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 an unusual structural property worth naming: because output pastes into the record rather than writing to it, the clinician necessarily handles every note before it enters the chart. The shallowest integration in this category therefore produces one of its firmest human gates, which is an accidental benefit rather than a designed one but a real defence against the automation bias that grows when notes file themselves. Held at B because no acceptance rate, edit burden figure or confidence threshold is published, and because problem thread assignment happens without any described review step of its own.

The architectural thesis is stated clearly, problem oriented rather than encounter oriented organisation, 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 sub processor list was located in two passes. This product inverts the assumption the rest of the category rests on, and that changes what the silence covers. Most peers compete on how fast content is discarded and treat minimal retention as the privacy posture.

This vendor cannot, 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 inside the store as well. So 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 community health centre deployment enlarges it further, placing medical, behavioural 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 a buyer should know who can reach it. Ask for a sub processor list, the hosting arrangement, and whether anything in the accumulated store is used for model improvement.

An important distinction the record should not blur. The founders published genuine peer reviewed research in JAMA Network Open in 2022, analysing 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, and it is a stronger intellectual foundation than most vendors in this category have.

But it is not evidence that Stream solves it. No evaluation of the product itself, no controlled comparison, no accuracy result and no third party assessment was located, and what remains is named clinician testimonials. Graded on evidence for the product, not for the thesis.

No retention schedule, de identification practice or training use statement was located, and health privacy compliance is asserted.

The earlier assessment identified why this matters more here, and the second pass confirms it is structural rather than incidental. This product inverts the assumption most of this category rests on. Nearly every peer treats minimal retention as the privacy posture, competing on how fast audio is discarded, and several delete 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 organised around a patient's active problems rather than around encounters, and a clinical thread maintained over years.

So the question this axis asks 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 centre deployment places medical, behavioural 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, and it is the thing to ask about. Establish the retention model, the deletion path, the export format on exit, and whether anything in that store is used for improvement.

HIPAA compliance is stated directly, on the same basis other vendors in this category are graded B. Note the buying path: this is self serve with a free trial requiring no credit card, so a clinician can begin using it before any business associate agreement conversation occurs, and the terms are not published for inspection.

No named or dated attestation and no trust centre were located.

The fairness point from the earlier assessment stands. This is a very small company, an attestation programme is a material cost at that size, and the grade records what a counterparty can verify rather than a judgement 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 belongs on the record, because it is specific to this product rather than generic. Most scribes in this category hold clinical content briefly. This one holds it permanently by design, structured, indexed and organised 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 scope of what an examination would cover is larger for this product at five people than for many larger vendors holding nothing.

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

Ask what external testing has been performed, whether a penetration test exists, and what the roadmap to an attestation looks like, since a buyer can reasonably accept a credible plan from a company this size.

No clearance claimed and none required for documentation. The thought raised in the earlier assessment stands and the second pass sharpens it in two directions.

The first is the one already identified. Surfacing problems that are falling off a clinician's radar is care gap identification, which 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 judgement about relevance rather than a record of what happened. Ask what determines that a problem needs surfacing, and 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 this record joins the coding gradient this index tracks across the category, and the longitudinal design makes it distinctive. Most peers infer risk adjusted conditions from a single encounter. 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 genuinely better continuity, which is the honest position. But it should be assessed when it ships rather than assumed benign, and a buyer should ask what evidence a condition is resurfaced on.

One candour credit belongs here. The vendor publishes comparison guides stating plainly where a competitor is the better choice, which is rare in this category.

No fairness statement, subgroup analysis or accent and dialect performance disclosure was located.

The fairness point from the earlier assessment stands and is worth keeping. No language coverage is claimed, so the absence is at least consistent with the product's stated scope rather than contradicting a marketing claim, which is more than can be said for several larger vendors in this category. A vendor that does not promise multilingual performance is not concealing evidence about a promise it made.

The second pass finds a reason the question still applies, and it comes from where the product is deployed rather than from what the vendor claims. The vendor markets specifically to community health centres, describing medical, behavioural and social care delivered under one roof with multiple disciplines contributing to a single record. Those centres 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 behavioural health need and complex social circumstance.

So the setting the vendor has chosen to specialise in is one where speech recognition performance across accents and languages matters more than average, not less, 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 in a way worth naming. 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.

Ask what evaluation has been done across the populations these centres actually serve.

Two passes located no accuracy or error figure, no published limitations and no warranty, indemnity or remediation commitment. The design introduces a failure mode that is specific to it and that nothing published measures, which is why the absence matters more here than for a conventional scribe. Documentation is organised around a patient's active problems rather than around encounters, and content is assigned to problem threads that persist for years.

Misfiling a finding under the wrong problem is therefore a different and more durable error than a wording mistake in a single note: it does not sit in one document to be corrected on review, it becomes part of a longitudinal structure that later clinicians read as established history, and every subsequent note inherits it.

Nothing is published about 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. Two recourse questions follow from the same architecture and are equally unaddressed: what happens to the accumulated store when a practice leaves, and whether a patient can have a thread removed. Ask for an assignment accuracy figure, the correction path for a misfiled finding, and the exit and deletion positions in contract language.

The shallowest integration model in this category, and the vendor says so plainly rather than dressing it up: output pastes into any electronic record system. Credit the honesty, because pastes into any EHR is a materially more truthful phrase than the works with all EHRs claims made elsewhere in this category, and for a small independent practice it means there is nothing to implement.

But it is copy and paste, with no write back, no structured field population and no chart context flowing in, which also means the problem oriented chart Stream builds lives in Stream rather than in the record system of legal record. A buyer should understand they are maintaining a parallel chart.

No hosting region, residency option or subprocessor information was located, 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, and the reason is the same one that shapes the stewardship axis. 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 a buyer would ask of a record system than of a scribe, which is appropriate, because the product is closer to a record system than most things in this category.

Two consequences follow that a buyer should raise directly. Persistent structured storage means the export question is real rather than theoretical: establish what format the accumulated problem oriented record comes out in on exit, and 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.

Ask for the hosting region, the subprocessor list, the model provider, the export format, and what happens to the accumulated record at contract termination.

Fully published and self serve: a 30 day free trial requiring no credit card, then 149 US dollars per month, stated on the front page rather than behind a demo request. For an independent clinician this is the entire commercial process, with no quote, no procurement and no implementation. In a category where enterprise vendors publish nothing, a small company publishing one clear number is doing exactly what this axis rewards.

Narrow and deliberately so. Named settings are primary care, paediatrics, 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 organised by problem compounds in value while an encounter scribe does not. Skilled nursing and direct primary care are both underserved by this category, which makes the narrowness useful rather than merely limiting. No specialty count, no language coverage and no acute or emergency capability.