Mendel

Clinical reasoning over unstructured records is the entire product. Hypercube abstracts, structures, and answers questions across pathology reports, genomic data, physician notes, claims, and electronic data capture records, and the company's stated design premise is that generic language models are insufficient for this task. There is no services or data brokerage layer; the platform is sold to operate on the customer's own data.

The assisted workflow is tested; the product's own limits are not published. In the Penn randomized evaluation (Nature Communications, 2026) the AI extracted eligibility elements and research coordinators decided, and the study separated that configuration from the AI working alone, which scored 59.9 percent chart-level accuracy against 76.5 percent with a person deciding.

The study found automation bias: staff followed wrong AI answers on performance status and scored worse than without the AI. What the product may do without a person in routine use, how uncertain answers are routed to a reviewer, and how the automation bias the study found is guarded against are not described in anything readable from outside the company.

The architectural claim is specific, unusual, and motivated by a stated failure mode: Hypercube pairs large language modeling with symbolic reasoning over a clinical hypergraph, adopted because the company argues pure neural approaches cannot discern clinical nuance and are susceptible to hallucination. This is the same neuro symbolic reasoning that RAAPID applies in risk adjustment.

Critically, every answer is tied to discrete highlighted evidence in the original record, which the company frames as making the algorithm's failure mode immediately visible rather than obscured. Publishing the reason for an architecture choice and building for inspectable failure is a materially higher standard than describing capabilities.

The customer hosted route is backed by two concrete mechanisms rather than assurances, and the other routes are not answered. The platform is cloud agnostic and can be hosted inside the customer's own environment, so an institution can process clinical records without them leaving infrastructure it controls, and a dedicated capability de identifies protected health information from clinical records, which is a control operating on the data rather than a statement about handling it.

Evidence linking helps too, since a system that records which source text produced each answer leaves an audit trail of what it drew on. What is absent is everything about the routes a customer might actually take. No model or model family is named, no foundation model provider is identified, no hosting arrangement is described for the vendor operated option, and no sub processor list is published, so the customer hosted path is bounded by the customer's own perimeter while the alternative paths are undescribed.

Nothing states retention, or whether customer records contribute to model development, which matters for a system whose reasoning layer sits over a clinical hypergraph that presumably improves as it sees more clinical language.

A randomized, peer reviewed evaluation led by an academic center, with modest results that it states plainly. Parikh and colleagues (Nature Communications, 2026; University of Pennsylvania and Emory investigators with seven Mendel staff among 17 authors; registered as NCT06561217, results posted) randomized 355 retrospective charts of patients with non-small cell lung or colorectal cancer to prescreening by research staff alone, staff with Mendel's neurosymbolic abstraction, or the AI alone, against a gold standard set by two to three blinded clinicians on 12 eligibility criteria.

Staff with AI reached 76.5 percent chart-level accuracy against 71.1 percent for staff alone (noninferior and superior); the AI alone reached 59.9 percent. Time per chart did not change (37.4 against 37.8 minutes). Gains were largest for biomarker, staging and response criteria, and automation bias lowered accuracy on performance status. The charts came from one health system and the trial was not powered for efficiency or subgroups. A second Penn study against the EA8191 prostate cancer trial (NCT06561230) is still running.

Two concrete mechanisms rather than assurances. The platform is cloud agnostic and can be hosted inside the customer's own environment so patient data never leaves, and a dedicated capability de identifies PHI from clinical records. Evidence linking, where each answer traces to highlighted source text, functions as a safety control as well as a transparency one, since an unsupported assertion is visibly unsupported. What is absent is published hallucination rate measurement or formal failure mode documentation.

The company's own pages show an architecture built for regulated data, and they publish no contractual posture.

The product documentation states the platform is cloud agnostic and can be hosted in the customer's own environment so that data never leaves, with vendor hosted options offered as an alternative. Redaction is listed among the platform's stated use cases alongside abstraction, chart review and cohort building. Native availability on a major data cloud means it can operate inside a customer's existing warehouse tenancy.

That matters because it changes which legal question applies. Where the platform runs entirely inside the customer's own environment and no records are disclosed to the vendor, the business associate framework is not engaged in the ordinary way, and the customer's own existing safeguards govern. Where the vendor hosts, it plainly is engaged, and nothing addresses it: no business associate agreement availability statement, no characterization of the company's role, and no identification of which entity would sign. The answer therefore differs by deployment and turns on which configuration is being sold.

