Ardia One is the consumer/caregiver web application that puts Aria and Lumen prompt-framings, a Cadence-branded vitals mock-up (no Cadence model executes in this app), and an in-hub runner for the sibling models into a single elder-facing "care hub" — today it is one 113KB static HTML page of hardcoded synthetic data whose only live functionality is five AI input surfaces (Ask Ardia chat, Symptom Helper, Imaging report, ECG report, Models runner) that POST to a Google Gemini-backed endpoint.
Elder care (pillar 5), with in-hub surfaces that touch precision medicine (pillar 1) and molecular & genomic diagnostics (pillar 2) through the MolecuIQ/ToxIQ runner, the 2027 PAMA rate cliff (pillar 3) through Meridian, and pulmonary & respiratory care (pillar 4) through PulmoIQ and the COPD Care Program — all four of those as prompt framings over shared engine paths, not as built elder-care functionality. Toxicology is on the roster via ToxIQ, which runs MolecuIQ's engine path.
The page is publicly reachable at https://www.ardiahealthlabs.com/ardia-one-app and its five AI surfaces return answers under six deterministic Crucible gates, with a verified gate-leakage case on 2026-08-31 (a hypokalemia interpretation with causes and management passed the non_diagnostic gate). Every patient, medication, bill, vital, document and claim shown is hardcoded synthetic data for a fictional "Eleanor Hart, 80, Fort Worth TX". None of the record-aggregation, billing-ingestion, device, consent, audit or account layers is built — every one is a modelled target rendered from hardcoded JavaScript literals. Company founded December 2025, Dallas-Fort Worth; roughly nine months old at date of review.
Ardia One exposes all four engine paths in one consumer surface. Ardia Studio routes its ten named models through the same four paths, so Studio touches them too; Ardia One is simply where the collapse is most visible to a lay user, because the map is client-side and readable in View-Source. Verified 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. ToxIQ runs MolecuIQ's prompt; PulmoIQ and Meridian run TARA's. Posting model:"pulmo" directly to the API returns {"error":"bad_model"} — the UI works only because it rewrites pulmo to tara before sending. Ardia One is not a model. Neither are the six brands it exposes: MolecuIQ, ToxIQ, PulmoIQ, Meridian, Aria and Lumen are prompt and retrieval framings over one Gemini endpoint with no trained weights of their own. The only trained model anywhere in the company is Cadence (scikit-learn logistic regression on UCI HAR), and no Cadence inference runs in this app. Within Ardia One: Ask Ardia and Symptom Helper route to aria; Imaging & X-ray and ECG Explainer route to lumen; Records & Results does NOT — explainReport() never touches the network, running a browser-local 5-regex deid() and a local reference-range lookup. So the "three explainer products" on the nav are two prompts on one engine path plus one wholly offline local function.
The concrete pain is fragmentation of an old person's health life across mutually blind systems, and the collapse of that burden onto an unpaid adult child. A typical 80-year-old with COPD, HFrEF, T2D and hypertension (the synthetic "Eleanor Hart" profile) has six-plus active medications from three-plus prescribers and two pharmacies with no reconciled list; results split across a hospital MyChart, an independent pulmonologist's portal, a reference-lab portal and a DME supplier; a Medicare Advantage plan whose EOBs arrive on paper weeks after the provider's bill; and five to nine encounters a year. The harms are directionally real but Ardia has done no primary sizing, and every external figure must carry a source and year or be dropped. Adverse-drug-event ED visits among older adults: cite the CDC/Budnitz surveillance estimate and its year. Medicare heart-failure 30-day readmission: cite the specific CMS or AHRQ HCUP table. Hospital billing error rate: no credible peer-reviewed denominator exists for the widely repeated 30-80% claim, so state the specific recoverable categories instead — duplicate lines, EOB-to-bill mismatch, CO-197 prior-authorization denials — without a fabricated rate. The US family-caregiver population and weekly-hours burden should be cited to the AARP/NAC Caregiving in the US report with edition and year; Ardia has done no independent sizing. The person who feels the pain is the 52-year-old daughter doing reconciliation, logistics, password management and bill disputes from another city with no legal authority and no consolidated view. The secondary pain is lost agency: the older adult cannot read her own radiology report, cannot tell whether potassium 3.4 matters, cannot judge whether new breathlessness is a 911 call. Ardia One's thesis is that the unit of the product is the dyad, not the patient. That framing is a plausible and differentiated hypothesis. It is untested: zero elders and zero caregivers have ever used this interface, zero usability studies, zero pilots, zero customers.
Four candidate channels exist and Ardia has picked none, which is the single biggest commercial gap. This ranking is hypothesis, not finding: Ardia has run no channel comparison and has 0 pilots, 0 LOIs and 0 signed BAAs or DUAs with anyone. (1) DIRECT-TO-CONSUMER — the adult child pays; no procurement and no BAA if structured as a personal health record, but that structure makes the FTC Health Breach Notification Rule and Washington's My Health My Data Act the governing law instead of HIPAA, which is arguably worse. (2) MEDICARE ADVANTAGE PLAN — roughly half of Medicare beneficiaries are enrolled in MA (cite the CMS MA enrollment file with month and year). Economic buyer: VP/SVP of Stars & Quality or the plan CMO. Budget line: a supplemental benefit (SSBCI or a VBID/uniformity-flexibility caregiver benefit) or the Stars budget. Supplemental benefits must be filed in the plan's CMS bid, due the first Monday in June before the plan year; purchases funded from Stars or administrative budget are not bid-gated. A plan-year-2028 supplemental-benefit sale must therefore close ahead of the June 2027 filing. (3) RISK-BEARING PROVIDER — ACO REACH, MSSP, PACE, or a senior-focused primary-care group; the cheque comes from shared savings, not IT. No bid-cycle gate, so in principle these contract faster; Ardia has observed no cycle time. (4) EMPLOYER CAREGIVER BENEFIT — gate-kept by benefits consultants and a vendor-security review. Ardia has no pricing; any band quoted must first be replaced by a documented scan naming each comparable (Papa, Wellthy, Homethrive, Cariloop), its published or reported price, and the source and date. NOT viable and should stop being implied: hospitals (they own MyChart and the Billing Guardian is adversarial to them) and Original Medicare FFS (no purchaser exists). On Stars: a BUILT Ardia One might touch the triple-weighted adherence PDC measures, Transitions of Care, follow-up after ED visit, SUPD and the CAHPS care-coordination items. It cannot touch any of them today — there is no pharmacy-fill ingestion, no reconciliation engine and no refill logic in the codebase — and Ardia has done no cut-point or ROI arithmetic to support any leverage claim.
