Triomics

Reading hundreds of pages of narrative oncology records and reasoning against trial eligibility criteria is the product, and there is no non AI version of it. The founding observation is precise and worth repeating: structured data, roughly 20 percent of the record, was already tractable with conventional software; the value is in the other 80 percent sitting in free text.

The design decision that earns this grade is citation. Every match is traced back to the pathology, biomarker, and note level evidence that produced it, so a research coordinator verifies the reasoning rather than trusting a ranked list, and the visit preparation output is described as verifiable patient summaries. Output is a screened, ranked candidate set for human confirmation, not an enrollment decision. In trial matching, where a false negative means a patient never learns an option existed and a false positive wastes scarce coordinator time, showing the evidence is the control that matters.

Among the most specific architecture disclosures in this index. Rather than referring to a proprietary model, the company states OncoLLM is a constellation of eight specialized models ranging from 3 billion to 72 billion parameters operating agentically at the patient level. Naming the count and the parameter range lets a technical buyer reason about where inference runs, what it costs, and why smaller models handle narrow extraction while larger ones handle synthesis. Peer reviewed validation in Nature Digital Medicine, rather than a vendor whitepaper, substantiates the approach.

One control here addresses a risk most of this index never names, and it is designed rather than promised. The evaluation pathway makes synthetic data the default: a prospective customer can work in a synthetic sandbox or bring a de identified record, and real data requires a signed business associate agreement first.

That ordering matters more than it sounds, because a demonstration environment is a recognised place for real patient records to end up inside a sales process, running with weaker controls and wider internal access than production, and this company has removed the temptation rather than relying on discipline.

In production, every output carries a citation back to the specific line in the record that produced it, with audit trails described as native, so a system that records what was read to produce each answer is materially more accountable than one returning conclusions without provenance. The open question is tuning and it is the important one.

The company builds institution tuned models rather than serving one general model, and nothing states whether the tuning artefacts remain specific to the institution whose data produced them, whether any learning flows back into a shared base, what is retained after tuning, or what becomes of an institution specific model at the end of an agreement. A vendor that tunes on customer records owes a clear answer on all four, and proximity to the data does not supply it. The screening surface is also every patient with an upcoming appointment rather than a consented cohort.

The strongest evidence position in the clinical trials category. Peer reviewed validation in Nature Digital Medicine with ASCO presentation, an independent pilot study at the Medical College of Wisconsin Cancer Center spanning four disease teams, and named deployment at Mount Sinai Tisch Cancer Center as the first NCI designated Comprehensive Cancer Center in New York City to run systemwide matching.

Reported outcomes are specific and multi dimensional rather than one headline: 40 percent more trial matches, over 30 percent more enrollments, 67 percent less chart review time, and 100 percent screening coverage at three times manual throughput at one center. Deployment across top 10 US cancer centers is claimed. Buyers should still confirm the baseline each percentage is measured against.

Concrete controls rather than a policy, with one significant question left open.

The evaluation pathway is designed correctly and the company publishes it. A prospective customer can work in a synthetic data sandbox, or bring a de identified record, and real data requires a signed business associate agreement first. That ordering matters more than it sounds. A demonstration environment is a recognised place for real patient records to end up inside a sales process, running with weaker controls and wider internal access than production, and this company has removed the temptation by making synthetic data the default path.

In production, every output carries a citation back to the specific line in the record that produced it and audit trails are described as native to the product. For an axis concerned with what happens to patient information, a system that records what was read to produce each answer is materially more accountable than one that returns conclusions without provenance.

The open question is model tuning and it is the important one. The company builds institution tuned models rather than serving one general model to everyone. Nothing published states whether the artefacts of that tuning remain specific to the institution whose data produced them, whether any learning flows back into a shared base model, what is retained after tuning, or what becomes of an institution specific model at the end of an agreement. A vendor that tunes on customer records owes a clear answer on all four, and a buyer should not assume the answer from the fact that the deployment sits close to the data.

Also worth establishing: the screening surface is every patient with an upcoming appointment, evaluated continuously as charts change, so the population in scope is the whole clinic rather than a consented cohort.

An explicit and vendor published business associate posture, which is uncommon in this category and is what earns the grade.

