River Records

Natural language processing is the whole product, and it does more work here than in a conventional scribe: the system must not only generate a note but classify content by medical problem and maintain those problem threads across encounters, which is a harder extraction task than producing a SOAP note from one visit.

Conventional draft and review, with one structural property: because output pastes into the record rather than writing to it, the clinician necessarily handles every note before it enters the chart. The shallow integration therefore produces a firm human gate, which is an accidental benefit rather than a designed one but a real defense against the automation bias that grows when notes file themselves. No acceptance rate, edit burden figure or confidence threshold is published, and problem thread assignment happens without any described review step of its own.

The architectural thesis is stated clearly, problem oriented rather than encounter oriented organization, and the reasoning behind it is published. What is not published is any measurement: no accuracy figure, no model card, no named models, no evaluation of how reliably content is assigned to the correct problem thread, which is the specific failure mode this design introduces. Misfiling a finding under the wrong problem in a longitudinal chart is a different and more persistent error than a wording mistake in a single note.

Nothing identifies any party in the chain: no model or model family, no foundation model provider, no hosting arrangement and no subprocessor list is published. The business associate agreement names none of them either, though it does contemplate agents and subcontractors receiving protected health information and requires assurances of substantially similar conditions from them. So parties exist and are contractually addressed; who they are is not disclosed.

The product keeps what it captures, and that changes what the silence covers. Fast discard is not available as a privacy posture here, because the product is the memory: persistent context across visits, a clinical thread maintained over years, and outside records converted into structured context so material from other institutions accumulates in the store as well. Whatever unnamed parties sit in this chain are holding a growing longitudinal object rather than processing a transient one, and the usual reassurance that audio is gone in minutes is unavailable by design.

The published terms widen the question rather than narrowing it. Uploaded content carries a perpetual, irrevocable, transferable and sublicensable license to the vendor for any purpose, subject to the privacy policy and applicable law. The business associate agreement permits use of protected health information for de identification and for data aggregation in connection with the practice's health care operations, and settles in advance that the information will be retained rather than returned or destroyed when the relationship ends. None of that names a recipient, and the sublicensing right means recipients need not be named.

The community health center deployment enlarges it further, placing medical, behavioral health and social care content in one record with multiple disciplines contributing, so social circumstances and mental health material persist together across years in a single structure. That is a more sensitive object than a note, and nothing says who can reach it.

The vendor's strongest asset is evidence about the problem, not about the product.

The founders published peer reviewed research in JAMA Network Open in 2022, analyzing 104 million clinical notes and finding that around half the text in an average record is duplicated verbatim from prior documentation. That is real evidence, rigorously establishing the problem the product addresses.

It is not evidence that the product solves it. No evaluation of the product itself, no controlled comparison, no accuracy result and no third party assessment is published. What remains is named clinician testimonials.

One published term bears on the thinness of independent assessment. The terms of service prohibit users, without prior written consent, from publishing or disseminating benchmarking or test results, or any comparison of the platform with alternative services. The absence of independent comparative results is therefore partly a contractual position rather than only a gap in the literature.

The vendor publishes no retention schedule, deidentification practice or training use statement, and asserts health privacy compliance.

This matters more here, and the reason is structural rather than incidental. This product inverts the assumption that scribe privacy usually rests on. Ambient scribes generally treat minimal retention as the privacy posture, competing on how fast audio is discarded, some within seconds or hours. This vendor cannot do that, because the product is the memory. Its value proposition is persistent context that keeps notes coherent across visits, documentation organized around a patient's active problems rather than around encounters, and a clinical thread maintained over years.

So the privacy question changes shape. It is not how quickly content is deleted, because deletion defeats the product. It is what is held, for how long, who can reach it, what happens when a practice leaves, and whether a patient can have a thread removed. None of that is published, and a buyer cannot substitute the reassurance a transient product offers.

Two things enlarge the store beyond encounter content. The vendor describes converting outside records into structured context, so material from other institutions accumulates inside it. And the community health center deployment places medical, behavioral health and social care in one longitudinal record with multiple disciplines contributing, which means social circumstances and mental health content persist together across years in a single structure.

That is a more sensitive object than a note. Nothing published sets out the retention model, the deletion path, the export format on exit, or whether anything in that store is used for improvement.

The business associate agreement is published in full at its own address, linked from the site footer alongside the terms of service, privacy policy and a security page, and readable without an account or a request.