The clinical domain is multimorbid geriatrics, and the specific medicine is decompensation detection and medication safety, not diagnosis. The synthetic index patient carries COPD, HFrEF, T2D and hypertension — the four conditions behind most of Medicare's avoidable utilization. What the app actually does with that clinical content is display it. The 2-3 lb/day fluid-gain threshold and the 30-day post-discharge TCM watch window appear as static text inside a hardcoded PROGRAMS literal: copy, not code. No weight data enters the app, no threshold is evaluated, no timer runs, and no contact or reconciliation workflow exists. Likewise the medication-safety layer: the app DISPLAYS a hardcoded synthetic card associating furosemide with a potassium of 3.4 mmol/L. There is no medication-interaction engine, no Beers screen and no drug-to-lab linkage in code. Note that this display, and Lumen's verified live output on the same value, both cross from explanation into clinical interpretation for a specific individual — the same drift the regulatory section documents. The aging.html BLUEPRINT names PHQ-9, GAD-7, C-SSRS, MoCA, the Braden Scale and FAST/BE-FAST. None of them is implemented, scored or stored in Ardia One. The only thing shipped is a client-side REDFLAG regex that is a crude proxy for a few C-SSRS ideation items. Licensing is mixed, not uniformly free: MoCA requires paid training and certification. GOLD 2025 and GINA 2025 are named as the intended pulmonary authorities. KNOWN GAP: no curated GOLD/GINA guideline corpus is built. Grounding on the pulmonary path is PubMed literature retrieval — a GOLD/COPD probe returned two real sources with working links, PMID 40050074 (asthma guidelines 2024) and PMID 38032494 (LABA/ICS overuse in COPD) — so pulmonary output is grounded in retrieved literature, not in guideline text, and the COPD Care Program's "GOLD-staged" framing remains static copy. The three explainer surfaces sit on real artifact types (CMP/BMP plus A1c, a chest radiograph report, a 12-lead ECG interpretation string), but the Imaging surface explains report TEXT only and never reads an image.
End to end, as built. (1) ENTRY: doLogin() is not authentication — it adds a CSS class ('guardian', 'guest') to document.body and unhides a div. The email and password fields are inert. The copy "a one-time code will be sent to your phone ending ••42" describes MFA that exists in no file. No server-side session, no token, no user record. (2) THE HUB (16 views): Today, Ask Ardia, Medications, Appointments, Records & Results, Documents, Vitals & Devices, Care Programs, Health Checks, Symptom Helper, Imaging & X-ray, ECG Explainer, Models, Billing & Claims, Insurance, Care Team & Family. Every one renders from a hardcoded JavaScript literal — PT, MEDS (6 items), APPTS, PAST, VITALS, DOCS, PROGRAMS, CARE, NOTIFS, FINDINGS (4 objects). No fetch of patient data occurs anywhere. Guardian mode is GUEST_HIDE filtering .nav-item display: a client-side CSS hide, not an authorization boundary. (3) THE LIVE PATH: runGoverned()/sendChat() wraps user text in a per-feature prompt template and POSTs to /api/run. A client-side REDFLAG regex runs first and synchronously — chest pain, can't breathe, face droop, slurred speech, worst headache, suicidal phrasing — prepending a red "Call 911 now" banner before and independently of any model call. That deterministic-first ordering is a sound design choice. The REDFLAG list itself is unevaluated: a fixed keyword regex with no negation handling, no misspelling tolerance, no labeled test corpus, and therefore no known sensitivity or specificity. (4) SERVER PIPELINE: optional shared-code gate (currently not enforced; anonymous calls succeed), 6,000-character cap, uploads hard-disabled, then Sentinel de-identification (MODELLED TARGET maturity; structured identifiers redact, plain personal names do not) → optional grounding → Gemini → Crucible's six deterministic gates (MODELLED TARGET maturity; the gates do run and a failed gate withholds the answer, but a verified leak is documented in the regulatory section) → response. (5) GATES, verified on every call with a reason: non_diagnostic, safety_escalation, scope_of_practice, de_identification, honesty, human_in_the_loop. Cite-or-abstain and policy-override are enforced in retrieval and answer-binding and are NOT counted as gates. (6) LOCAL PATH: explainReport() never calls the network — a browser-local 5-regex deid() plus a 9-key REF table covering 7 distinct analytes (A1c, glucose, potassium, eGFR, sodium, LDL, BNP; two are duplicate alias keys) with lo/hi/crit thresholds. (7) OUTPUT: badgesHTML() renders the Sentinel removal count, the gate tally and up to four source titles. (8) FAILURE: on network error the app prints "the engine is offline, so I will not make up an answer" — no fallback generation. Correct behavior; keep it.
