Edison Scientific

The agent is the entire product. Kosmos is described as autonomously reading literature, executing analysis code, generating hypotheses, branching investigations, and steering mid run, with independent coverage citing roughly 200 agent rollouts and tens of thousands of lines of executed code within a single campaign lasting up to twelve hours. There is no non AI version of this offering.

This is one of the most autonomous systems in the index, running unattended research campaigns for hours, which raises rather than lowers the oversight burden. Two things earn credit. The output is fully cited, so every claim traces to a source a scientist can check, and the workflow is stated to be fully auditable with analyses, reasoning steps, and decisions traceable.

More notable is the founders' own published caution that the system sometimes goes down rabbit holes or chases statistically significant yet scientifically irrelevant findings, and that they often run it multiple times on the same objective. That is an unusually honest limitation disclosure and it tells a buyer exactly what the review obligation is.

The architecture is described in public technical writing rather than asserted: a structured world model coordinates agents and maintains coherence across hundreds of steps, with a published technical report and engineering write ups on how the platform was built. An open source implementation is reported to exist allowing self hosting across multiple model providers, which is the strongest form of transparency available and rare anywhere in this index. Underlying foundation model providers are not named on the commercial platform.

Self hosting changes what this axis is asking, and it is the reason for the grade. An open source implementation is reported to exist that can be run across multiple model providers, so an organisation with a reason to care can eliminate the chain rather than audit it: they choose the provider, they hold the deployment, and no enumeration is needed because there is no third party estate to enumerate.

That is the same structural answer this index credited where a de identification proxy could be self hosted, and it is stronger than any published sub processor list because it removes the dependency instead of documenting it. The core product also carries no patient data surface, operating on research data and literature, and data is stated to be encrypted in transit and at rest with account deletion rights.

Held below the top grade on the commercial path, which is the one most customers will actually take. Foundation model providers are not named on the commercial platform, so a customer using the hosted service is back to the ordinary position of not knowing who processes their queries.

And where partners feed translational and clinical data into the platform, the governing controls are enterprise deployment terms rather than anything published, so the arrangement that matters most for clinical material is the one a reader cannot see. Ask which providers serve the hosted platform, and for the partner deployment terms.

Evidence is research productivity rather than clinical outcome, which is the right measure for this product. The company reports collaborators finding a single 20 cycle run equivalent to roughly six months of their own research time, and a named partner describes 292 independent analysis trajectories in 24 hours surfacing insights their team had missed. Specific scientific findings have been described in the Alzheimer's research literature. What is missing is independent controlled evaluation of output quality, and the vendor's own caveat about irrelevant findings means throughput claims should not be read as accuracy claims.

No PHI surface in the core product, which operates on research data and literature. Where partners feed translational and clinical data into the platform, the governing controls are the enterprise deployment terms rather than a published PHI framework. Data is stated to be encrypted in transit and at rest with account deletion rights under the privacy policy.

No HIPAA or BAA commitment located. The buyer is a biopharma R&D organization rather than a covered entity, so this axis is largely inapplicable, but any deployment touching identifiable clinical trial data would need it established contractually.

SOC 2 Type II compliance is stated on the main site, and the company describes enterprise deployment on SOC 2 compliant infrastructure with fully auditable workflows. The engineering write up is candid that supporting agents executing arbitrary code is a substantial security challenge in a regulated environment, which is the correct thing for a buyer to be thinking about. No trust center with downloadable artifacts was located.

No FDA pathway applies. This is a research acceleration platform operating upstream of any regulated claim, and it makes no diagnostic or treatment recommendation.

Graded B on category non application rather than C. The axis genuinely does not reach this business model, there is no substitute vendor level regime standing behind it the way payer accreditation stands behind prior authorisation software, and the company does not overclaim clinical or regulatory standing to imply otherwise. Declining to assert a regulatory position a product does not need is treated here as accurate rather than as a gap.

The boundary worth watching: this holds only while output stays upstream of a regulated claim. A platform whose results begin supporting a submission, a label or a clinical decision would move into a different conversation, and that is the change to re test on.

No formal governance framework is published, but the substance that governance is meant to produce partially exists: the founders publicly document failure modes, the outputs are cited and auditable, and the technical report allows external scrutiny. Credit is limited because none of this is a stated governance program with monitoring commitments, and no evaluation of systematic bias in hypothesis generation was located.

The architecture is described in public technical writing rather than asserted, with a published technical report and engineering write ups covering how a structured world model coordinates agents and maintains coherence across hundreds of steps.

That level of detail lets a reader reason about the specific failure mode of long horizon agent work, which is drift: a system taking hundreds of dependent steps can go wrong early and elaborate confidently on the error for the rest of the run, and a described coherence mechanism is the thing a reviewer would interrogate.

An open source implementation is reported to exist allowing self hosting across multiple model providers, which is the strongest form of transparency available and rare anywhere in this index, because a claim about how a system behaves can be tested by anyone who disagrees rather than only by the vendor. Held below the top grade on three points.

The underlying foundation model providers are not named on the commercial platform, so the open artefact and the commercial product are not the same object and an evaluation of one may not transfer. No evaluation of research output quality was located, which for a system producing scientific analysis is the measure that matters and is genuinely hard to define, so its absence is understandable and still leaves a buyer without one. And no warranty, indemnity or remediation commitment attaches. Ask what differs between the open implementation and the commercial platform, and how output quality is assessed.

No EHR surface. Integration is with internal research tools and datasets on the enterprise tier, and collaboration happens through Slack, Teams, and email rather than clinical systems.

The strongest disclosure on this axis among the vendors added in this batch. The company states enterprise deployment happens in the customer's own environment so the customer retains full governance, and explicitly that model training on customer data is retained only as it serves the customer.

For biopharma organizations whose proprietary experimental data is the asset at risk, an in environment deployment with an affirmative training data position is the material term, and it is stated plainly rather than left to negotiation.

Published pricing in a category where essentially nobody publishes. Tiers are listed with credit allocations, a free tier with 10 expiring credits per month, a paid tier with 650 expiring credits per month including Kosmos access and additional non expiring credits purchasable at one dollar per credit, and a separately quoted enterprise tier. Independent reporting notes academic users received the first six Kosmos runs free at roughly 200 dollars per run thereafter. A buyer can estimate spend before contacting sales, which is the point of this axis.

Domain coverage is genuinely broad, spanning neuroscience, metabolomics, genetics, and materials science, with the platform positioned across biology, chemistry, and physics. The healthcare relevant surface is narrower than that range implies: the buyer is a pharma or biotech R&D organization, with the disclosed enterprise partnership focused on target discovery, validation, and translational biology. This is not a provider facing product and should not be evaluated as one.