BenchSci

EMET is a research environment built around learned scientific reasoning. It breaks a research question into smaller tasks, chooses scientific tools and combines results with a biomedical knowledge graph. Removing those models would remove the product's main ability to connect evidence, propose explanations and plan follow up analysis. BenchSci's earlier antibody selection tools and ASCEND platform established the same focus on extracting experimental knowledge.

The present offer extends that work into agent workflows and computational analysis. Dedicated scientists help implement the software, but the commercially licensed product performs the research operations rather than simply supplying a consulting team or access to computing infrastructure.

Scientists direct EMET through research questions and can inspect the supporting evidence. Its orchestration layer chooses tools and performs sequences of research tasks, while the privacy policy describes human review when monitoring identifies use outside the scientific remit. Those controls do not provide complete scientific confidence thresholds, laboratory execution approvals or rollback rules.

The home page also makes an unqualified claim that EMET does not hallucinate, separately from its numerical accuracy statement. That assertion concerns generated answers, not workflow speed. It creates tension with the terms' explicit possibility of incorrect outputs and with the need for scientists to verify consequential findings.

The system combines a biomedical knowledge graph, generative reasoning and specialized scientific tools. The home page names ESM 2, AbLang2 and RDKit alongside a changing set of frontier models, while the security FAQ identifies Anthropic as the primary AI model provider. Public and licensed literature, experimental resources and customer research can all supply context. This is more specific than an unnamed proprietary engine, but it is not a fixed model inventory.

Routing can change with the task and available providers. Version identifiers, release qualification criteria and advance notification of model substitutions are not specified for a particular customer deployment.

Anthropic is identified as the primary model provider, and the privacy notice also lists Google Cloud Platform for model inference. Azure and Amazon Web Services appear as infrastructure providers, alongside other service subprocessors. Internal data is excluded from shared model training, and identifiable personal information requires consent for external model training. The full material subprocessor list is available by request.

Specific provider retention periods, deployed model versions and the routing of each type of research input are not mapped publicly. The named model layer makes the chain partly visible, while the combination of changing engines and global inference endpoints leaves a customer specific processing path unresolved.

The October 2026 case involving University of Cape Town researcher Melissa Chengalroyen describes two weeks using EMET for tuberculosis research, a review article, a grant and gene essentiality questions. A separate case names Kyle Medak and describes finding mouse strain and dose evidence for kidney experiments. These accounts give identifiable users and concrete tasks, but their time savings are retrospective impressions without a measured comparison.

Sanofi's October 2025 agreement establishes a three year ASCEND license, not an EMET outcome study. The current accuracy claim exceeds 95% across hundreds of tests, without a reproducible test set or independent evaluation accompanying it.

Internal research data is isolated by organization and excluded from shared model training. The privacy notice separately requires consent before identifiable personal information trains external models, while allowing certain internal research uses. Routine retention follows purpose and legal need rather than a fixed deletion schedule; records associated with safety incidents have a minimum five year period.

Automated monitoring can flag inappropriate use for human review and access restriction. These controls address misuse, but there is no complete account of how erroneous scientific conclusions trigger correction across prior outputs. A general promise of confidentiality does not remove the distinction between proprietary research and identifiable participant data.

BenchSci operates from Canada and states GDPR, CCPA and applicable US privacy commitments. Its June 2026 privacy notice reaches the applications, agents and platform rather than only the marketing website. It explains controller responsibilities, international transfers, access rights and restrictions on service providers. That creates a substantive privacy position for the research service.

A US business associate agreement and a specific HIPAA service scope are not identified in these materials. Scientific use alone does not establish that every customer dataset excludes identifiable information, particularly when internal clinical trial data can enter the product. The documented position is its broader privacy regime.

BenchSci states that it holds a SOC 2 Type II attestation. Enterprise customers can request the report from their account team or privacy contact under a nondisclosure agreement; the public badge links to explanatory SOC material rather than the audit itself. The current security page also specifies encryption, access logging and scoped integrations. Those controls describe the service but do not reveal the auditor, report period, exceptions or systems included in the examination.