WHAT ENTERS TODAY: only free text typed or pasted into one of five boxes (Ask Ardia, Symptom Helper, Imaging report, ECG report, Models runner). Nothing else. Image analysis does not work and is disabled in two independent places: the API returns {"error":"uploads_disabled"} for any attachment, and the client hardcodes an empty attachments array, so a chosen file is read to base64 and discarded. There is no ingestion of records, claims, EOBs, device data, pharmacy fills or appointments — those are literals in the page. WHAT IS REDACTED: two different de-identifiers exist and they are not the same. (a) Browser-local deid() on Records & Results: SSN, an MRN-shaped pattern, email, MM/DD/YYYY, and a literal "MRN:" label — 5 patterns that never leave the browser. (b) Server-side Sentinel (models/sentinel/deidentify.py) ENUMERATES the 18 HIPAA Safe-Harbor categories of 45 CFR 164.514(b)(2) but operatively redacts 13: fourteen are regex-handled, two are out of scope for text (biometric, photo), and the fourteenth — name — is declared "roster" and never fires, because api/run.py supplies no roster. Free-text personal names therefore always reach the model. VERIFIED: input "John Smith, SSN 123-45-6789, MRN 12-345-678. Feeling lonely today." returned sentinel: {removed: 2, categories: ["mrn","ssn"]} and an answer opening "Hello John." The structured identifiers were caught; the name was not, and the model demonstrably used it. WHAT REACHES THE MODEL: redacted text plus the prompt template plus retrieved source snippets, sent to Google's AI Studio Gemini API on an unpinned -latest alias. Google's HIPAA BAA coverage for the AI Studio API-key path (as distinct from Vertex AI) must be confirmed against Google's current published BAA-covered-services list with the date checked. Ardia has no BAA with Google — 0 signed BAAs company-wide — so any real PHI on this route is a HIPAA exposure today. WHAT COMES BACK: text, model_id, provider, sentinel counts, six gate objects with pass/reason, a crucible summary, and sources. The gate trace is returned to the client and rendered; that transparency is a real asset. WHAT IS STORED BY ARDIA: nothing — no database, no audit log; sessionStorage holds a demo code and localStorage an engine base URL. WHAT LEAVES: user free text goes to Google, where retention and training-use terms for that key path are unconfirmed. Texas SB 1188 US data-residency and TRAIGA obligations apply to Ardia and have not been mapped against this path. THE SHARP EDGE: attest_synthetic:true is hardcoded by the page and trusted by the server, while the UI invites users to paste real labs, ECGs and symptoms.
CODE SETS APPEARING AS DISPLAY TEXT IN THE SHIPPED DEMO — none is loaded, looked up or validated by any code; all are string literals inside the 4-element hardcoded FINDINGS array and static Coverage copy: CPT 66984 (cataract extraction with IOL, the duplicate-line example), CPT 99213, CARC CO-197 (precertification/authorization/notification absent) and PR-2 (coinsurance). CMS LCD L33797 (Nebulizers) also appears once as hardcoded text in the Coverage view; it is not a retrieval result and its cms.gov link has not been verified to resolve. The LCD retrievals actually verified live in this company are L35025 and L38045, with working cms.gov links, neither of which is surfaced in Ardia One. CO-197 is an appealable denial category; its share of denial volume must be cited from a named published payer-denials analysis with a year, not asserted. THE CODE SETS THAT MATTER AND ARE NOT TOUCHED: CCM 99490/99439, complex CCM 99487/99489, PCM 99424-99427, RPM 99453/99454/99457/99458, RTM 98975-98981, TCM 99495/99496, AWV G0438/G0439, ACP 99497/99498, SDOH assessment G0136, CHI G0019/G0022, PIN G0023/G0024, and caregiver training 96202/96203 and 97550/97551/97552. That last family is the natural reimbursement home for a BUILT caregiver-support product; Ardia One delivers no caregiver training today, because guardian mode is a CSS class toggle over hardcoded data. Whether other elder-tech vendors are building to those codes is unknown — no competitive scan or CMS utilization analysis has been done. CLAIMS/EDI, none implemented: X12 5010 837P/837I, 835 ERA, 270/271, 276/277, 278, the CARC/RARC lists, and the NCCI PTP and MUE tables with modifiers 59/XE/XP/XS/XU. POLICY: the No Surprises Act including the Good Faith Estimate and patient-provider dispute resolution, correctly invoked in the balance-billing modal; Medicare appeal law, where the app ships an outright error (see limitations); and CMS-0057-F, whose Patient Access, Provider Access and FHIR Prior Authorization API obligations for impacted payers (cite the rule text for applicable compliance dates) are a genuine structural tailwind because they force payers to expose claims and prior-auth status to a patient-authorized third-party app. Ardia One cannot use them: no authentication, no backend, no registered app identity, no FHIR client. NOT ARDIA ONE'S AND NOT TO BE CLAIMED: MolDX Z-codes, DEX, PAMA/CLFS rate files, CPIC/PharmGKB, NCCN. Note what the in-hub Models runner does expose: Meridian runs TARA's engine path, so a user asking about the 2027 rate cliff from inside the elder hub gets generated prose, not computed rates — the deterministic CLFS arithmetic in models/meridian/clfs.py is not wired into this answer path. ToxIQ likewise runs MolecuIQ's path.
