HOPPR

The foundation model is the entire product. There is no clinical application to buy, only Grace and the environment for adapting it, which makes this the most AI native position in the index by construction: a customer purchases model weights and the means to fine tune them.

The axis does not reach the Foundry in its literal form, since a development environment takes no clinical action, but it reaches Presto Agent directly and the equivalent question for the Foundry is left open as well.

For Presto the mechanism is specific. The product places AI results into the radiologist's existing reporting environment, and a radiologist quoted by the company describes no longer having to dictate measurements into reports. Automatic insertion of a numeric measurement produced by a third party model into a report the radiologist will sign is a different oversight problem from a drafted sentence. A reader can evaluate prose against the image and notice a claim that looks wrong. A number placed in a template looks identical whether it is right or wrong, and reviewing it properly means going back to the image and remeasuring, which is the work the feature exists to remove.

Nothing published states what Presto may insert without explicit confirmation, whether inserted content is visually distinguished from dictated content at the point of signing, or whether the signed report retains any record of which content originated from a model and which model produced it. That last question is the one to press, because the report is the permanent clinical and legal artefact.

For the Foundry the translated question is what the platform requires of downstream builders by way of human oversight design, and nothing addresses it. Graded C because a domain equivalent plainly applies on both products and neither is answered.

Among the strongest disclosures in the index, and appropriately so given that developers building regulated products downstream inherit whatever the base model does. Training corpus is quantified and characterized: over a petabyte of permission based anonymized imaging studies paired with corresponding reports, spanning 2D, 3D, and longitudinal series across modalities.

The narrative model ships with training data traceability records explicitly to support downstream evaluation, and with version locking so teams can reproduce results across development and deployment. Third party models on the platform are named rather than white labeled. Publishing provenance and pinning versions is what a developer needs to defend their own regulatory submission later.

This is a base model provider and it publishes what a base model provider should, which matters more than for an ordinary vendor because developers building regulated products downstream inherit whatever the base does. The training corpus is quantified and characterised: more than a petabyte of permission based anonymised imaging studies paired with corresponding reports, spanning two dimensional, three dimensional and longitudinal series across modalities, so the consent basis is stated rather than implied and the composition is described rather than asserted.

The narrative model ships with training data traceability records explicitly to support downstream evaluation, and with version locking so teams can reproduce results across development and deployment. Those two artifacts are the ones a developer needs to defend their own regulatory submission later, and shipping them is a decision to make downstream accountability possible rather than to leave it to the customer.

Third party models on the platform are named rather than white labelled, which closes the gap this index has recorded repeatedly where a distributor obscures whose engine is running. Three residuals to obtain rather than assume: the de identification standard applied, the mechanism by which permission was obtained, and whether customer data used in fine tuning is isolated from the base model. For a platform whose business is other people's imaging, those are the decisive terms.

Evidence appropriate to infrastructure is thin here. The narrative model is stated as evaluated against internal benchmarks with no published results, comparison, or external validation, and no downstream clinical application built on Grace was identified with outcome data. Investment from Health2047, the American Medical Association venture studio, and a strategic partnership with RadNet are credibility signals rather than performance evidence. For a base model that others build regulated products on, published benchmark performance is the reasonable expectation and it is absent.

The consent basis for training data is stated rather than implied, described as permission based and anonymized, which is more than most model builders in healthcare disclose, and development runs inside a stated HIPAA compliant environment. Held back from A because the de identification standard, the permission mechanism, and whether customer data used in fine tuning is isolated from the base model were not detailed in retrieved materials. For a platform whose whole business is other people's imaging data, those are the decisive terms.

No business associate agreement terms are published and nothing states whether an agreement is standard, negotiated or separately priced. What lifts this above a bare assertion is HITRUST e1 certification, which is a healthcare specific framework assessed independently and mapped to the Security Rule's control expectations, and the company's care in describing it accurately rather than claiming a HIPAA certification that does not exist.

The substantive gap is that this company runs three distinct data flows with three different postures and never distinguishes them.