The agreement covers status, availability, tier and subcontractors. Business associate status is stated, with River Records, Inc. named as Business Associate. The agreement is available in full. The tier is universal and automatic: the agreement is incorporated into the terms of service by reference and binds on acceptance, so a clinician starting a self serve trial is covered without a negotiation. And subcontractors are addressed, with the vendor required to obtain assurances that any agent or subcontractor receiving protected health information adheres to substantially similar restrictions.

Two qualifications apply, and the second is the more consequential.

The subcontractor clause sits at the regulatory floor rather than above it. It requires reasonable assurances of substantially similar conditions rather than a written agreement to the same conditions.

The agreement settles in advance that protected health information will not be returned or destroyed. One clause states the standard position, that on termination the vendor returns or destroys all protected health information and retains no copies, with infeasibility as an exception requiring notice. A later clause then states that the parties acknowledge and agree in advance that, except in limited circumstances, return or destruction will not be feasible and the vendor will need to retain the information, governed by the terms and privacy policy between the vendor and its consumer users. A practice signing this is agreeing at the outset that the exception applies, and that what happens to the retained record later is set by an agreement it is not party to. For a product whose value is a longitudinal store accumulated over years, that is the most consequential term in the document.

Breach reporting runs to thirty calendar days for a written report, where the covered entity's own clock is sixty days from discovery, and time spent responding to a Secretary compliance request is chargeable back to the practice at the vendor's hourly rate.

There is no named or dated attestation and no trust center.

This is a very small company, an attestation program is a material cost at that size, and the gap reflects what a counterparty can verify rather than a judgment that controls are absent. A health system review would still stop here, and that is the practical consequence.

One argument cuts the other way, and it is specific to this product rather than generic. Ambient scribes typically hold clinical content briefly. This one holds it permanently by design, structured, indexed and organized around each patient's problems across years, with outside records converted into the same structure. A persistent longitudinal clinical store is a materially more attractive target than a pipeline that discards audio after producing a note, and it is also harder to secure, because the exposure is continuous rather than momentary.

So the usual proportionality argument, that a small vendor handling transient data can reasonably defer certification, applies less well here. The data is not transient. The company has five people, and the scope of what an examination would cover is still large.

The deployment setting adds to it. Community health centers hold behavioral health and social care content alongside medical records, and the vendor's own design puts all of it in one structure.

No clearance claimed and none required for documentation. The product leans toward decision support in two directions.

The first is care gap identification. Surfacing problems that are falling off a clinician's radar sits closer to decision support than to documentation. The product's stated purpose is that persistent context makes patterns visible and keeps threads intact across visits, and that is a real clinical benefit. It is also a system deciding what a clinician should be reminded of, which is a judgment about relevance rather than a record of what happened. Nothing says what determines that a problem needs surfacing, or whether anything is ever suppressed.

The second is new and prospective. The product already generates diagnosis and procedure code suggestions, and the vendor states hierarchical condition category support is coming. That is the point at which the product meets the risk adjustment coding question, and the longitudinal design makes it distinctive. Inferring risk adjusted conditions from a single encounter is one thing; a product holding structured problems across years is positioned to surface conditions for annual recapture that were documented in a previous visit, which is exactly the mechanism a published policy analysis identified as driving measured increases in documented risk adjusted diagnoses.

That capability could equally support better continuity. Nothing yet says what evidence a condition is resurfaced on.

The vendor publishes comparison guides stating plainly where a competitor is the better choice.

No fairness statement, subgroup analysis or accent and dialect performance disclosure is published.

One point counts in the vendor's favor. No language coverage is claimed, so the absence is at least consistent with the product's stated scope rather than contradicting a marketing claim. A vendor that does not promise multilingual performance is not concealing evidence about a promise it made.

Still, the question applies, and the reason comes from where the product is deployed rather than from what the vendor claims. The vendor markets specifically to community health centers, describing medical, behavioral and social care delivered under one roof with multiple disciplines contributing to a single record. Those centers exist to serve populations who face barriers to care, and their patients include a disproportionate share of people whose first language is not English, alongside high rates of behavioral health need and complex social circumstance.

So in the setting the vendor has chosen to specialize in, speech recognition performance across accents and languages carries extra weight, whatever the product claims. That is the same argument that applies to the free scribes reaching underserved clinics, arriving here through deliberate market focus rather than scale.