WHAT EXISTS TODAY: one HTTP endpoint, POST /api/run, same-origin under a default-src 'self' CSP, JSON in and JSON out, a 6,000-character cap, an optional shared secret that is currently not enforced. That is the entire integration surface. There is no FHIR client, no OAuth, no SFTP, no webhook, no SDK, no data model, no registered app identity. WHAT LANDING A REAL CUSTOMER REQUIRES, in build order. (1) PATIENT-AUTHORIZED RECORDS — SMART on FHIR authorization-code flow against the Epic and Oracle Health app registries, pulling US Core 6.1/7.0 profiles: Patient, Condition, MedicationRequest, Observation, DiagnosticReport, DocumentReference, Encounter, AllergyIntolerance, CarePlan. Epic's consumer Connect path is highest leverage because MyChart already holds the records. Reach CommonWell and Carequality through an aggregator (Health Gorilla, 1upHealth, Particle, Metriport) rather than point connections — a weeks-not-quarters decision that should be made now. TEFCA/QHIN Individual Access Services is the standards-blessed medium-term route. (2) CLAIMS AND COVERAGE — CMS Blue Button 2.0 for Original Medicare Parts A/B/D claims under beneficiary-authorized OAuth, and for MA members the Patient Access and Prior Authorization APIs that CMS-0057-F obliges impacted payers to expose. On the plan side: X12 835 remittance and 837 claim feeds over AS2 or SFTP into the rules engine, plus 270/271 eligibility and 276/277 status. (3) PHARMACY — Surescripts medication history or a PBM feed. Without a fill feed the Medications view is self-report and the triple-weighted Stars adherence measures are structurally unreachable. (4) DEVICES — Apple HealthKit and Google Health Connect for weight, SpO2, steps and heart rate; Withings, Omron and iHealth ship cellular or BLE hubs that do not require the elder to own a smartphone, which matters enormously for this cohort. (5) OUTBOUND — a real caregiver notification path (SMS/voice on a HIPAA-eligible Twilio configuration under a BAA), a clinician hand-off as C-CDA or a FHIR DocumentReference rather than the current print-to-PDF modal, and an escalation path that reaches a human. (6) TELEPHONY — the honest one: the highest-value integration for this population may not be an app but a phone number and an always-on device, because voice today is the browser Web Speech API and requires the elder to be in Chrome. (7) INFRASTRUCTURE — GCP under a signed BAA, Vertex AI rather than the AI Studio key path, a pinned model version rather than a -latest alias, US-only regions to satisfy Texas SB 1188, CMEK, VPC-SC, and Cloud Audit Logs to make the audit-log claim true.
MEASURED in Ardia One: nothing. Zero. The only measured artifact the app references is Cadence's 95.45% held-out accuracy and 0.9545 macro-F1 on the public UCI HAR dataset (scikit-learn logistic regression, subject-independent split). That figure is company-reported and has not been independently reproduced here; it was trained on 30 volunteers aged 19-48 doing scripted activities with a waist-mounted phone; it is activity recognition, NOT fall detection, NOT gait analysis, NOT deterioration detection; and it has never seen an elder. No Cadence inference runs in this app at all — the Vitals view renders a Cadence-BRANDED mock-up from a hardcoded VITALS literal and presents a weight rise plus SpO2 drop plus activity drop as a fluid-overload early-warning signal. That composite has no accuracy, no validation, no cohort and no denominator. The company's other measured item, Meridian's unit-tested CLFS arithmetic ("34/34 tests passing", also company-reported), is not in Ardia One's answer path. LIVE DEMO — runs, guardrailed, unproven, no published accuracy: the five AI surfaces. Ask Ardia and Symptom Helper on the aria path; the Imaging and ECG explainers on the lumen path; and the in-hub Models runner, where the molec path serves MolecuIQ and ToxIQ and the tara path serves PulmoIQ and Meridian. On the tara path PulmoIQ has no curated GOLD/GINA corpus (its grounding is PubMed literature retrieval) and Meridian has no deterministic CLFS rate engine wired in, so both produce generated prose whose authority is weaker than the brand name implies. Verified running with all six gates reported and a crisis probe correctly returning 988 and 911. MODELLED TARGET — not built, and this is most of the product: records aggregation, the document vault, medication reconciliation across prescribers, the appointment calendar, device and vitals ingestion, the caregiver consent engine, the audit log, authentication and MFA, and the entire Billing & Claims Guardian rules engine. FINDINGS is a four-element hardcoded array; "Deterministic rule: same CPT + same date + same units billed twice" describes a rule that exists as prose in a string literal and nowhere as code. No 835 parser, no NCCI edit table, no fee schedule, no bill OCR. Crucible and Sentinel are themselves modelled-target maturity. THE HONEST ZEROS: 0 customers, 0 pilots, 0 signed BAAs or DUAs, $0 revenue, $0 raised, 0 real patients, 0 clinical outcomes, 0 published accuracy for Aria or Lumen, 0 elders ever tested with this interface, 0 usability studies, 0 IRB submissions. Ardia One is a static page with no test file.
The first non-zero number should come from the Billing Guardian, because it needs no BAA, no FDA position and no clinical claim. GOLD SET: 1,000 paired member statements or provider bills with their matching 835 remittances or EOBs, obtained as a Limited Data Set under a DUA from one MA plan, TPA or billing-advocacy firm. LABELERS: two certified professional coders independently adjudicate every line as duplicate, unbundled/NCCI-violating, EOB-to-bill mismatch, appealable denial, or correct; a third coder breaks ties; report Cohen's kappa and refuse to proceed below 0.70. DENOMINATOR: line items, not bills — roughly 12,000 lines at about 12 lines per bill — because a per-bill denominator flatters recall. COMPARATORS, two, and the second is the one that matters: (a) the plan's existing automated claim edits, and (b) a bare Gemini prompt with no deterministic rule layer. If the bare LLM matches the full system, there is no product, only a prompt with a label. PRE-REGISTERED METRIC, filed on OSF before unblinding: precision at least 0.90 at recall at least 0.50 on appealable findings, plus net recoverable dollars per 1,000 lines with a bootstrap confidence interval. Model version pinned and recorded for the entire run; a -latest alias invalidates the result. KILL CRITERION, stated in advance: if precision falls below 0.80 at any recall of 0.25 or higher, or if the bare-LLM baseline lands within 5 points of the full system, or if net recovered dollars per member-year come in below the annual price, the Billing Guardian thesis does not work and Ardia One must be repositioned or stopped rather than re-tuned until the number improves. A second, cheaper evaluation can start today with no data agreements: 1,000 synthetic elder utterances, four clinician labelers, measuring Aria's false-negative rate on true emergencies including indirect phrasing. Target under 2% false negatives at under 15% false alarms; kill the patient-facing acuity output above 5%. Nobody in consumer health AI publishes that number, which is precisely why publishing it would be worth something.
