The verification layer for clinical AI
Every fact in the note is found in the recording first, or it is thrown away. Nothing leaves the room.
Every major vendor ships this. None publishes what it gets wrong.
Ambient documentation arrived in clinics with distribution behind it rather than evidence. It is now default in the largest electronic record in the world, and the people relying on it cannot tell you how often it is wrong.
of large US health systems have deployed an ambient scribe or are piloting one
of 565 notes from three deployed scribes carried a failure that reviewers verified
vendors publishing a fact-level failure rate for what they ship
How it arrived
- Epic shipped native AI charting
- February 2026
- athenahealth bundles it
- at no marginal cost
- The best randomised trial
- found no significant time saving
An unmeasured benefit, a documented harm, and no way to tell them apart.
And none of it can be compared
Same 565 notes. Change only the instruction given to reviewers, and the share of candidate errors that gets verified moves:
9.3%to79.0%
An eight-fold swing from the counting rule alone.
Adoption via secondary reporting of KLAS 2025. Deployed-scribe failure census: npj Digital Medicine, scribe error rates, 2025Oxford University Hospitals, ambient voice technology evaluation, February 2026Epic AI Charting general availability, February 2026
We asked a model for its evidence
On our own hardware, a 3B model read a consultation and reported that the patient was on amlodipine. That was correct. We asked it which words in the transcript proved it. This is what it returned.
“The chest pain hasn’t come back, which is good.”
A correct fact. A confident citation. A sentence that has nothing to do with it.
That is the failure this company exists to remove. Not the wrong fact, which a clinician can catch. The right fact with invented proof, which looks exactly like the right fact with real proof.
- 94%
- of frontier-model citations resolve to something real
- 39-77%
- actually support the claim they are attached to
- 0
- ways to tell, from the output, which kind you are holding
Scale narrows this. It does not close it. The frontier ceiling on citation precision is about 77%, not 100%.
Aura internal benchmark, measured on our own hardware, October 2026 (data room)Onweller et al., arXiv:2605.06635, preprint, May 2026Leung et al., arXiv:2607.08700, preprint, July 2026
Try it
No model decides what a fact is, or where its evidence lives
Rules read the transcript for drug names, doses, allergies and the word no. Each candidate fact is then matched back against the text. If the words cannot be found, the fact does not exist. The note is written from what survives.
Facts 11 of 11
- SymptomVerify
Negatedno chest pain
finding chest pain
- MedicationVerify
amlodipine 5 mg every morning
name amlodipine dose 5 mg frequency every morning
- SymptomVerify
Negatedno swelling
finding swelling
- MedicationVerify
aspirin 75 mg at night
name aspirin dose 75 mg frequency at night
- AllergyVerify
penicillin — rash
substance penicillin reaction rash
- SymptomVerify
Negatedno cough
finding cough
- SymptomVerify
Negatedno wheeze
finding wheeze
Also extracted
- Observationblood pressure 138/84 mmHg
- Exposurenoise
- Exposurehand-arm vibration
- PlanI'd like to arrange a hearing test, and we'll review in three months
Transcript, with evidence
Doctor: Morning. How have things been since the last review? Patient: Not too bad. The chest pain hasn't come back, which is good. Doctor: Good. And you're still on the amlodipine? Patient: Yes, five milligrams every morning. Doctor: Any swelling in the ankles? Patient: No, nothing like that. Doctor: And the aspirin, seventy-five milligrams? Patient: Yes, I take that one at night. Doctor: Any allergies we should know about? Patient: Penicillin. I come out in a rash. Doctor: Your blood pressure is 138 over 84 today. Any cough or wheeze? Patient: No, none of that. Doctor: Do you get noise exposure at work? Patient: Yes, I'm around loud machinery and vibrating tools most days. Doctor: I'd like to arrange a hearing test, and we'll review in three months.
Every shaded run is a character offset the extractor matched. Point at a fact to see the words it came from.
Draft note
Reject a medication and it leaves the note. Reject the allergy and the heading stays, reading “Not discussed”, because an absent heading would read as “no allergies”.
key f3264c133fdcb2b1
This panel is the shipped review screen, running on output frozen from the real pipeline: the transcript, the facts, the character offsets and every note you can reach were produced by the product, not written for this page.
The limit, stated before you find it: speech recognition is a model, and its errors propagate. Aura guarantees that a fact traces to words in the transcript. It cannot guarantee the transcript heard the room correctly.
Aura internal benchmark, measured on our own hardware, October 2026 (data room)Koenecke et al., FAccT 2024, on speech-recognition hallucination
Everything runs here. Nothing phones home.
There is no cloud tier, no fallback endpoint and no telemetry. The clinic keeps working when the line drops, and so does the privacy.
- 1Microphone arraySix channels, in the room
- 2Speech to textwhisper.cpp on the box, at 0.24x real time
- 3The rules engineDrug names, doses, allergies, and the word no
- 4MatchingFinds each fact in the transcript, or drops it
- 5Encrypted storeAES-GCM. Deleting burns the key first
There is not one outbound call anywhere in the code, and a test fails the build if somebody adds one.