The company's own product page offers a prospective customer two routes into an evaluation: use a synthetic data sandbox, or sign a business associate agreement to upload real data. Elsewhere it offers evaluation on a de identified record. That is a clear statement that the company executes business associate agreements, and that it treats a signed agreement as the precondition for handling real records rather than something negotiated afterwards. Most vendors in this lane say nothing at all on this axis; this one puts the answer in front of a buyer before the first call.

Held at B rather than A on three points. The company does not characterise its role, so a buyer cannot tell from published material whether it positions as a business associate in all configurations or only some. It does not identify which legal entity signs. And the agreement terms are not published, so the substance behind the offer is unknown.

A second legal instrument applies here and is not addressed. Prescreening every patient with an upcoming appointment against every open trial means screening records before anyone has consented to anything. The governing instruments for that are institutional review board approval and an authorisation or waiver, not a business associate agreement, and the two answer to different authorities. Establish under whose authority the screening runs and who holds the approval, alongside the agreement.

No SOC 2, HITRUST, ISO 27001 or equivalent attestation was located and there is no trust centre. The search reached the company's own home and product pages, and no security or compliance page was published there.

Two things partially offset that and neither substitutes for an attestation.

The company does execute business associate agreements and says so publicly, which is graded on the health privacy axis and is a contractual commitment rather than an examined control. And the reported architecture places the system inside the customer's own record data estate, which means an in boundary buyer has substituted architecture for assurance. That legitimately changes what a security review should concentrate on, from the vendor's own environment to what the deployed component transmits, what telemetry leaves and what the vendor can reach during support.

It does not cover the vendor's corporate environment, its development and release practices, its own supply chain, or the access its engineering staff hold to deployed instances. Those are exactly what an attestation examines and none of it is evidenced here.

The absence carries real weight given the deployment footprint. This system reads complete oncology charts continuously across entire cancer centre populations at named academic and designated comprehensive cancer centres. Ask whether any attestation exists and what it is scoped to, what the deployed system sends outbound and on what schedule, and what access company staff hold to a live instance with what logging and what customer visibility over it.

A scoping determination, and it closes, though the boundary is closer here than for a pure matching tool and is worth watching.

The core products prescreen patients against trial criteria, abstract structured data from notes for registry reporting, and prepare summaries for a visit. None diagnoses, none directs treatment, and no device pathway is claimed. Matching output is a candidate list a research coordinator works from, which places it in research operations rather than clinical care.

The boundary is worth naming because the product line has moved toward it. A summary prepared for a clinician to read before a visit sits closer to clinical decision support than a trial match does. The exemption that keeps such software outside device regulation depends on the clinician being able to independently review the basis for what is presented, rather than relying primarily on the software's output. This company's design meets that condition unusually directly: every output carries a citation clickable back to the line in the record that produced it, which is precisely the reviewability the exemption contemplates. The traceability is therefore not only a governance feature, it is load bearing for the regulatory position.

What follows for a buyer. If the product line extends further into recommending or prioritising care rather than summarising and citing what is already documented, the analysis changes. Ask what the company's position is on that boundary, and whether the citation and review path is preserved in every module including any newer ones.

Among the better governance records in this lane, and notable for addressing the one question the lane otherwise avoids.

The domain relevant risk for any trial matching system is that it inherits the record's own coverage. Candidates are surfaced from documentation, so patients with thin documentation are less matchable regardless of actual eligibility, and thin documentation correlates with populations trials already under enrol. A system can therefore raise match volume while reproducing under representation, and the volume figure will read as success. Almost every vendor in this category reports the volume figure alone.

This company's published work engages the question directly. One study is framed on equity in access through automation, reporting an increase in treatment trial accruals. Another examines unrealised enrolments following matching, which is a study of the company's own failure mode: matches that were made and did not convert. Publishing an analysis of where your output does not produce the outcome it promises is rare in any lane of this index. The company also states that it partners with academic cancer centres to develop generative AI performance and safety benchmarks, which places benchmark design with parties who can use it against the product.

Traceability is built rather than asserted. Every output carries a source citation clickable back to the line in the record that produced it, and audit trails are described as native rather than added. A governance property implemented in the product is worth more than one described in a document.

