Aignostics

The foundation model is the company. There is no scanner, no laboratory, no image management platform and no service layer underneath it, and the products sold are either the model itself for integration by others or applications built directly on top of it.

The commercial logic follows the model rather than the other way round. Successive releases are announced as model results, the partnerships with a large pharmaceutical company and two academic medical centres exist to obtain training data and validate performance, and the newest offering is a self service application exposing what the model can already do.

Low autonomy, and appropriate to what this is. The models produce representations and measurements from tissue images for researchers and drug developers, and a scientist decides what to do with the output. Nothing acts on a patient and nothing issues a diagnosis.

The oversight that is described sits in development rather than in operation. The company describes a pathologists in the loop approach, with medical expertise built into curation and model development at every stage rather than applied as review afterwards, and the first model's published method makes the same argument, that existing self supervised approaches do not leverage pathologist knowledge by design. That is a real design commitment and it is a statement about how the model was made, not about who checks its output in use.

The most transparent record in this index by a clear margin, and the standard other model vendors should be measured against.

What is published, in preprints with named authors rather than in marketing copy: the architecture and its parameter count, the training dataset size in slides, cases and extracted tiles, the magnifications sampled, the sampling algorithm, the institutions the data came from, the number of contributing laboratories, tissue types and staining modalities, and full benchmark results. Crucially the benchmarks are comparative and the rivals are named, with their own parameter counts and dataset sizes given alongside, so a reader can judge whether a claim of state of the art is meaningful rather than selected. Evaluation runs across 21 public benchmarks for one model and 80 for its successor.

This is the exact inverse of the transparency anti benchmark this index recorded for Isomorphic Labs, where every figure is vendor generated on vendor selected benchmarks and unfalsifiable from outside. Here an outside party can check the claim on public data.

The corpus provenance is enumerated to a level nothing else in this index matches, and the precision is what makes the harder question unavoidable. The models were trained on one point two million and later more than five million whole slide images drawn from the archives of named academic medical centres, representing over four hundred and ninety thousand patient cases, with contributing laboratory counts, tissue types and staining modalities all stated.

That is the corpus half of this axis answered properly rather than with an adjective, and a reader can judge whether the training population resembles their own. It also makes visible what most vendors keep invisible. Those slides are real people's tissue, taken and examined for their own care, and the resulting models are commercial assets.

Nothing published describes what governance covered the transfer from those archives, what legal basis applied, or what those patients were told, and this index should be explicit that publishing the provenance so precisely is to the company's credit rather than the reason it faces the question: a vendor that says nothing faces it too, and simply cannot be asked.

Held below the top grade for that gap and because no operating chain is named, with no hosting arrangement or sub processor list located. Ask what governed the archive transfers, and for the hosting and sub processor position.

The evidence is technical and comparative rather than clinical, which is the right kind for what is being sold, and it is unusually strong of its type. Performance is reported across public benchmark suites against named competing models, published in preprints, so the results are reproducible by third parties on the same data.

What is absent is evidence at the level of use. No study shows a diagnostic or drug development outcome improved by these models in practice, no clinical deployment is described, and the biopharma partnerships are named without published results. Benchmark leadership demonstrates that the representations are good; it does not demonstrate that a target was found, a trial enriched or a patient better treated. Held at B rather than A on that distinction alone.

Graded on an honest basis. No published position on data handling, retention or de identification for customer material was located in this pass.

The substantive question is the same one the electrocardiogram records in this index raise, and it is more visible here because the numbers are published. The models were trained on 1.2 million and later more than 5 million whole slide images drawn from the archives of named academic medical centres, representing over 490,000 patient cases. Those are real people's tissue, examined for their own care, and the resulting models are commercial assets. Publishing the provenance so precisely is to the company's credit and it also makes the question unavoidable: nothing describes what governance covered the transfer or what those patients were told.

Graded on an honest basis and flagged for re verification. No compliance statement or agreement posture attributable to the company was located in this pass.

The jurisdictional position is worth noting because it is not the usual one. This is a German company with a United States presence, training on data from both a German and an American academic medical centre and selling to global pharmaceutical companies, so European data protection law and United States health privacy law both bear on it, and the relevant frameworks differ. A buyer should establish which applies to which data flow.

Recorded honestly: the dedicated trust and security search that this index requires before assigning a grade on this axis was NOT run in this pass. The grade is provisional and should not be quoted until it has been.