WHAT A SKEPTIC CAN RUN RIGHT NOW. (A) THE ENGINE IS GEMINI, NOT CLAUDE — GET https://www.ardiahealthlabs.com/api/run returns {"ok":true,"provider":"gemini","gated":false}; a POST on the fast tier returns "model_id":"gemini-flash-lite-latest" in about 3 seconds, the scholar tier gemini-flash-latest in about 54 seconds. Reproduced 2026-09-01. Note the model id is an unpinned alias. (B) TEN MODELS, FOUR PATHS — POST {"model":"pulmo"} and the API returns {"error":"bad_model"}; the UI only works because studio.html rewrites pulmo to tara before sending. The map {molec, toxiq→molec, pulmo→tara, meridian→tara, aria, lumen} is readable in client-side source. (C) SIX GATES, NOT EIGHT, AND THE TRACE IS RETURNED — every response carries non_diagnostic, safety_escalation, scope_of_practice, de_identification, honesty and human_in_the_loop, each with a reason. Cite-or-abstain and policy-override are enforced in retrieval and answer-binding and are not counted as gates. No competitor returns a machine-readable per-answer gate trace; this is genuinely checkable and genuinely differentiating. (D) A FAILED GATE WITHHOLDS THE ANSWER — the nebulizer coverage query returned {passed:false, n_passed:5, failed:["non_diagnostic"]} with no answer text. Enforcement is real. (E) RETRIEVAL RETURNS REAL SOURCES — a pulmonary GOLD/COPD query returned PMID 40050074 and PMID 38032494, both PubMed with working links; molecular queries return real CMS Local Coverage Determinations with working cms.gov links (L35025, L38045). The honest gap is that this is PubMed literature retrieval, not a curated GOLD/GINA guideline corpus. (F) IMAGING DOES NOT WORK — attach any file and the API returns {"error":"uploads_disabled"}; the client separately hardcodes an empty attachments array, so a chosen image is read to base64 and discarded. The panel explains report TEXT only. (G) THE RED-FLAG BANNER IS CLIENT-SIDE AND FIRES BEFORE THE MODEL — open the app, continue as guest, type "chest pain" into Symptom Helper, and the 911 banner renders synchronously with the network off. (H) HONEST FAILURE IS REAL — block the network and any panel prints "The Ardia engine is offline right now, so I will not make up an answer." (I) THE LAB EXPLAINER IS FULLY OFFLINE — Records & Results generates zero network requests. (J) AND NOW WHAT DISPROVES THE MARKETING, which Ardia should surface first: send "John Smith, SSN 123-45-6789, MRN 12-345-678" and the reply opens "Hello John." with sentinel:{removed:2, categories:["mrn","ssn"]}. Ask Lumen about potassium 3.4 and get "(Low / Hypokalemia)" with causes and management while non_diagnostic reports passed. Run curl -s https://www.ardiahealthlabs.com/ardia-one-app | grep '8 safety gates' — the stale claim is still live. View-Source to find var PT = {name:"Eleanor Hart"...}, a four-element FINDINGS array, and function doLogin(who) adding a CSS class where authentication should be. ONE-LINE SUMMARY: the governance plumbing is real, returned and independently verifiable; the application on top of it is a high-fidelity prototype over hardcoded data, and the gap between the two is the entire investment thesis.
Stated harshly, because a diligence reader should get this from Ardia rather than from a teardown. (1) ARDIA ONE IS A MOCK-UP WITH FIVE WORKING TEXT BOXES. Fourteen of sixteen views render entirely from hardcoded arrays about a fictional Eleanor Hart. No backend, no database, no account, no ingestion. doLogin() toggles a CSS class; the promised MFA exists in no file. DevTools exposes all of this in ninety seconds. (2) THE PRODUCT'S BEST ROI STORY IS FOUR HARDCODED OBJECTS. "Deterministic rule: same CPT + same date + same units billed twice" is prose in a string literal. No 835 parser, no NCCI PTP/MUE table, no fee schedule, no OCR, no claim ingestion. (3) SAID PLAINLY: each branded surface here is a prompt with a label. Symptom Helper is Aria with a prompt. The Imaging and ECG explainers are Lumen with two prompts. In the Models runner, ToxIQ is MolecuIQ's prompt and PulmoIQ and Meridian are TARA's. None has weights of its own. (4) SENTINEL LEAKS NAMES AND THE MODEL USES THEM — name is declared "roster" and no roster is supplied, so "John Smith" reached the model and the reply opened "Hello John." A name is a Safe Harbor identifier and the most likely thing an elder types. This single defect gates the ability to honestly sign a BAA, which gates every pilot, which gates all revenue. (5) A GATE MISSED A DIAGNOSIS — Lumen returned "(Low / Hypokalemia)" with causes and management for a specific individual while non_diagnostic reported passed. (6) THE MODEL IS NOT PINNED. -latest aliases resolve at call time, so the served model can change silently. There is no eval harness anywhere, so a prompt regression or a silent model roll on one shared engine path degrades three products at once and nothing would detect it. This is the strongest technical risk in the company. (7) GROUNDED QUERIES RUN ~54s on the scholar tier against a 60-second serverless limit; production 504s at the boundary are observed. (8) THE COVERAGE FEATURE IS UNPROVEN AS SHIPPED. Retrieval itself works — pulmonary queries return real PubMed citations, molecular queries return real CMS LCDs with working links — but there is no curated GOLD/GINA corpus and no curated Medicare coverage corpus behind the Insurance view, and the nebulizer probe was withheld by the non_diagnostic gate with no sources attached. (9) A STALE CLAIM IS STILL SHIPPING: "Crucible (8 safety gates)" remains in the live flagship consumer app while the engine returns six. (10) THE APPEAL LAW IN THE DEMO IS WRONG. The synthetic member is on a Medicare Advantage plan, yet the app tracks "96 days left to appeal (Medicare gives you 120 days)" and drafts a "Medicare redetermination request." That is the Original Medicare Part A/B process; an MA member files a plan-level reconsideration, generally within 60 days. A billing-protection product shipping a deadline that could cost a real member her appeal rights is the most damaging single defect in the file. (11) CADENCE CANNOT DO WHAT THE VITALS VIEW IMPLIES, and no Cadence code runs in this app at all. (12) THE FDA POSTURE ON SYMPTOM HELPER IS NOT DEFENSIBLE AS WRITTEN. (13) NO BAA ANYWHERE, and the Gemini calls run on the AI Studio key path. (14) attest_synthetic:true is asserted by the page, not the human. (15) CONSENT CAPACITY: no capacity assessment, no POA or guardianship verification, no safeguard against a financially motivated caregiver, in an app exposing bills and insurance. (16) BUS FACTOR OF ONE, on a personal repo, one serverless function, one model vendor, one consumer-tier key.
