TORTUS

The models and the safety layer around them are the entire product. There is no non AI business underneath, and the company's differentiation is explicitly technical rather than distributional: the OSLER agent, the Shell verification layer and the CREOLA evaluation system are all proprietary model work.

The only vendor assessed here that filters before the human rather than only after. The Shell verifies every generated statement against the consultation itself and removes anything unsupported before a clinician ever sees it, which addresses automation bias at its source: a reviewer cannot fail to catch a fabrication that was never presented. It then publishes the rate at which that filter works, satisfying the published threshold benchmark this index rewards.

Clinician approval remains the final gate, framed by a deploying clinician as professional responsibility rather than a workflow step. One capability warrants ongoing scrutiny rather than reassurance: OSLER uses vision and click based interaction to act inside the record, which is a higher autonomy surface than note generation, and the published safety metrics cover documentation content rather than action execution.

TORTUS publishes its own error science, which almost no vendor in any category of this index does. It reports a live platform metric of 92.7 percent of detected hallucinations removed by the Shell, and a 75 percent reduction in major hallucinations published in npj Digital Medicine rather than asserted in marketing.

CREOLA, its clinician in the loop labelling platform, produced what it describes as the first clinical ontology of hallucination and omission in ambient scribes, categorising errors and quantifying each by its potential to affect diagnosis or treatment, with a published study reviewing over 49,000 transcript sentences and 13,000 clinical note sentences and over 100 clinicians engaged.

Publishing a taxonomy of your own failure modes, with rates attached, is the strongest form of model transparency available and the exact inverse of the unfalsifiable vendor benchmarking this index flags elsewhere.

The boundary half of this axis is answered as well as anywhere in the index and the enumeration half is not answered at all, which is what places this in the middle band rather than higher. On the boundary, the company states no model is trained and no data is retained, and unusually that position is corroborated by the deploying institution, with a hospital telling its own patients directly that no patient data are held or used to train the system.

A customer restating a vendor commitment in its own patient communications is materially stronger than a trust page, because the institution carries the regulatory exposure if it is wrong. Operation under United Kingdom data protection law and a published trust centre complete that side. On enumeration, no located material names a foundation model provider, a model class or version, a hosting or cloud arrangement, or the parties in a sub processor list.

That is a notable gap specifically because this company discloses more about its own failure modes than any other vendor in the lane, publishing an error ontology and peer reviewed error rates, so reticence about the chain looks like an oversight rather than a policy. On that basis the information is probably obtainable, and this record is a refresh candidate. Ask for the sub processor list and for whether any third party model provider is invoked at inference.

Among the deepest evidence bases in the ambient category, and unusually it is prospective, multi site and publicly funded rather than vendor commissioned. TORTUS co ran what is described as Europe's largest ambient voice evaluation across nine clinical sites and more than 17,000 patients with 500,000 pounds of NHS England backing, independently evaluated, with results feeding the UK Government NHS Ten Year Plan for ambient voice technology and a Lancet submission under review.

A Phase II evaluation at Great Ormond Street Hospital, published with a named author list, is the first NHS study of ambient AI for paediatric documentation and reports improvements in note quality and clinician cognitive load. Separate peer reviewed work in npj Digital Medicine underpins the safety claims. Graded on the existence, venue, scale and independence of the evidence; this index does not re verify the underlying studies.

The clearest data commitment in this category, and it is corroborated by the deploying institution rather than resting on vendor assertion. TORTUS states no model trained and no data retained, and Great Ormond Street Hospital independently tells its own patients that no patient data are held by the organisation or used to train the AI.

Customer corroboration of a no training and no retention commitment is materially stronger evidence than a vendor trust page, because the hospital carries the regulatory exposure if the statement is wrong. Full UK GDPR compliance and a published trust centre complete the position.

The earlier assessment treated this axis as not applicable, on the basis that the product is built for and deployed in the NHS and its governing frame is UK GDPR, DTAC and the UK medicines regulator. That reading is superseded. HIPAA alignment is asserted for this product and the claim appears consistently across third party listings, alongside Cyber Essentials, DTAC, GDPR and the NHS Data Security and Protection Toolkit. Once the claim is made the axis applies, and the question becomes what a counterparty can actually verify.

Very little. No business associate agreement template, no scope statement covering which processing activities the claim reaches, no subprocessor list and no United States customer or deployment were located. The verifiable footprint is entirely British: London registration, a research partnership with Great Ormond Street Hospital, and NHS trust deployments. The electronic record integrations named include EMIS and TPP SystmOne, which are British primary care systems.

Two cautions belong with this. There is no certification for the United States health privacy rule, so an assertion of compliance with it is a self description rather than an examination by an independent party. And the compliance lists circulating for this vendor mix things that are not comparable: Cyber Essentials is a certification, DTAC is an NHS assessment framework, the Data Security and Protection Toolkit is a self assessment, GDPR and the health privacy rule are laws, and ISO 27001 appears as pending rather than held. Six items in one line reads as more assurance than it contains.