That distinction matters because it is different from the case where a search was run and returned nothing. Companies at this stage of maturity, selling to large pharmaceutical partners who run vendor security reviews as a matter of course, frequently hold attestations that never appear in product material. Flagged as an open work item.

No clearance and none claimed, which is correct for a research and drug development tool. The company does not sell a diagnostic to a clinician.

One disclosure here is genuinely novel for this index and deserves recording. The company states that its latest model ships with clinical grade regulatory documentation intended to support integration into medical devices built by others. That is a component supplier position: rather than seeking its own clearance, it prepares the evidence package a downstream manufacturer would need in order to include the model inside a regulated product. This index has repeatedly found certifications that do not transfer between parties. Here a vendor is deliberately building the paper trail so that its model can transfer, which is the same problem approached from the opposite direction and a considerably harder thing to do well.

Better than almost anything in this index, and the reason it is B rather than A is a distinction worth drawing carefully.

What is published is preanalytical diversity, deliberately and in detail. The first model was trained across more than 15 laboratories, 58 tissue types and 129 histochemical and immunohistochemical staining modalities; its successors span multiple scanners, diseases, staining types and four magnifications; and robustness is treated as a headline property rather than an afterthought, described as a critical requirement for clinical deployment. That directly addresses the failure mode that has embarrassed pathology models before, where performance collapses on a different scanner or a different laboratory's staining protocol.

What is not published is patient population diversity. Nothing reports performance by patient sex, age or ancestry, and the training archives come from a small number of academic medical centres in Germany and the United States, which are not demographically representative of the patients these models will eventually be applied to. Variation across laboratories is documented; variation across people is not.

This is the strongest published evidence position located anywhere in the backfill, and the feature that puts it there is comparative benchmarking against named rivals. Preprints with named authors publish the architecture and its parameter count, the training dataset size in slides, cases and extracted tiles, the magnifications sampled, the sampling algorithm, the contributing institutions and laboratory count, the tissue types and staining modalities, and full benchmark results across twenty one public benchmarks for one model and eighty for its successor.

Any of that alone would be unusual. What matters most is that the benchmarks are comparative and the competitors are named, with their own parameter counts and dataset sizes given alongside, so a reader can judge whether a claim of leading performance is meaningful or merely selected. A vendor choosing its own benchmark can always win; a vendor reporting on public benchmarks against named alternatives has given every one of those competitors the means to answer back.

That is the exact inverse of the pattern this index has recorded where every figure is vendor generated on vendor selected tests and unfalsifiable from outside. Held below the top grade because nothing attaches commercially: no warranty, indemnity, service level or remediation commitment was located, and benchmark performance is not a committed error characteristic for a laboratory's own case mix. Ask what the deployed configuration achieves on your tissue types and stains.

Largely not applicable in the usual sense and graded accordingly rather than penalised. The buyer is a pharmaceutical research organisation or a device manufacturer, not a clinician working inside a record system, so there is no chart to write back to.

The interoperability that does matter is with the pathology imaging estate: whole slide image formats, scanner outputs and image management platforms. Working across multiple scanners is claimed as a model property and the newest application is described as self service, which implies a hosted route to access. No named platform integration, image standard or application interface detail was located.

Not described. No hosting model, region or retention position was located, and no statement covers whether customer slides are processed in the company's environment or the customer's.

One architectural fact points in a useful direction: distilled smaller versions of the newest model were released specifically to enable adoption across diverse settings on constrained compute. That implies the models can run somewhere other than a single central service, which is the beginning of an answer to where a pharmaceutical company's proprietary trial images would be processed, and the company has not spelled the answer out.

Nothing published. No price, no pricing mechanism and no unit of sale, across what appear to be three quite different commercial routes: model licensing for integration, biopharma research collaborations, and a self service application.

Those three would normally be priced entirely differently, by licence, by programme and by usage respectively, and none is described. The self service framing of the newest product implies a usage based or subscription route exists, which makes the absence of any published rate more notable than it would be for a pure enterprise partnership business.

Wide in tissue and narrow in setting, and the second bounds the grade. The models cover 58 tissue types and 129 staining modalities, so the biological range is as broad as pathology itself.

The buyer, however, is almost entirely the pharmaceutical industry: drug discovery, translational research, clinical trials and companion diagnostic development, with a secondary route to device manufacturers integrating the model. There is no clinical deployment, no diagnostic use at the point of care, and no hospital buyer. That is a deliberate and coherent position rather than a shortfall, and it does mean the product reaches patients only indirectly, through the therapies and tests it helps develop.