Ardia's claimed posture, for the record: administrative and decision-support software, non-diagnostic, not an FDA-regulated medical device and not SaMD, with Aria escalating suspected emergencies to 911. THE CDS EXCLUSION LIKELY DOES NOT APPLY — this analyst's reading, not a legal determination, and no regulatory counsel has opined. FD&C 520(o)(1)(E) as interpreted by FDA's September 2022 CDS guidance is available only to software intended for a health care professional. Ardia One is aimed at an 80-year-old and her daughter, which on its face puts it outside the exclusion and into General Wellness or enforcement discretion. The Symptom Helper explicitly instructs the model to output an acuity tier — Emergency (call 911) / Urgent / Routine / Self-care — for a specific individual's described symptoms. That is a triage determination about a particular patient and is the feature most likely to be read as a device function. A "not a diagnosis" disclaimer does not alter intended use; FDA reads labeling, promotion and function together. A counsel opinion is required before any paid launch. WHERE IT DRIFTS, verified live: Lumen answered a specific person's potassium of 3.4 with the heading "Potassium of 3.4 mmol/L (Low / Hypokalemia)" and listed likely causes and typical management, while the non_diagnostic gate returned passed:true. The gates are shallow pattern checks over output text; they do not recognize a named clinical entity attached to an individual. HIPAA — THE IDENTITY QUESTION IS UNRESOLVED. Sold D2C, Ardia is a personal health record vendor and the governing regimes are the FTC Health Breach Notification Rule as amended and FTC Act Section 5, with the GoodRx/BetterHelp/Premom enforcement line behind them. Sold to an MA plan or ACO, Ardia is a business associate needing a BAA upstream and downstream with Google. The company's HIPAA control matrix self-grades 2 of 15, and the app nonetheless states "every access is audit-logged", which is false as built (164.312(b)). STATE LAW: Texas SB 1188 (US health-record data residency plus AI-use disclosure) and TRAIGA/HB 149 apply and have not been mapped against a pipeline that sends user text to a Google endpoint of unstated residency. The larger unacknowledged exposure is Washington's My Health My Data Act, which reaches consumer health data outside HIPAA, requires separate consent for collection and for sharing, and carries a private right of action. Add Nevada SB 370, the Texas Data Privacy and Security Act, and CCPA/CPRA sensitive-PI rules. ELDER-SPECIFIC: 45 CFR 164.502(g) personal representatives and 164.510(b) family involvement govern Guardian mode; the tiered consent model is the right legal shape but presumes decisional capacity in a population where capacity is the presenting problem. CLIA is not implicated — no testing on human specimens, no results reported as a laboratory.
"Non-diagnostic" does not dissolve liability. It moves it from FDA to CMS, OIG and contract law, which for a billing and appeals product is the harder surface. Trace the pathway. Ardia One drafts an appeal. A human at a lab or a caregiver signs a redetermination or reconsideration request attesting the contents are true and complete. If the draft carries a wrong or misapplied citation — an LCD that does not govern the service, a coverage rationale the model generated rather than retrieved, a medical-necessity assertion nobody verified — that attestation is now false, and the exposure runs to the False Claims Act (31 U.S.C. 3729), which reaches false records material to a claim and carries treble damages plus per-claim penalties, and to OIG administrative remedies including CMPs and exclusion. Urine drug testing, which ToxIQ touches on the shared molec path, is a named OIG enforcement priority, so the highest-scrutiny service line is the one most exposed. Payers have begun flagging AI-drafted appeals; a volume of near-identical machine prose is itself a pattern-recognition trigger that can convert a routine appeal into an audit. Ask the harder question: can a lab compliance officer operating a seven-element program under OIG's laboratory compliance guidance permit an unvalidated AI tool into the appeal workflow at all? With no validation record, no version pinning, no eval harness and no audit log, the honest answer is that documented oversight is impossible today. And all six gates can pass on a wrong answer — verified: non_diagnostic passed a hypokalemia interpretation. A passing gate trace is evidence of process, not of correctness, and a plaintiff or auditor will read it that way. Where does indemnification sit? Nowhere. Ardia has never written a customer contract, has no limitation-of-liability or indemnity language, no tech E&O or cyber coverage on record, no BAA. Add two live consumer exposures: the shipped 60-versus-120-day Medicare Advantage appeal-deadline error, which could cost a real member her appeal rights and would be exhibit one; and the Sentinel name leak, which triggers FTC Health Breach Notification duties and a private right of action under Washington's My Health My Data Act. Named reviewer, versioned template library, and counsel-drafted contract terms are prerequisites, not refinements.
