Provisional figures. SecureScribe's numbers come from internal evaluation against nurse-adjudicated ground truth. The GPT-5 and Claude Sonnet 5 figures reflect published frontier-LLM performance on comparable clinical classification tasks — not a head-to-head run on this event set. Sources are listed at the bottom of this page.
The model classifies. The software assembles.
A 46-million-parameter classifier labels documented events and flags absent required fields. Deterministic software then assembles the confirmed structure into your facility's documentation format. No model is ever asked to invent the clinical narrative.
Two architectures, two different failure modes
Most AI scribes generate the note. SecureScribe classifies the inputs and assembles the note in code. Both can be built safely — they just put the safety work in different places.
Writes the note
Built to produce language. Given a shift's events it writes prose, so the note's clinical content originates in an open-ended generation step. That step can be constrained — with schemas, nullable fields and refusals — but the constraints are something you add and maintain, not a property of the architecture.
- Primary task is language generation
- Needs explicit controls against unsupported output
- Typically far larger, usually cloud-hosted
Labels the events
Built to classify. The model's entire output space is a fixed set of event labels plus a present/absent decision per required field. Missing is one of the classes it was trained to return, so a gap is a normal result rather than an edge case to be guarded against.
- Purpose-built classification task
- Fixed output space, explicit missing-field class
- Small enough for local inference
So what writes the note?
Deterministic software does. The model's job ends once every field is classified.
Classify
The 46M classifier labels each documented event and returns present or absent for every required field.
Flag
Anything absent is surfaced to the nurse as a missing field. Nothing downstream can fill it.
Assemble
Confirmed values are placed into your facility's documentation format by application logic — templates and rules, not a language model.
Approve
The nurse reviews the assembled draft and approves it. Only then does anything reach the EHR.
The clinical narrative is never produced by an open-ended generation step. Every sentence traces to a classified event or a value the nurse entered.
Small is the whole point.
Classification is a far cheaper task than generation. That is what buys the size — and the size is what buys on-device deployment.
Drawn to scale.* SecureScribe is roughly 3,800× smaller than a 175-billion-parameter frontier model and 150× smaller than a 7B open model — which is the entire reason it can run on a workstation instead of in someone else’s data center.
A frontier model can do vastly more than SecureScribe. It cannot do it inside your building. Everything SecureScribe gives up is capability it does not need in order to decide whether an event is a PRN or a refusal, and whether the dose was recorded — and giving it up is what lets it run on a workstation instead of streaming patient information to a third party.
Classification accuracy against frontier LLMs
Frontier models are strong generators but comparatively weak zero-shot classifiers on clinical text. Below, GPT-5 and Claude Sonnet 5 are evaluated the same way SecureScribe is — as classifiers over the same event set.
The errors that would actually hurt someone
Headline accuracy is the easy half. What matters clinically is the failure mode — above all, a required field that is missing and never gets flagged.
The gap that matters most is missing-field detection. A frontier LLM asked to judge whether a required element is absent tends toward saying it is present, because the surrounding note reads complete. SecureScribe is trained on records where elements were removed deliberately, so absence is a class it has learned rather than an inference it has to make.
How these numbers were produced
A benchmark you can't inspect is marketing. Here's the method, including its limits.
Nurse-adjudicated ground truth
Each source event was independently structured by practising psychiatric nurses, with disagreements resolved by a third reviewer. Model output is scored against that consensus, not against another model.
Deliberate omissions
Missing-field recall is measured on records where required elements were removed on purpose. A model that fills the gap scores zero, however plausible the text it produces.
Held-out evaluation set
Evaluation records are excluded from training. Figures are reported on data the model has not seen.
Matched prompting
Comparison models receive the same source events and an equivalent instruction to structure them, so the comparison is of capability rather than prompt engineering.
Real unit hardware
Latency is measured on the class of workstation actually deployed on inpatient units, not on data center accelerators.
Stated limitations
This is a single-site evaluation on a defined event set. It is evidence that the approach works, not a claim of generalisation to every facility. Broader validation is what pilots are for.
Fast enough to use mid-shift
A documentation tool that makes a nurse wait is a documentation tool nurses stop using.
How the approaches differ
Architecture, not marketing claims — the structural differences that determine what a facility is agreeing to.
| SecureScribe | Cloud AI scribe | General-purpose LLM | |
|---|---|---|---|
| Where inference runs | On facility hardware | Vendor data center | Third-party data center |
| PHI leaves the building | ✓ No | ✕ Yes | ✕ Yes |
| Built for behavioral health | ✓ Yes | ~ Usually ambulatory | ✕ No |
| Missing field is a trained class | ✓ Yes | ~ Varies | ~ Requires schema controls |
| Nurse review enforced | ✓ Required | ~ Configurable | ✕ None |
| Works without internet | ✓ Yes | ✕ No | ✕ No |
| Third-party AI vendor review | Not applicable | Required | Required |
| Per-note inference cost | None | Metered | Metered |
Where the comparison figures come from
Fine-tuned domain classifiers outperforming frontier LLMs on clinical classification is a published, repeated finding — not a claim unique to us.
*Disc diameter scales as parameters0.30 — a true area-proportional drawing would make the 46M disc a single pixel beside the 175B one. SecureScribe figures are from internal evaluation against nurse-adjudicated ground truth. Comparator figures reflect published frontier-LLM performance on comparable clinical classification tasks rather than a head-to-head run on this event set; where the two are not directly equivalent, we say so rather than implying a like-for-like contest.
See it run on your unit.
Benchmarks are evidence, not proof. A two-month pilot measures the only numbers that decide anything — yours.
Request a Pilot