Pymander Health

Nothing is left to sell once the models are gone. A person texts a question and an AI answers it, asks the one follow up question that best separates the remaining possibilities, and says what level of care the person needs, with no clinician in the conversation today. The cash price catalog is a lookup of other companies' prices, and clinician visits, prescriptions and labs are announced rather than live.

What surrounds the models is engineering in service of the AI rather than a second product. Off the shelf frontier models from an outside provider propose facts, possible diagnoses and wording, and a deterministic pipeline of ordinary program logic accepts, rejects or corrects each proposal before a message goes out. The company sums the design up as the model proposes and the pipeline disposes. That split is the core of the offering, and it is also what makes the safety behavior something a reviewer can inspect stage by stage.

Every answer reaches the patient without a person reviewing it, and the oversight lives in software. Program logic routes each message, a separate pass sets care urgency apart from the reasoning that produced the list of possible diagnoses, and an audit stack checks tier routing, emergency language, pediatric scope and repeated questions before a draft is sent, falling back to a fixed, pre approved message for that tier when redrafting fails. Diagnoses, prescriptions and exams are reserved for a licensed clinician, and the product tells people to get emergency care when something sounds urgent.

The weak point is a claim the controls do not back. The safety page says every answer is based on trusted medical sources, yet nothing in the described pipeline retrieves a source or shows one to the patient, and the configuration behind the headline exam score runs without retrieval. No clinician can override a disposition, and the criteria for each care tier are not given.

Off the shelf frontier models from an outside provider do the language work, used without fine tuning: a stronger tier for reasoning and drafting and a faster tier for pulling facts out of messages. Everything around them is described stage by stage, from chart assembly and quote verified extraction through routing, a danger ranked differential, the falsification pass, the urgency vote, tightly framed drafting and the audit stack, with each stage's contract defined over text so the models can be swapped.

The provider and model names are not given, nor are versions or a schedule for model changes, although the company notes that results have to be measured again with each model generation. A cost and accuracy comparison of three setups on the USMLE questions shows the shipped exam configuration as a single model without retrieval, ahead of a cheaper two model cascade and a retrieval setup that cost more and scored lower.

Three kinds of outside party touch consult content: hosting, message delivery and the AI infrastructure behind the AI doctor. Each works under contracts requiring it to protect the data, use it only to serve Pymander and never train its own models on it.

None of them is named, the model provider included; it appears only as an outside provider of off the shelf frontier models in two tiers. Text messages also pass through third party delivery providers and mobile carriers before reaching Pymander's systems, and licensed providers would join the chain once clinical services launch.

Without a subprocessor list, a person cannot see which companies hold their consult history, and a reviewer cannot read the model provider's own retention terms against the training restriction Pymander describes.

Benchmark results arrive with their method, denominators and run data. On 244 text only questions from the official USMLE sample exams, each answered once by a single model without retrieval in two independent runs, the shipped configuration got 479 of 488 right, or 98.2 percent. A companion working paper reports 98.9 percent (93 of 94) on a held out set sealed before any tuning.

A sealed, pre registered HealthBench Hard run over 1,000 open ended clinical conversations scored a mean of 0.3331 (N=999), with protocol, generations and judge grades deposited on Zenodo so anyone can rerun it; the company places that result in the top five of published HealthBench Hard results. An internal suite of 300 synthetic red flag cases backs the safety work.

All of it is company run, the working papers are not peer reviewed, and the company itself says an exam score does not carry over to open ended care. No result from real patients, no clinician blinded comparison and no outcome measure exists yet.

Identifiable health information is not used to train generalized, foundation or third party models, and the outside companies that host data, deliver messages and run the AI are bound by contract never to train their own models on it. Improvements to Pymander's own features draw on aggregated, de identified data, and any program wanting identifiable data needs a separate, explicit consent.

Retention runs for the life of the account plus legal, medical record and tax obligations, after which data is deleted or de identified; no fixed window is given. Plain SMS is not encrypted in transit, a point the policy makes outright while offering the app for sensitive topics.

The safety engineering is laid out in unusual depth: quote verified fact extraction, a pass that tries to break the leading diagnosis, an independent vote on care urgency, and a post draft audit with a fixed safe fallback. How a safety event in live use is reported and handled is not described.

The position is plain: Pymander calls itself a medical AI system rather than a licensed provider, and says that what people share with it today is not created or held under HIPAA provider arrangements. The HIPAA aligned label on the homepage means safeguards modeled on the rule, not coverage under it.

For Washington, Nevada and other states with consumer health data laws, the privacy policy doubles as the consumer health data notice, naming the categories collected, the purposes, the recipients and the rights. It commits to no sale of personal information, no advertising use of health data and no use in insurance or underwriting models, and gives people access, export and deletion rights.

Records held by licensed providers, once clinical services launch, would fall under those providers' own obligations. No business associate agreement exists, since no covered entity buys the service.

No security certification is claimed: no SOC 2 report, no HITRUST, no ISO 27001 and no trust center. The controls listed in the privacy policy are TLS 1.3 in transit, AES 256 at rest, access limited to authorized staff, audit logging and regular security testing, with breach notification as the law requires.