The base model corpus is described as permission based and anonymized, which if genuinely de identified to a recognised standard sits outside the framework altogether. Customer data brought into the Foundry for fine tuning is protected health information processed on a customer's behalf, which is ordinary business associate territory. And Presto Agent operates inside the reporting environment on live report content, which is a third relationship again, with a different customer and a different set of systems.

A buyer needs to know which agreement covers which flow, whether one agreement spans all three, and what happens to fine tuning data at the end of an engagement. Ask specifically whether the agreement covering Foundry access extends to Presto, given the company states the two are independent products.

Among the more precisely stated security postures in this lane, and the precision is the point. The company publishes SOC 2 Type I and Type II attestation with the assessing firm identified on the badge, and HITRUST e1 certification for the Foundry following an independent assessment. It names the HITRUST tier rather than saying certified without qualification, which matters because e1 is the entry level of three and a vendor that names it is not inviting a buyer to assume r2. The security programme is enumerated rather than asserted: access management and least privilege, continuous monitoring and logging, incident response, vendor risk management, and governance. Infrastructure is named as Amazon Web Services, with controlled data access and isolation stated. A named security lead publishes explanatory posts on what each credential covers.

Two things hold it at B. There is no trust centre, no statement that the reports are available on request, no penetration test disclosure, no subprocessor list and no period or criteria stated for the attestation.

More importantly, the scope precision that earns this grade also reveals a gap. The certifications are stated for the Foundry, and the company states elsewhere that Foundry and Presto Agent are independent products. Presto is the newer product, it sits inside the radiologist's live reporting environment, and nothing published places it inside the certified boundary. Ask directly whether Presto is in scope for the SOC 2 and the HITRUST certification, or whether it is a separate environment awaiting its own assessment.

No clearance is claimed and none would apply to the Foundry, which is correct rather than a gap. The company supplies models and development infrastructure, and regulatory responsibility for any resulting clinical product sits with the developer who builds and submits it. Buyers should understand that clearly, since using a cleared component does not confer clearance and using an uncleared foundation model does not prevent it. The traceability records and version locking exist precisely to support that downstream submission burden, which is the company naming the regime it sits adjacent to and building for it rather than ignoring it. That is what earns a B in a position where no vendor level regulator exists.

A second product complicates the picture and the company has not addressed it. Presto Agent places AI results into the radiologist's live reporting environment. It is deliberately model agnostic and does not supply the analysis, so it is not itself the device, and the practice chooses which models run. But it is the layer through which another manufacturer's cleared output reaches the clinician and enters the signed report.

That raises a question this category has not settled. A cleared imaging device is authorised with particular labelling, particular output formatting and particular assumptions about how a radiologist encounters the result. An independent integration layer that re renders that output into a different environment sits between the manufacturer and the user, and can alter the conditions under which the cleared product is actually used, without being the manufacturer of anything.

So the company now sits at both ends of the imaging AI supply chain and at neither point is it the regulated middle: upstream of the device as a model supplier, and downstream of it as the delivery path. Ask what the company holds itself responsible for when a third party model's output is presented through its integration layer, and whether device manufacturers whose outputs it carries have agreed to that presentation.

No subgroup analysis, no bias assessment, no fairness evaluation and no model card describing the training distribution was located. In this lane that is a common finding. Here it is the most consequential instance of it in the index, because of where this company sits.

Grace is a base model sold so that others can build clinical products on it, several of which will be submitted for regulatory clearance. The training corpus is quantified as over a petabyte of permission based anonymized studies with paired reports across modalities, which is real provenance disclosure and is credited on the transparency axis. But quantity is not composition. Nothing states the geographic origin of the studies, the mix of scanner manufacturers and field strengths, the institutional mix between academic and community settings, or any patient demographic characteristic of the population represented.

The consequence is structural rather than theoretical. A developer building a device on this model inherits its distribution, and must characterise their own product's performance across subgroups to support a submission and to satisfy the source attribute expectations now attached to decision support interventions. They cannot reason about inherited bias in a base model whose composition is undisclosed, and the model provider is the only party positioned to describe it. Publishing corpus composition is a foundation model obligation in a way it is not for an application vendor.

