Deciphex

Graded the same way this index has graded Indica Labs, and for the same reason: the platform came first and the models were added to it. Patholytix is described by its own partner announcement as a non clinical workflow, into which artificial intelligence classifiers were integrated. Slide management, study organisation and peer review would all function without a model.

The other half of the business points the same way more strongly. Diagnexia's asset is a network of more than 250 subspecialty pathologists, which is the moat is the network case this index has applied to credentialing and to Reveleer. What is being sold there is human expertise routed efficiently, with artificial intelligence assisting.

A stated programme to build pathology foundation models on the company's own image repository would change this reading if it produces a model that is itself the product. It has not yet.

Low autonomy throughout and unambiguously so. In the preclinical product the stated purpose is accelerating primary evaluation and peer review by a study pathologist, who signs the assessment. In the clinical service a named subspecialty pathologist is the deliverable, not the software.

That makes this one of the few records in the index where a human being is contractually the output rather than a safeguard around it. What is not published is the shape of the assistance: whether classifiers pre annotate slides before the pathologist looks, which would risk anchoring the reader, or surface findings afterwards as a check, which would not. In peer review specifically that distinction determines whether the tool reduces error or propagates it.

No model, architecture or training description is published. The classifiers inside the preclinical decision support tool are referred to collectively and never characterised, and the announced foundation model programme is stated as an intention with no technical detail.

The one quantitative claim, that pathologists work up to 40 percent faster, is an up to figure and therefore sets a ceiling rather than reporting a result. It is the third instance of that construction in this sweep and the treatment is settled: an up to number is an absence of a number.

Nothing identifies any party in the chain: no model or model family, no hosting arrangement and no sub processor list was located in two passes, and no retention, encryption or model training position was found for either half of the business. That last phrase carries the finding, because the two halves sit under different bodies of law and public material does not separate them.

The preclinical platform processes tissue from laboratory animals in drug safety studies, which is not protected health information at all and falls outside health privacy law entirely, so the obligations attaching to it are contractual and scientific rather than statutory. The clinical service handles human diagnostic material across five jurisdictions and falls squarely inside health privacy law, with different obligations in each of those five.

A buyer is therefore evaluating two different risk profiles under one company name, and a diligence process that establishes the position for one learns nothing about the other. Anyone reviewing this should ask which entity and which platform their material passes through, and should not let assurances given about the preclinical side stand in for the clinical one. Ask for the processing position per platform and per jurisdiction, for a sub processor list covering both, and for the retention and training position on human diagnostic material specifically.

The commercial validation is genuinely strong and the published evidence is not. Charles River Laboratories, the largest preclinical research organisation in the sector, co launched a product on this platform and then extended into an exclusive image management arrangement, and Novartis partners on drug discovery pathology. Neither of those organisations adopts lightly.

Against that, no study, accuracy figure, denominator or comparison against unaided pathologists was located. The efficiency claim is unsubstantiated as written. And one structural point belongs on the record neutrally: Charles River is both the flagship partner and an investor in the Series C, so the most prominent validation is not fully arm's length. That is the fifth instance of investor and customer overlap in this sweep.

Graded on an honest basis, with a structural distinction that is unusual enough to be worth stating plainly.

The two halves of this business sit under different bodies of law. The preclinical platform processes tissue from laboratory animals in drug safety studies, which is not protected health information at all and falls outside health privacy law entirely. The clinical service handles human diagnostic material across five jurisdictions and falls squarely inside it. A buyer evaluating this company is therefore evaluating two different risk profiles, and public material does not separate them. No retention, encryption or model training position was located for either.

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

The clinical service raises a question most software records do not. A network of pathologists reporting cases for a healthcare provider is providing a professional service on identifiable patient material, so the arrangement is closer to a laboratory or a locum agreement than to a software licence, and the obligations follow accordingly. Operating that across the United States, United Kingdom, European Union, Canada and Japan means several regimes apply at once.