The longitudinal design compounds it. Where a scribe produces a note, a recognition error is contained in that note. Where a product carries structured problems forward across visits, an error entered once can persist and be resurfaced at every subsequent encounter.

The published terms say a great deal about what happens when the system is wrong, and almost all of it runs away from the vendor.

Nothing published gives an accuracy or error figure, a limitation, or any warranty, indemnity or remediation commitment. Accuracy is expressly disclaimed, the service is provided as is, and aggregate liability for any claim arising from use of the platform is capped at $1,000. Indemnification runs one way, from the practice to the vendor. Disputes go to binding arbitration with no class actions, and no claim may be brought more than one year after it accrues.

The correction mechanism is real but reaches only one party. The business associate agreement obliges the vendor to amend records it maintains within fifteen business days of a request from the practice, and to mitigate known harmful effects of an improper use or disclosure. It also states that the vendor is not obliged to provide anything directly to an individual, and the agreement confers no rights on anyone who is not a party. So a mechanism exists, nothing stands behind the output, and the patient has no route.

The design introduces a failure mode specific to it, which raises the stakes of these terms. Documentation is organized around a patient's active problems rather than around encounters, and content is assigned to threads that persist for years. Misfiling a finding under the wrong problem is a durable error rather than a wording mistake: it does not sit in one document awaiting review, it becomes part of a structure later clinicians read as established history, and every subsequent note inherits it. Nothing published describes how reliably assignment works, what a clinician sees about why content landed where it did, or how a misassignment is found and unwound once it has propagated.

The one year limitation sits against that: an error of this kind is latent by construction, and may surface long after the year has run.

Integration is deliberately minimal and the vendor describes it plainly rather than dressing it up: output pastes into any electronic record system. The plainness helps: pastes into any record system is a checkable description where works with all systems is not, and for a small independent practice it means there is nothing to implement and nothing to wait for.

What it is, though, is copy and paste. No write back, no structured field population, and no chart context flowing in from the record system. The consequence is structural rather than inconvenient: the problem oriented chart this product builds lives in the product, not in the record system that is the legal record. The practice maintains a parallel chart, and the more valuable the accumulated thread becomes, the more of the practice's working picture sits outside the system of record.

That compounds with the exit position set out in the vendor's agreements, where the accumulated store is retained rather than returned on termination.

No hosting region, residency option or subprocessor information is published, and nothing establishes which model service processes the encounter or what it retains.

The gap has a different weight here than for a transient product. Where a vendor discards audio within seconds, residency governs a brief pipeline and the practical exposure is small. This vendor operates a persistent longitudinal store by design, structured around each patient's problems and accumulating across years, with outside records converted into the same structure.

So residency is not a question about where a conversation was transcribed. It is a question about where a growing clinical database physically sits, under whose jurisdiction, and for how long. That is closer to the question asked of a record system than of a scribe, which fits, because the product is closer to a record system than to a scribe.

Two consequences follow. Persistent structured storage makes the export question real rather than theoretical: nothing says what format the accumulated problem oriented record comes out in on exit, or whether it can be moved into a record system or is only readable inside this product. And a persistent store means subprocessor retention compounds rather than resets, so if an external model service holds anything, it holds it against an accumulating corpus rather than a single note.

Fully published and self-serve. A 30 day free trial requiring no credit card, then $149 per month, or $99 per month on annual billing, all stated on the front page rather than behind a demo request. For an independent clinician that is the entire commercial process: no quote, no procurement, no implementation and no sales conversation.

There are two rates, and the annual option is a third off the monthly one, which a buyer sizing an annual budget needs.

The unit of charge is per month rather than per clinician seat as such, and the terms describe accounts as individual, so it is not stated how multiple clinicians are billed or whether a single subscription covers them.

One term sits behind the price and belongs in a commercial assessment. The agreements pre-agree that the accumulated record is retained rather than returned when the relationship ends, so the exit cost of this product is not the subscription but the store built up inside it. A price that is easy to start is not the same as a commitment that is easy to leave.

Narrow and deliberately so. Named settings are primary care, pediatrics, direct primary care and skilled nursing, which share the property that matters to this product: the same clinician follows the same patient over years, so a chart organized by problem compounds in value while an encounter scribe does not. Skilled nursing and direct primary care are both underserved by ambient documentation tools, which makes the narrowness useful rather than merely limiting. No specialty count, no language coverage and no acute or emergency capability.