The substantive position is strong inside its home regime and undocumented outside it. A United States buyer should ask for the business associate agreement and its scope directly rather than relying on the compliance claim, and should not read the NHS position as transferring.

A published trust centre exists and the product is DTAC compliant, meaning it has been assessed against the NHS Digital Technology Assessment Criteria covering clinical safety, data protection, technical security, interoperability and usability. That is a genuine external assessment framework rather than a self declared badge, and more than most vendors in this category hold. Held at B because no named and dated commercial attestation such as ISO 27001 or SOC 2 Type II was located in this pass.

The most regulatorily engaged vendor in this category, graded carefully so the engagement is not confused with approval. TORTUS was selected for Phase II of the MHRA AI Airlock, the UK regulator's sandbox for AI as a medical device, and co developed the regulatory pathway for ambient voice technology with the MHRA, with the Phase 2 Programme Report published 9 June 2026. That is real and unusual.

However its current status is UKCA Class I medical device REGISTRATION, and this index holds a standing precedent that registration is not clearance: Class I is self certified by the manufacturer rather than reviewed by the regulator, the same distinction applied to Neurotrack. TORTUS is openly pursuing Class IIa, which would involve notified body assessment. Graded B for the combination of genuine regulatory co development and an honest, publicly stated intention to move to a reviewed classification. No FDA pathway.

The only ambient scribe assessed here with documented accent testing, and it was performed by the deploying hospital rather than the vendor. Great Ormond Street Hospital states publicly that its extensive evaluation tested a wide variety of accents and found the tool transcribed accurately across them.

That directly addresses the failure mode this index tracks across every speech product, where accuracy variation by accent and dialect falls hardest on populations already underserved, and it is the difference between a vendor naming the risk and an institution measuring it. CREOLA reinforces the position by classifying errors according to their potential clinical consequence rather than by frequency alone, which is the correct way to weigh a documentation error.

This is the strongest position in the band and the closest any vendor in this lane comes to the top grade, so it is worth setting out what it has and what it would still need. What it has is falsifiable error science published by the company about its own product.

A live platform metric of 92.7 percent of detected hallucinations removed, a 75 percent reduction in major hallucinations published in a peer reviewed journal rather than asserted in marketing, and a clinician in the loop labelling programme that produced a clinical ontology of hallucination and omission in ambient scribes, categorising error types and quantifying each by its potential to affect diagnosis or treatment across tens of thousands of reviewed sentences.

Publishing a taxonomy of your own failure modes with rates attached is the most useful disclosure a buyer can receive on this axis, because it converts a vague worry into a defined list of things to monitor for, and it is the exact inverse of the undefined percentages published elsewhere in this lane.

The data position is corroborated rather than asserted, with a deploying hospital telling its own patients that no patient data are held or used to train the system, and a customer that carries regulatory exposure for a false statement is better evidence than a vendor trust page. Operating under United Kingdom data protection law also gives the recorded patient a statutory route to have inaccurate records corrected, exercisable through the practice with the processor obliged to assist.

What would move this to the top grade is the one thing absent: a commercial commitment attached to the published rates. No warranty on output, no remediation obligation and no indemnity toward the customer was located, so the company measures its errors rigorously and does not yet stand behind them contractually. Ask whether the published error rates can be written into the agreement as a service level.

Integrated into NHS electronic patient records, deep enough at Great Ormond Street to pick up patient identity automatically rather than relying on manual selection, and distributed into primary care through the X-on Health partnership as Surgery Intellect.

The OSLER approach is architecturally notable: vision and click based interaction lets the agent operate a record system through its interface rather than requiring an integration to be built, which is a route to breadth that does not depend on vendor cooperation. Held at B because integration coverage is concentrated in the NHS estate and no named integrations outside the United Kingdom were verified in this pass.

The retention answer largely resolves the residency question: if no data is retained and no model is trained on customer content, the surface that residency protects is much smaller. UK deployment under UK GDPR with NHS governance approval at trust level. Held at B because hosting regions, sub processors and the treatment of data in transit during processing were not documented in this pass.

No published rate card. Procured through NHS routes including integrated care board approvals and partner channels such as X-on Health, where rollout follows NHS guidance on AI scribes rather than a published price list. One reported operational figure rather than a price: early adopters describe saving an average of four minutes per consultation.

The most unusual setting spread in this category. Documented across the ambulance clinical hub, the emergency department and outpatients as three deployments on one platform, plus paediatrics through Great Ormond Street and primary care through the X-on partnership. Pre hospital and ambulance dispatch documentation appears nowhere else in this index and is a genuinely harder acoustic and clinical environment than an outpatient room. Note the geographic bound: coverage is deep in the NHS and unevidenced elsewhere.