Recorded honestly: the dedicated trust and security search this index requires was not run in this pass, so the grade is provisional and should not be quoted until it has been.

One consideration will bear on the result. An exclusive image management arrangement with a large listed research organisation, handling that company's client study data, will have been preceded by a supplier assessment of some depth. Whatever it produced is not public.

No device clearance and none claimed, which is correct for both halves. Preclinical workflow software is not a medical device, and a diagnostic service delivered by a licensed pathologist is regulated as the practice of medicine rather than as a product.

The framework that does apply to the preclinical side is one this index has not encountered before and it is worth naming: toxicologic pathology conducted to support a regulatory submission falls under Good Laboratory Practice, which governs study conduct, data integrity and auditability rather than device safety. A tool inserted into that workflow inherits those requirements, including traceability of every change to a study record. Nothing published describes how the platform satisfies them, and for a pharmaceutical buyer that is the first question rather than a later one.

Nothing published on model evaluation, monitoring, error handling or subgroup performance.

The preclinical setting displaces the usual equity framing without removing the underlying problem. Variation there is by species, strain, tissue preparation, staining protocol and contributing laboratory rather than by patient demography, and a classifier that performs unevenly across those still distorts a safety assessment that regulators and eventually patients rely on. On the clinical side the ordinary questions return in full and are equally unaddressed.

Two passes located no model, architecture or training description, no evaluation methodology, no operating characteristics and no warranty, indemnity or remediation commitment. The classifiers inside the preclinical decision support tool are referred to collectively and never characterised, and the announced foundation model programme is stated as an intention with no technical detail behind it. The single quantitative claim is the construction this index treats as settled.

Pathologists working up to forty per cent faster sets a ceiling rather than reporting a result: a deployment delivering no speed improvement whatever satisfies it, so the claim cannot fail and therefore says nothing. This is the third instance of that construction in this sweep and the treatment does not change with repetition. An up to number is an absence of a number. The consequence side is worth naming because it differs between the two halves of this business.

In the clinical service a missed or miscalled feature is a diagnostic error reaching a patient; in the preclinical platform it is a safety signal missed in a toxicology study, which reaches patients later and through a longer chain. Neither has a published error characteristic. Ask for sensitivity and specificity by tissue and finding type, what the speed figure was measured against, and what the vendor commits to when a case is called wrongly.

The relevant systems here are laboratory information systems, slide scanners and study data platforms rather than the electronic health record, and the record is graded on that basis rather than penalised for an integration it has no reason to build.

The strongest evidence is the exclusive image management arrangement with a large research organisation, which implies the platform ingests and manages whole slide images at industrial scale across many client studies. No named interface standard, scanner compatibility list or laboratory system integration was located.

Not described. No hosting model, regional arrangement or retention schedule was located.

Stated expansion across the United States, United Kingdom, European Union, Canada and Japan implies residency has been addressed for several regimes, and the pharmaceutical customer base makes it more pointed than usual: preclinical study images are commercially sensitive intellectual property about compounds in development, so where they sit and who can reach them is a competitive question as much as a privacy one.

Nothing is published for either half of the business, and the two would price on entirely different logics: a software platform by licence, seat or study volume, and a pathologist network by case, by turnaround commitment or by subspecialty.

That second one is the more interesting omission, because the service is positioned explicitly as an improvement on the locum model, which has a known and quotable cost. A comparison against a named alternative is being invited without either figure supplied.

Unusually wide because the company spans two markets that rarely share a vendor. On the preclinical side it reaches pharmaceutical and biotechnology drug safety programmes through the largest research organisation in that sector. On the clinical side it reaches healthcare providers through a network of more than 250 subspecialty pathologists, which is itself a breadth claim since subspecialty coverage is precisely what a single hospital laboratory lacks.

Geographic reach spans the United States, United Kingdom, European Union, Canada and Japan. Held at B rather than A because everything sits within pathology, and because the clinical half is a service whose reach is bounded by how many pathologists are in the network rather than by software.