Who performs the testing, how often and against what scope are not given, so nothing independent stands behind the list. The safety architecture is described in far more depth than the security program, and that gap matters because consult text passes through outside hosting, messaging and model providers on every turn. Plain SMS adds a carrier link outside the company's control, which it acknowledges, pointing people to the app for sensitive topics.

The terms say Pymander does not practice medicine, gives no medical advice, diagnosis or treatment, and that nothing it produces is a diagnosis or a clinical instruction. The product as marketed goes further. Its safety working paper describes a list of possible diagnoses ranked by danger, a pass that tests the leading diagnosis before a conclusion is shared, and an independent vote that sets how urgently the person needs care, all delivered straight to the patient by text. The homepage sells it as an AI doctor that reasons through symptoms and says when someone needs a clinician.

No FDA position on the software is stated beyond a general line that statements about the service have not been evaluated by the agency. The company has drafted state legislation, the Clinical AI Provider Licensing and Patient Safety Act, for step by step licensing of AI doctors in New Hampshire's 2027 session. That shows engagement with how such products should be governed, and it is not a regulatory position on this one.

Changes to the pipeline ship through a release gate. A 60 case sample from a 300 case suite of synthetic red flag presentations has to earn a ship verdict, and categories that flipped between runs are retested with repeated sampling before they count as fixed. Results come by category: three flip prone presentations went from 8 of 15 correct escalations to 15 of 15 after the independent urgency vote was added, and a 40 case probe built to tempt false reassurance found no under escalation. The company names its remaining failure, over escalation, and spells out its costs in emergency department load, anxiety and lost trust.

Nothing is broken out by age, sex, race, language or any other patient group, children are covered by scope rules rather than tested on their own, and no outside audit exists. The results are also tied to the specific models in use and need rerunning whenever the model generation changes.

Error rates for the triage behavior come with method and denominators. On three presentations that had flipped between runs, rabies risk after a bat exposure, an airway burn and a leg clot after childbirth, correct escalation rose from 8 of 15 sampled runs to 15 of 15 once the independent urgency vote was added. A 40 case probe built to provoke false reassurance found no under escalation, and the remaining error is over escalation, sending people to a higher level of care than they need. On the USMLE sample questions the shipped configuration missed 9 of 488.

Every turn also leaves a SOAP note and a full decision trace that can be replayed without rerunning a model. Nothing stands behind an individual answer, though: the terms say the service can be wrong, guarantee no accuracy, accept no liability for harm from relying on it, cap damages and send disputes to individual arbitration, with a 30 day window to opt out.

Connections run to the consumer's own data rather than to a provider's record system. With an account, a person can link medical records, lab results and wearables. Whoop, Oura Ring and Garmin are named, Apple Health is coming, and wearable data syncs only with consent and is not watched live.

How medical records are pulled in, through which network or standard, is not stated, and no hospital or clinic record system is named. Nothing flows out to a clinician's EHR either. Instead every consult lands in care notes, written as a structured SOAP visit note, which a person can open through a signed link without an account and pass to their own physician.

For a practice or health system there is nothing to integrate, because the product is sold only to consumers and has no enterprise connection or API offering.

The service runs as a single hosted product reached over SMS, iMessage and an app, with data encrypted in transit with TLS 1.3 and at rest with AES 256. Hosting, message delivery and AI processing sit with outside service providers described by role rather than by name, and no data center, cloud region or country of storage is committed.

The audience, emergency numbers and legal terms are American, with Delaware law governing the terms, which implies US processing without promising it. Model calls go to an outside provider, so a person's message text leaves Pymander's own systems on every turn of a consult.

People can delete their data subject to legal retention duties. There is no private, regional or single tenant deployment, since nothing is sold to organizations.

Texting the AI doctor costs nothing. A first consult starts from a phone's messaging app with a prefilled text, needs no account and runs at any hour; standard carrier message rates are the only charge mentioned. An account is optional and adds connected records, labs and wearables.

There is no paid tier yet, because the clinician services that would carry a fee, diagnosis, prescriptions and lab orders, have not launched. The cash price catalog lists 886 medications from the Cost Plus Drugs price list, 124 lab tests and 26 imaging studies at published source rates or Medicare benchmarks, labeled as reference prices rather than local quotes, and the company does not bill insurance. Nothing is sold to health plans, employers or health systems, so there is no enterprise contract to price and no sales conversation stands between a person and the product.

Adult primary care questions are the core. The topic menu runs from fever, cough, sore throat, sinus and ear pain through stomach complaints, urinary and vaginal symptoms, back and joint pain, sprains, rashes and moles, shortness of breath and heart rhythm worries, blood pressure and blood sugar, to low mood, anxiety, sleep and focus.

The channel is SMS and iMessage in the United States, with emergency instructions pointing to 911 and the 988 crisis line. Children are handled under explicit scope rules in the pipeline rather than by a pediatric trained model, and the service does not examine anyone or take on problems that need physical findings.

Validation covers part of that range: licensing exam questions, physician graded conversations, and a red flag suite built around presentations that are dangerous to miss, such as a bat exposure, an airway burn or a leg clot after childbirth. No language coverage is stated.