Held at B rather than A for three reasons. There is no model card for the underlying models, no formal governance framework and no statement of intended and unsuitable use. And while the equity framed study exists and is pointed the right way, the specific comparison this axis asks for, the demographic profile of candidates surfaced by the system set against those found by manual screening at the same sites, could not be confirmed from the material retrieved. A buyer should ask for that analysis by name.

The architecture is specified to a level that lets a technical buyer reason about it, and the specification is peer reviewed. The system is described as a constellation of eight specialised models spanning a stated parameter range, operating agentically at the patient level, and the count arrives with the range and a rationale rather than alone.

That distinction matters and this index has been strict about it elsewhere: a bare count of models or parameters is a scale signal that reads as sophistication while remaining unfalsifiable, whereas a count paired with sizes and a division of labour is a description, because a reader can work out why smaller models handle narrow extraction and larger ones handle synthesis, where inference runs and what it costs.

Peer reviewed validation in a named journal rather than a vendor whitepaper substantiates the approach, so the design can be examined by people with no commercial relationship to the company. Held below the top grade because nothing attaches commercially: no warranty, indemnity, service level or remediation commitment was located, and no error characteristic is published for the deployed configuration as distinct from the validated one.

The failure direction worth pressing is the patient never surfaced, since an eligible patient the system does not flag generates no artefact and never learns a trial existed. Ask for recall against a manually reviewed cohort, and per model performance on the extraction steps your protocol depends on.

The deepest integration position in this category, and the decisive element is where the output lands rather than where the data comes from.

Most tools in this lane read from the record and present findings in their own interface, leaving a coordinator to work in two systems. This one delivers matches as an electronic health record in basket message, in the clinician's existing inbox, alongside everything else demanding attention, or in its own worklist if a team prefers that. Writing into the destination workflow rather than beside it is what this axis is actually asking about, and very few records in this index can claim it.

The read side is correspondingly specific. Integration is reported to use recognised interoperability standards to pull demographics, pathology reports and biomarker data in real time from the two dominant electronic record platforms by name, with the model reported to sit natively within the record system's clinical data warehouse rather than in a separate environment. The system also reads unstructured notes, which is where most eligibility criteria are actually documented and where structured extraction alone fails.

Re evaluation is continuous rather than batch. Eligibility is reassessed whenever the chart changes, so progression, a new biomarker or a change in line of therapy surfaces the patient at the moment they qualify. That requires a live connection rather than a periodic extract, which is a materially harder integration to build and operate.

One qualification on sourcing. The in basket delivery and worklist are stated on the company's own pages. The specific standards, the named record platforms and the data warehouse residency come from trade coverage rather than from the company's own documentation, so confirm the integration architecture in writing.

Substantively answered on architecture, unanswered on the specifics a security review will ask for.

The reported architecture places the models natively within the record system's clinical data warehouse rather than in a separate vendor environment, drawing data through recognised interoperability standards in real time. Institution tuned models reinforce that reading, since tuning to a specific institution's documentation implies operating close to that institution's data rather than pooling across customers. For a product reading complete oncology charts continuously, that is the right shape and it is a stronger position than a hosted service with a residency promise.

What is missing is the detail that turns a shape into a commitment. Nothing published states whether model inference executes inside the institution's boundary or in the company's own infrastructure, which regions serve which customers, whether tenancy is dedicated or shared, what the retention period is for anything the system holds, or which subprocessors are involved. Nor is there a statement of what the deployed system transmits outbound, which is the question that matters most when a component sits inside a customer's data estate.

One qualification on sourcing. The data warehouse residency and the interoperability standards come from trade coverage rather than the company's own documentation. Confirm the deployment topology in writing rather than relying on the reported description, and ask specifically where inference runs.

No public pricing. Contact the vendor. Enterprise agreements with cancer centers, with a separate life sciences partnership line. Note the strategic investor: Precision Health Informatics, a wholly owned subsidiary of Texas Oncology, participated in the Series B, and a customer investing while extending its own use is a stronger signal than a logo.

Deliberately confined to oncology and tuned for it, which is the point rather than a limitation: oncology charts are unusually long and narrative heavy, trial criteria are unusually strict, and a general model performs worse on both. Workflow coverage has broadened sensibly along one dimension, from trial matching into visit preparation, registry abstraction, and real world data generation, all of which depend on the same underlying chart reading.