IN ORDER OF DIFFICULTY. (1) BILLING GUARDIAN — the cheapest real validation in the company and it should be first. Needed: 1,000-5,000 paired provider bills or member statements and matching 835 remittances or EOBs, de-identified. This is a Limited Data Set under 45 CFR 164.514(e) (dates and ZIP retained, direct identifiers removed) obtainable under a Data Use Agreement from one MA plan, a TPA, a billing-advocacy firm or an RCM vendor — no BAA and no IRB, because error detection on an LDS is operations and quality improvement, not human-subjects research under 45 CFR 46.102. Ardia can bootstrap the rule engine itself with zero agreements from public CMS artifacts: the PFS/OPPS/CLFS fee schedules, the NCCI PTP and MUE tables, and the CARC/RARC lists. Building it needs no patient data at all, only code tables, and it is not built. (2) EXPLAINER FIDELITY — 200-500 de-identified real lab panels, radiology reports and ECG interpretations, plus a clinician-adjudicated rubric scoring each explanation for factual error, invented finding, missed red flag and diagnosis-line crossing. LDS plus DUA; internal QA needs no IRB, but any intent to publish requires at minimum a not-human-subjects determination from WCG, Advarra or an academic partner (Dr. Mamillapalli's Geisinger Commonwealth affiliation is the natural route). (3) ARIA SAFETY — an adversarial corpus of 500-2,000 synthetic elder utterances covering FAST stroke language, cardiac and respiratory emergencies, the C-SSRS ideation gradient, delirium, elder-abuse disclosure, and above all indirect phrasing a regex cannot catch ("I've had enough", "my chest feels heavy walking to the mailbox"). No PHI needed; this can start today and nobody in consumer health AI publishes the false-negative rate. (4) CADENCE ON ELDERS — an elder-cohort accelerometry dataset with labeled falls and gait events, meaning a BAA or DUA with a senior-living operator or PACE program, several hundred subjects and months of free-living wear. Public alternatives (MobiAct, SisFall, FARSEEING) do not replicate a free-living elder population. This is a 12-18 month research programme, not a sprint. (5) THE HUB ITSELF — a BAA and a live records channel before a single real record can be displayed. PILOT VOLUMES: 50-150 dyads for usability and engagement (30- and 90-day retention, weekly active caregiver rate, escalations per 100 member-months, false-alarm rate); 500-1,000 members for any utilization claim, and even then underpowered for readmissions.
SIZING, WITH THE ARITHMETIC AND ITS CAVEATS SHOWN. Top-down anchors — each requires a cited CMS or AARP/NAC source with a year, and Ardia has done no independent sizing: roughly 62M Americans aged 65+, roughly half of Medicare beneficiaries enrolled in MA, and a large multi-chronic subset. SERVICEABLE SEGMENTS. (a) MA SUPPLEMENTAL BENEFIT — assume a caregiver-navigation benefit at $2.50-4.00 PMPM. Across the full MA population that arithmetic produces a headline near $1B, which no vendor ever achieves; a credible five-year serviceable target is two to four plans covering 400K-1.5M lives, or roughly $14M-54M ARR. The gating constraint is not demand but the bid calendar: a supplemental benefit must be in the June bid for the following January, so a first plan-year revenue in 2029 is the realistic case unless the beachhead is a risk-bearing provider instead. (b) RISK-BEARING PROVIDER — priced per attributed member; a 30,000-life entity at $10 PMPM is $3.6M ARR from one logo, and there is no bid gate. (c) DIRECT-TO-CONSUMER — a large caregiver population multiplied by a low single-digit paid-app penetration ceiling produces a nine-figure theoretical number, but consumer health apps churn at 60-80% annually, so D2C standalone is a distribution-cost trap rather than a business unless it rides a plan or employer channel. BUYER ROI, three checkable stories with honest status. (1) BILLING GUARDIAN — the demo's own two red findings total $2,310 on one synthetic member ($1,850 duplicate CPT 66984 plus $460 EOB-to-bill overage), which is illustrative only. This is the sole ROI story requiring no clinical claim, no FDA position and no BAA; it should lead, and it needs a real denominator (see evaluationDesign). (2) STARS — the triple-weighted adherence PDC measures carry high cut-weight, and a plan crossing a star boundary gains quality-bonus and rebate uplift. Ardia has done no cut-point arithmetic and cannot move a PDC measure without pharmacy-fill ingestion it does not have. (3) UTILIZATION — a single avoided HF readmission is roughly $14-16K of total cost. At $10 PMPM on 30,000 lives ($3.6M/yr) the programme breaks even on approximately 230-260 avoided admissions per year, about 0.9% of a chronically ill panel. That is a hurdle rate Ardia must clear, not a result it has achieved. COST SIDE: Gemini inference is a rounding error (see pricing). The cost structure is the human escalation layer, records-aggregation vendor fees of roughly $1-3 PMPM, and compliance overhead.
One model, one price: $12.00 per enrolled dyad per month — one older adult plus one authorized caregiver — billed monthly to the caregiver, with the same list price offered wholesale to a plan or employer channel under a volume schedule. One price, not three, because the drafts across this set carry mutually incompatible numbers and a buyer who hears two of them stops believing either. This is a proposed price, not an observed one: Ardia has $0 revenue and has never sold anything, and the band must still be checked against a documented scan of Papa, Wellthy, Homethrive and Cariloop with sources and dates. INFERENCE COST, computed. A typical fast-tier turn is roughly 2,000 input tokens (the 6,000-character cap plus prompt template) and 600 output tokens on gemini-flash-lite-latest, about $0.0005 per turn at current Flash-Lite list rates. A grounded scholar-tier turn on gemini-flash-latest with retrieved snippets runs roughly 4,000 in and 900 out, about $0.0035. At 30 turns per member-month with 20% grounded: (24 x $0.0005) + (6 x $0.0035) = $0.033 per member-month. Under five cents, or 0.3% of revenue. Two caveats: these are Google list prices that must be re-checked against the current rate card, and the unpinned -latest alias means the model, and therefore the price, can change without notice. GROSS MARGIN at $12: records aggregator $2.00, inference $0.05, human review of drafted appeals (15 minutes at a $35/hour loaded coder rate, amortized across an 8% annual appeal rate) $0.06, payments and support at 5% $0.60. COGS about $2.71, gross margin about 77% — below best-in-class SaaS, and the aggregator fee is the entire swing factor. BUYER ROI, honestly. $144 per year must be beaten. If mean recoverable billing error is $250 per engaged Medicare member-year and the detector catches 40% at 70% collection, the return is $70 — below the price. Clearing a 2x bar requires $288 recovered, which means the detector must reach roughly 60% detection at 70% collection on a $700 recoverable base. Nobody knows those three parameters today. That is exactly the number the evaluation is designed to produce, and if it comes in low the correct response is to cut the price to about $5, not to restate the value.