GDPR and CCPA commitments address privacy obligations rather than substitute security attestations. The report access arrangement provides a procurement route, with the detailed assurance scope remaining behind that process.

EMET's clinical development FAQ explicitly limits use to sponsor teams researching investigational compounds, trial populations and program safety. It excludes use at the point of care and treatment recommendations for individual patients. This boundary is relevant because sample workflows assess interactions, biomarkers and adverse event signals that could otherwise resemble clinical decision support.

The general terms also restrict the service to scientific research and exclude medical advice. No device clearance is attached to the current research offer. References to FDA labels, FAERS and approved medicines identify evidence inputs; they do not imply that the EMET platform itself has regulatory authorization.

A safety monitoring program examines interaction patterns and content for use outside the research remit. Flagged interactions can pass to a safety and compliance team, then through executive and board governance points, with user notifications or suspension among the responses. This is a described organizational process for handling model related incidents.

The scientific quality system also combines knowledge graph checks with expert calibration, but detailed performance across disease areas, experimental modalities or poorly represented biological targets is not available. Security compliance does not establish scientific fairness. The incident process gives review structure without showing how consistently the research models perform across the populations of evidence they analyze.

Source references and inspectable research findings let scientists challenge a result before using it. The kidney experiment case also shows a user questioning an apparent gap and obtaining additional literature. Those are practical review mechanisms, not promises to compensate for an incorrect conclusion. General terms disclaim accuracy and any assurance that all errors will be corrected; the customer provides the stated indemnity. Separate enterprise contracts may change those terms.

No public obligation specifically requires remediation of an erroneous research output, replacement of a failed experiment or coverage of downstream development costs. Traceability helps review but does not allocate those losses to BenchSci.

EMET connects to research systems rather than primarily to hospital records. The July 2026 internal data release names live connections to Benchling, Veeva, electronic laboratory notebooks, D360 and GDB, alongside document uploads and a private knowledge graph option. It distinguishes reading existing systems from ingesting material into that graph. These are identifiable research data pathways, although a named customer implementation guide and evidence of writing results back are absent.

Epic, Oracle Health and clinical FHIR workflows are not part of the stated connection set. For this product, interoperability chiefly concerns making institutional experimental evidence available to scientific reasoning.

The service runs in cloud infrastructure from Google, Microsoft Azure and Amazon Web Services. Its security FAQ identifies the United States as the primary hosting location and says inference may run outside the country depending on endpoint availability. Internal data can remain in connected systems or enter a private organizational graph, with access following existing permissions. This gives both a deployment approach and a meaningful cross border caveat.

It does not commit each inference request to a named region or enumerate all transfer locations. Enterprise residency requirements are handled individually, and no standard installation option on customer premises is described.

Academic access is free for eligible academic, nonprofit and government users, with a limited feature set and institutional verification. Enterprise access is custom priced and starts with a free pilot. The enterprise package includes broader content access, internal data connections, advanced reasoning, team controls, onboarding and dedicated support. General terms permit subscription or usage charging but leave the actual schedule to the selected agreement.

The free academic offer therefore has a stated entry cost, while an enterprise buyer cannot calculate a production budget from the plan page. Contract duration, user allowances, usage limits and separate implementation charges remain unspecified.

The product serves biopharma discovery and development teams, alongside academic researchers. Its workflows cover disease biology, molecular design, biomarkers, experimental planning, safety research and related portfolio decisions. The Cape Town tuberculosis example and the kidney experiment case support actual use in two research settings.

Sanofi's ASCEND license extends to its global preclinical organization across therapeutic areas, although that agreement does not validate every newer EMET workflow. Clinical development examples concern investigational programs and population evidence. They do not establish performance in direct patient care, and no comparable validation accompanies every chemistry, translational or regulatory workflow in the current catalog.