Presto adds a second and different question. It is deliberately model agnostic and the practice selects which models run, which places validation and monitoring of those models with the customer. Nothing describes what a practice must establish before wiring a model into live reporting, who signs that off, or whether performance is monitored afterwards.

The accountability mechanism here points downstream, which is the right direction for a base model and rarer than it should be. Version locking means a developer can reproduce results across development and deployment rather than discovering that the foundation shifted under a product already in the field, and shipped training data traceability records exist explicitly so downstream teams can evaluate what they built on.

Together those let a developer answer a regulator's questions about their own product, which they otherwise could not, because a submission resting on an unversioned undocumented base model is resting on something nobody can characterise. That is a concrete transfer of the means of accountability to the party who will carry the regulatory burden, and it is worth naming as its own route into this band.

The general principle is worth stating: a base model vendor's disclosure decisions become every downstream product's disclosure ceiling, so opacity here multiplies. Held below the top grade because nothing attaches commercially. No accuracy commitment, no service level, no indemnity toward downstream developers and no stated position on what happens if a base model defect surfaces in a cleared downstream device was located. Ask what the vendor commits to when a defect is found after downstream clearance, how long a version remains available, and what notice a developer receives before a version is retired.

Interoperability is the substance of one of the two products rather than a feature attached to it, which is unusual in this lane. Presto Agent is positioned explicitly as the last mile problem: getting model results to the radiologist without changing the picture archiving system, the reporting platform, the templates or the clinical workflow. Radiologists using it describe it running inside their existing reporting application, with one naming the platform, and describing setup as requiring nothing on their part. The Foundry side sells to radiology picture archiving vendors as a development environment, so integration is the customer relationship in both directions.

Held off A because the compatibility matrix is unpublished, and for an integration product that is the central fact. Nothing lists which reporting platforms and archiving systems are supported, which versions, or what the integration mechanism actually is. No standard is named anywhere, so a buyer cannot tell whether this rides established imaging and messaging interoperability standards, a vendor interface, or something operating at the presentation layer of the reporting application. Ask for the supported platform list and the integration method before assuming a given stack is covered.

More is disclosed than most vendors in this lane manage, but the decisive terms are missing and for this business model they are the ones that matter.

What is stated: the platform runs on Amazon Web Services, with controlled data access and isolation named as design properties, and the environment described as a controlled development platform assessed independently. Presto deploys into the customer's existing reporting environment without changes to surrounding systems.

What is absent: no cloud region, no data residency commitment, no statement of whether the Foundry is single or multi tenant, and no description of where fine tuning compute physically runs. The isolation claim is asserted at the level of a design property rather than described.

This is a near miss rather than a bare absence, and the specific things that would move it are worth naming. For a platform whose entire business is holding and processing other organisations' imaging data, a buyer needs the region, the residency terms, the tenancy model, and above all a statement of whether data brought in for fine tuning is cryptographically and logically separated from other customers' work and from the base model. That last question is already flagged on the stewardship axis and it belongs here too, because the answer is an architectural fact rather than a policy one.

No public pricing. Contact the vendor. Sold to developers, AI companies, and radiology PACS vendors as platform access, with Forward Deployed Services as a separate professional services engagement pairing HOPPR staff with customer teams. Buyers should establish whether compute, model access, and services are priced separately, since a fine tuning heavy program can shift cost substantially toward the services line.

Scoping is clear on modality and on buyer, which is more than most of this lane offers. Grace is stated to span X-ray, computed tomography, magnetic resonance and echocardiography, across two dimensional, three dimensional and longitudinal series, with paired reports. The buyer set is named on both sides of the business: developers, AI companies and radiology picture archiving vendors for the Foundry, and radiology practices and health systems for Presto Agent, with named radiologists at named private practices describing daily use.

Held off A on two counts. No subspecialty scoping is stated anywhere, so nothing indicates whether the model performs comparably on neuroimaging, musculoskeletal, thoracic or abdominal work, and for a multimodal base model that variation is the practical limit on what a downstream developer can build. And the modality list describes what the model was trained across rather than what it has been evaluated on, which are different claims. Ask which modalities carry evaluation results rather than training coverage.