem. Two independent inputs are required. The first is free and public: the CMS CLFS annual and quarterly public use files, ingested with effective dates and version stamps to replace the five hard-coded floats. No agreement, no counterparty, no PHI. That this has not been done is the clearest available signal of how early the product is. The second is one design-partner lab's trailing twelve months of Medicare Part B activity, aggregated by the lab before transmission to code, billed units, allowed amount and paid amount. Ardia's position is that such a code-level aggregate is de-identified on its face and would need only a commercial data use agreement — no BAA, no Limited Data Set agreement, no IRB review under 45 CFR 46. That has not been confirmed by counsel or any IRB, and no design-partner compliance officer has reviewed it. With 0 signed BAAs and 0 signed DUAs, the path has never been walked once; expect the partner's own privacy review, not Ardia's reading, to set the terms. A proposed starting set is 3-5 labs spanning toxicology, molecular and general chemistry. Ardia has not defined what would constitute a pass — no reconciliation tolerance, no acceptance criteria, no argument for why that number of labs supports any generalisation. Until those are written down it is a data-collection plan, not a validation plan, and zero design partners are engaged. Only if 837P and 835 ingestion ships does the requirement escalate to a full BAA, a security risk analysis under 45 CFR 164.308(a)(1)(ii)(A), encryption in transit and at rest, and a minimum-necessary review — which is why closing the 2-of-15 matrix comes first.
MODELLED, from unsourced inputs, and it should be read as a thought experiment rather than a sizing. Medicare CLFS spending is commonly cited near $8B/year [source needed]; CLIA-certified independent laboratories number in the several thousands [source needed]; Ardia assumes 1,500-3,000 of them have material CLFS exposure, concentrated in toxicology, molecular and reference chemistry [assumption, unvalidated]. Applying an untested price band of $12K-$60K per lab per year — no lab has ever been quoted, and Ardia has $0 revenue and 0 customers — gives an illustrative ceiling of $24M-$120M, of which only a fraction is winnable against Excel, free ACLA member analyses, and free CMS files. Every term is an assumption. The honest conclusion an investor should hear is that Meridian is not a business; it is a wedge whose job is to make an urgent, dated, quantified problem legible to a CFO and hand off to the products where dollars change hands. That handoff does not exist yet: ToxIQ runs MolecuIQ's engine path, neither has published accuracy on denial prediction or recovery, no recovery has ever been achieved, and the roughly 15%-of-recovered-revenue price is a proposal no customer has accepted. The company's own note that contingency pricing on Medicare recoveries invites percentage-of-collections scrutiny applies before that is quoted to anyone. The value hypothesis to test — not a return to underwrite — is that the output changes one decision worth more than the fee: repricing a commercial contract before the Medicare reference resets, shifting test menu toward durably reimbursed codes, or pricing a sale or lender covenant correctly. Ardia has never produced an output that informed a real negotiation and has no evidence that a PAMA forecast changes contract outcomes.
One model, chosen and defended: a fixed-fee PAMA Exposure Assessment at $25,000 per engagement, with an annual re-run at $10,000 once the CLFS ingestion and backtest exist. Fixed fee rather than SaaS because the buyer has no line item for planning software and does have a line item for professional services; fixed fee rather than contingency because contingency on Medicare-related dollars invites percentage-of-collections scrutiny and, worse, gives Ardia a financial incentive to inflate the exposure number it is being paid to compute. That conflict has to be stated to the buyer, and the fee has to survive the answer 'your exposure is small.' Cost side: inference is noise. A scholar-tier Gemini call at roughly eight thousand input and two thousand output tokens costs well under a cent at flash-tier list pricing; even a thousand calls across an engagement is under ten dollars. Gross margin is governed entirely by human hours. At roughly forty hours of founder and analyst time per engagement at a fully loaded $150 per hour, direct cost is about $6,000 — a 76% gross margin, but a services margin, and one engineer supports perhaps two engagements a month, which caps this at a few hundred thousand dollars a year and does not compound. Buyer arithmetic, labelled as scenario rather than forecast: a lab with $4M of annual Medicare CLFS revenue, under an assumed uniform 15% annual cut, ends 2029 at 61.4% of baseline with a cumulative three-year shortfall against a flat baseline of roughly $3.25M. A $25,000 fee is 0.77% of that. That ratio justifies nothing on its own; it earns its keep only if the output changes one decision worth more than the fee, which Ardia has never once demonstrated.
The honest first answer is that the incumbent is Microsoft Excel and it wins most of the time, because a compounding formula is a thirty-second exercise and the CFO already trusts her own spreadsheet. Any pitch that does not confront that is not credible. Commercial competitors, all materially ahead: XIFIN's RPM is widely regarded as the dominant lab-specific revenue-cycle platform and markets PAMA reporting support; Telcor and Ligolab occupy the same layer at smaller scale; hc1 sells lab analytics on real volume data. The specific claims about any incumbent's customer counts, data holdings, or forecasting capability are unsourced here and should be substantiated from their own published materials before use. The strategic point survives regardless: any incumbent already holding a lab's adjudicated claims can compute a per-code forecast from data Ardia does not have and cannot obtain without a design partner and, for raw files, a BAA. On rate transparency, Turquoise Health, Serif Health and Payerset have ingested payer machine-readable files under the Transparency in Coverage rule and can answer what a named payer pays for a code in a market; Ardia's mocked-up rate-intelligence screen is illustrative UI, not data the company holds. On the advisory side, ACLA publishes PAMA impact modelling to members at no cost, CMS publishes the complete CLFS files free, and lab-focused consultancies sell PAMA exposure studies as a standard workstream staffed by people with twenty years of lab P&L experience. Ardia's differentiation is narrow but real: an open, readable, unit-tested model that runs client-side with no procurement and no BAA, usable inside a first sales meeting, sitting in a portfolio that also proposes to work the denial side. That is a go-to-market motion, not a technology position. Ardia has won zero competitive evaluations because it has entered zero.
Sequenced with the dependency chain exposed, and with what would count as failure attached. Step 0, this week, is an honesty fix costing about a day: either wire the Studio's Meridian to the deterministic module as a tool call — parse the mix, call clfs.py, have the model narrate a number it did not compute — or remove Meridian from the Studio picker and point it at model-pama.html. Alongside it, correct any site copy naming Claude as the engine, and pin the Gemini model id to an explicit version rather than a '-latest' alias. Failure condition: the contradiction survives another diligence conversation. Step 1 is the CMS CLFS public use file — ingest full national payment amounts with effective dates and a version stamp, retire the hard-coded floats, and re-point tests at the ingested table. It is free, public, needs no agreement, and converts a slider into something that knows what a code pays. Failure condition: not shipped within 30 days, since nothing blocks it. Step 2 is the retrospective backtest described under evaluation design, which is the first number Ardia can publish that is not company-reported arithmetic about itself. Step 3 is the applicable-laboratory logic from 414.502 plus the real reporting calendar, turning a penalty multiplication into a compliance answer; step 4 is the ADLT and new-code exception layer plus per-code modelling; step 5 is CSV upload with Excel and board-PDF export. Step 6, gated on a BAA and on moving the HIPAA matrix off 2 of 15, is 835 ERA ingestion — the highest-value integration and correctly the last. The real blocker is none of these: it is one design-partner lab, which is the only thing on this list a single engineer cannot manufacture.