A second legal question sits alongside the first and is specific to trial prescreening. Screening records to identify candidates happens before anyone has consented to anything, so the operative instruments are institutional review board approval and an authorization or waiver rather than a business associate agreement. A vendor can be correctly positioned on one and silent on the other. Nothing published says under whose authority the screening runs, who holds the approval, or whether the vendor acts as the institution's agent or on its own account.

No SOC 2, ISO 27001, HITRUST or equivalent attestation is published, and there is no trust center.

What the company does publish is architecture rather than assurance, and it is substantive enough to change what the gap means. Its own product documentation states the platform is cloud agnostic and can be hosted in the customer's own environment so that data never leaves, with vendor hosted options offered as an alternative. It is available natively on a major data cloud, meaning it can run inside a customer's existing warehouse tenancy, and it is listed on a major cloud marketplace under a first party seller profile.

That produces two kinds of customer with two different exposures, and the missing attestation does not affect them equally. A customer running the platform inside its own environment has substituted architecture for assurance: nothing transits, so the question narrows to what the software does inside the boundary, what it transmits outbound, what telemetry leaves, and what the vendor can see during support. A customer taking the vendor hosted option has none of that and no attestation either, which is the weaker position by some distance.

The sources disagree on trial environments. The company's own product page states that trial environments use synthetic data, while earlier trade coverage of the same feature described it as using a real world clinical dataset. A demonstration environment is a recognized place for real patient data to end up inside a sales process, and nothing published settles which description is accurate today or what governs the environment either way.

The product is scoped to chart abstraction, cohort building and trial prescreening, with no device claim made for it. It does not diagnose or direct treatment, and candidates it surfaces are confirmed by a research team. The University of Pennsylvania's registrations of its two studies of the product (NCT06561217 and NCT06561230) record the intervention as neither an FDA-regulated drug nor an FDA-regulated device, which is the study sponsor's statement. Prescreening records before consent is governed by IRB approval and an authorization or waiver, and who holds that at each site is not stated.

Results are analyzed within clinical groups; demographic results are not yet published. The Penn randomized evaluation (Nature Communications, 2026) analyzed accuracy and efficiency by cancer type (lung against colorectal), chart length, complexity and order, with the per-group figures in its supplement, and reported where the AI helped (biomarker, staging, response) and where automation bias hurt (performance status).

The second Penn study, against the EA8191 prostate cancer trial, states the hypothesis that human plus AI prescreening improves the diversity as well as the efficiency and accuracy of prescreening; it is still running. No governance framework or demographic subgroup results are published.

Answers are meant to be checked against the record; nothing stands behind them. The company describes pairing language models with symbolic reasoning over a clinical hypergraph and tying each answer to highlighted evidence in the source record, and in practice a coordinator confirms eligibility.

An outside error figure exists: in the Penn randomized evaluation (Nature Communications, 2026) the AI alone was 59.9 percent accurate at chart level on 12 eligibility criteria against a blinded clinician gold standard, across 355 charts. The company's website was unreachable on 29 Sep 2026, so whether it cites that result is open. No warranty, indemnity or remediation commitment is known.

Integration is oriented toward the analytics stack rather than the clinical one, and is named specifically: upstream and downstream connectivity with data warehouses including Databricks and Snowflake, and business intelligence tools including Tableau and Qlik. The platform ingests structured and unstructured sources spanning notes, pathology, genomics, claims, and electronic data capture.

Named connectors are a stronger disclosure than asserted interoperability, though direct EHR integration is not enumerated, which fits a product operating on warehoused clinical data rather than at the point of care.

The deployment answer is explicit and puts residency in the customer's hands. The platform is cloud agnostic and can be hosted in the customer's own environment so that data never leaves, with vendor hosting available as an alternative rather than a requirement. For an institution unwilling to move patient records to a vendor cloud, that removes the primary objection outright. Try Me environments with synthetic data are offered for evaluation, which lets a buyer test before exposing real records.

No pricing is published. The company does make a comparative cost argument, that domain specific models can outperform generic large language models at a scalable cost, but no rates, licensing structure, or per record economics are disclosed. Evaluation access through synthetic data environments lowers the cost of assessing fit before committing, which is a partial offset.

Broader than trial matching alone. Distinct product configurations address chart review, cohort building and analysis, and trial eligibility mapping against templates that can incorporate prior authorization criteria and NCCN guidelines, serving clinical care, pharmaceutical, and diagnostics customers. Published evaluation work is concentrated in oncology, which is where unstructured record complexity is greatest, but the underlying capability is not disease specific.