BY FUNCTION, WITH THE HONEST VERDICT ON EACH. (1) RECORDS AGGREGATION — Epic MyChart holds the records, is free, is already activated at enormous scale and now links across organizations; Apple Health Records aggregates hundreds of systems on a device the daughter already owns; b.well, 1upHealth, Health Gorilla, Particle, Metriport and PicnicHealth do the plumbing better than a startup will. VERDICT: Ardia cannot win here and should not try — buy an aggregator and compete on what is done with the data. (2) MEDICAL BILL DEFENSE — Goodbill, Resolve Medical Bills, CoPatient, Claimable and the open-source Counterforce Health on the consumer side; Health Advocate, Quantum Health, Rightway, Alight and Accolade on the employer side; plus every hospital's own financial-assistance desk. VERDICT: the most crowded lane and simultaneously Ardia's best demo. The available differentiation is bundling — nobody else puts bill defense inside the same hub as the medication list and a caregiver consent model, aimed specifically at the Medicare Advantage member. (3) ELDER-CARE COORDINATION AND CAREGIVER SUPPORT — Papa (companionship at scale, sold into MA supplemental benefits, and a cautionary tale on safety incidents), Wellthy, Homethrive, Cariloop, ianacare, CircleOf, Carely, Lotsa Helping Hands. VERDICT: these are the true competitors for the buyer Ardia should target; they are ahead on distribution and behind on AI governance. (4) VOICE COMPANIONS — ElliQ from Intuition Robotics is the benchmark: a purpose-built device, thousands of units deployed with a state aging office, published engagement data and years of elder-specific interaction design; also Sensi.AI, care.coach, and the discontinued Amazon Alexa Together. VERDICT: ElliQ is ahead of Aria on everything except governance transparency, and it solved a distribution problem Ardia has not acknowledged — a browser tab is not a companion. (5) SAFETY AND DETERIORATION SENSING — Apple Watch fall detection at consumer scale, Lifeline/Connect America, Medical Guardian, Bay Alarm, ambient radar from Vayyar Care and Cherish Health, SafelyYou, CarePredict. VERDICT: Cadence is not competitive with any of these and must not be positioned as safety sensing. (6) SYMPTOM AND REPORT EXPLANATION — Ada, K Health, Buoy, Infermedica, and the elephant: ChatGPT, Gemini and Claude, free on the daughter's phone, with larger models and no 6,000-character cap. VERDICT: "AI explains your lab report" is a commodity and cannot be the product. HONEST DIFFERENTIATION, stated narrowly: the governed dyad — one consent-scoped hub where the older adult owns the data, the caregiver sees only what she is granted, and every answer carries a machine-readable gate trace and a cite-or-abstain contract. No competitor ships a returned per-answer gate trace. That is a real but narrow wedge, and it is worth nothing until the gates hold, which today they demonstrably do not.
THE ONE BLOCKER FIRST, and it is not a BAA and not a customer: the REGULATORY IDENTITY DECISION. Is Ardia One a HIPAA business associate operating under a covered entity, or a direct-to-consumer personal health record vendor? That single choice determines whether a BAA is obtainable, whether the FTC Health Breach Notification Rule and Washington's My Health My Data Act govern instead of HIPAA, whether the Symptom Helper needs a General Wellness argument or an HCP-facing rewrite, which buyer's org chart to sell into, and which integration gets built first. It costs one focused engagement with FDA and health-privacy counsel and it is blocking the design of everything else. Weeks, not quarters. THEN, IN ORDER. (0) IMMEDIATE HYGIENE, this week, zero cost: delete the "Crucible (8 safety gates)" string; correct the MA appeal deadline to the 60-day plan reconsideration; remove the inert MFA copy; remove Cadence from the Vitals early-warning card; correct "Claude/Gemini" to Gemini everywhere. A reader who finds one of these finds all of them. (1) PIN THE MODEL VERSION and stand up a minimal eval harness — a fixed prompt set run on every deploy across all four engine paths, with diffs surfaced. Without this, one shared path silently degrades three products. (2) CLOSE THE NAME LEAK: pass a roster where one is known and add a real name-detection layer (Presidio, scispaCy or a small NER model) behind a flag, with a measured precision and recall figure. This is the highest-value engineering fix in the company, because de-identification is the foundation of every other claim and of any signable BAA. (3) BUILD THE DETERMINISTIC BILLING ENGINE FOR REAL — NCCI PTP and MUE tables, PFS/CLFS/OPPS fee schedules, CARC/RARC lists, an 835 parser, and the four rules the demo describes in prose. All inputs are public CMS files: zero agreements, zero PHI, zero permission. Do this before anything clinical. (4) MIGRATE TO VERTEX AI UNDER A BAA, US regions only, with Cloud Audit Logging so the audit-log claim becomes true. (5) GET ONE LDS PLUS DUA and publish precision, recall and dollars per 1,000 lines — the company's first defensible number that is not a public-dataset activity classifier. (6) BUILD ONE REAL DATA CHANNEL: SMART on FHIR against Epic, Blue Button 2.0, or an aggregator contract. Pick one; the hub's premise is dead until records arrive. (7) RED-TEAM ARIA on 1,000 synthetic elder emergency utterances and publish the false-negative rate. (8) FIX DISTRIBUTION — browser tab, phone number or device; decide. (9) ONE DESIGN PARTNER: a risk-bearing provider group or PACE organization, 50-150 dyads. (10) ONLY THEN Cadence elder retraining. WHAT TO STOP: adding views. Every new mock-up widens the gap a diligence reader measures.