What that takes off the table
- The recording leaving the building
- It cannot. Nothing is sent anywhere.
- A made-up drug or dose
- Rules read it, and a clinician confirms it before export.
- A missed no
- Denials are written out in full. They are never dropped.
- Invented proof
- We find the words, or we bin the fact.
- Data you cannot delete
- Burn the key and the recording is gone.
Speeds measured on our own hardware. Aura internal benchmark, measured on our own hardware, October 2026 (data room)Healthwatch England and YouGov, July 2026
Three guarantees nobody else publishes
We reviewed every competitor in four markets. Not one of them states how often its output is wrong at the level of a single clinical fact.
Provenance
A fact is located in the transcript before it is stored. Every evidence span is re-checked against the text, and anything that no longer matches is moved to a dropped list rather than quietly kept.
No untraceable medication fact can reach the note
Determinism
Run the same consultation through the same template twice and you get the same note, character for character, with the same key. There is no sampling and no temperature.
Reproducibility key f3264c133fdcb2b1
Audit
A log that cannot be edited and survives deleting the record itself. The proof that permission was given outlives the encounter; the patient's own words do not. Signing and audio replay are funded, not built yet.
The clinician's verdict is kept, not just the note
Competitor review, October 2026. Heidi raises $340m, HLTH, September 2026Omi Scribe, offline Mac modeEka Care and NVIDIA offline multilingual scribe, February 2026
Four regulators, four answers, all moving the same way
The four markets we have studied disagree about whether a scribe is a medical device. They agree on what they are starting to ask for: show your working.
Not a device
United Kingdom
Not a device since July 2026. But hallucination has to be mitigated by design, and a disclaimer is not an acceptable mitigation.
Class IIa minimum
European Union
Sweden's regulator found Class I is not enough. The test is the output used in care, not how it was produced.
Not a device
United States
Not a device under the revised FDA guidance. Which also means no regulator has reviewed anybody's evidence.
Explicitly excluded
India
Excluded more clearly than in the UK. Documented clinical observations sit outside the device rules.
Provenance is becoming a requirement rather than a differentiator. Aura has it natively. Everyone else has to retrofit it.
MHRA, ambient voice technology guidance, 29 July 2026Swedish MPA inspection findings, as reported by Tandem, May 2026FDA, Clinical Decision Support Software guidance, rev. 29 January 2026CDSCO, Guidance on Medical Device Software, 2026
The appliance is the reference implementation. The layer is the company.
Deterministic extraction and mechanical provenance do not care who captured the audio. The box proves the architecture in a real room under a real regulator. What it proves is sellable well beyond the box.
Shipping now
Documentation you can check
- Evidence-linked notes, with every material fact tied to its offsets
- A review gate the clinician has to clear before anything is exported
- Byte-identical notes from the same consultation and template
- Spoken permission, recorded as proof that outlives the encounter
Next
The ambient clinic, still entirely local
- Coding suggestions that trace to something actually said
- Random audit mode, comparing stored audio against signed notes
- Referral letters and plain-English patient summaries, drafted on the box
- Live interpreting that never phones home
The company
The assurance layer
- Verification sold as software, wherever the audio was captured
- Checking any vendor's output against the source recording
- Per-encounter licensing rather than per device
The device is the only place in the clinic where the ground truth of the consultation exists. Everything downstream can be checked against it.
The first column ships today. The other two are the plan, and they are written down as a plan rather than as a promise.
Aura three-horizon plan and blueprint, October 2026 (data room)Derksen et al., npj Digital Medicine, 2026
What would prove us wrong
We publish what we got wrong, including about ourselves
One of our own findings was falsified on hardware, and we corrected it in writing rather than quietly. The sample behind our benchmark is small and the corpus is ours. These are the five things that would break the argument on this page.
- 01
Local quality stalls
If a 4B model running on a device still visibly underperforms the cloud on note quality by mid-2027, privacy has to carry the whole product on its own. It is not enough on its own.
- 02
A major vendor ships a real on-premise tier
Omi Scribe already ships an offline, evidence-linked Mac mode, so local-first is not unique. If a large vendor productises a clinic-local appliance, the moat narrows to execution.
- 03
Deletion gets mandated with no carve-out
Retention is defensible today because monitoring tool safety is a recognised purpose. If that closes, the audit mode goes with it.
- 04
Liability is legislated away
A statutory safe harbour for AI-assisted documentation would remove much of the demand for provenance.
- 05
Camera capture becomes standard
If competitors ship camera or screen-capture medication capture, that advantage disappears inside a release cycle.
A benchmark you cannot fail is marketing. The instrument matters more than the number it produces.
Aura internal benchmark, measured on our own hardware, October 2026 (data room)Omi Scribe, offline Mac modeEka Care and NVIDIA offline multilingual scribe, February 2026
For clinics that have to defend what they wrote
We are looking for sites willing to run the device in real consultations and hold it to the benchmark it claims. The evidence pack, the bench results and the regional studies are available under NDA.
Get in touchWho this is built for
- Occupational health providers writing statutory, employer-facing reports
- Medical directors who need to audit what an AI wrote against what was said
- Clinics where the recording is not allowed to leave the building