Edison Scientific

The agent is the product. Kosmos, Edison's AI scientist, takes a research objective and a dataset and runs for up to 12 hours, reading literature, executing analysis code, generating and branching hypotheses and steering the investigation mid run, then writes a cited report; the company's November 2025 announcement describes about 1,500 papers read and 42,000 lines of analysis code in a typical run.

The enterprise offer adds forward deployed scientists who work alongside the customer's team and custom world models built on the customer's data, but those exist to run and adapt the agent. There is no non AI version of what is sold.

The oversight structure is described and its limits are stated; what is missing is any point at which the system stops. Kosmos runs unattended for up to 12 hours, and review happens after the run: every statement in the report is cited to code or literature, and conclusions can be traced through the platform. The company's own guidance tells users what that review has to catch: Kosmos "often goes down rabbit holes or chases statistically significant yet scientifically irrelevant findings", longer runs are more likely to go astray, and the team runs it several times on the same objective to sample different avenues. The launch post compares it to a reagent kit run on high value targets rather than a chatbot.

No checkpoint, budget limit or condition that halts a run or brings in a person mid run is published, and nothing describes how findings move from a report into a customer's decisions.

The approach is published in detail; the models underneath are not named. The technical report and the November 2025 announcement describe the architecture: a structured world model shares information between a data analysis agent and a literature search agent, letting Kosmos pursue one objective over about 200 agent rollouts and tens of millions of tokens, with every statement cited to code or literature. The company traces the lineage from ChemCrow, PaperQA and Aviary to Robin and now Kosmos.

The home page calls the platform "model agnostic" and says it uses "the best frontier models", but the language models Kosmos runs on are not identified, their versions are not given, and no change log tells a customer when a model was swapped; the report itself notes a finding that predated the models' training cutoff, which is exactly why model identity matters.

The architecture is described and no model provider is named. Kosmos runs on outside frontier language models ("model agnostic", "the best frontier models"), and the company does not say which providers receive the prompts, datasets and literature a run sends them, or on what retention and training terms. The privacy policy (effective 30 October 2025) lists its other recipients by category only (cloud hosting, email, customer service, content delivery and other technology providers) and names Stripe for payments.

The enterprise option deploys in the customer's environment, which removes Edison's hosting from the chain, but it is not stated whether the models run there too or are called from outside.

Published research short of independent validation. The company's technical report on Kosmos (arXiv, November 2025, written by Edison and FutureHouse staff with academic collaborators) describes seven discoveries: three reproduced findings from unpublished or preprinted work that Kosmos did not access at runtime (nucleotide metabolism in hypothermic mouse brain, humidity during annealing as the key factor in perovskite solar cell efficiency, and neuronal connectivity rules, the last with the caveat that the preprint predated the models' training cutoff), and four new contributions, including a proposed role for reduced flippase expression in entorhinal cortex neurons with age that the company checked in a human Alzheimer's single cell dataset and says it is validating in the lab. Independent scientists judged 79.4 percent of report statements accurate; beta users estimated a 20 cycle run at 6.14 months of their work on average across seven scientists, a survey figure the company itself calls "intrinsically suspect".

No candidate has entered preclinical or clinical development on the platform's output, and the report is a vendor led preprint. Named enterprise partners are Incyte (May 2026) and Population Health Partners (June 2026).

A categorical commitment on the customer's scientific data, without a retention schedule. The privacy policy (effective 30 October 2025), which covers the platform's inputs and outputs, says Edison does "not and will not intentionally train on any user's uploaded data, such as any scientific data that may be uploaded". It does train on what it calls trainable data, meaning agent trajectories and user interaction data, unless the user opts out in the account. Data is transferred to servers in the United States, and aggregated or de-identified data may be used for analytics. For enterprise customers the home page offers deployment in the customer's own environment, where the customer retains governance and "can retain model training only as it serves you".

Retention is stated only as "as long as reasonably necessary", with no period, and trajectories can embed the substance of a customer's analysis even where the uploaded files are excluded from training. Safety engineering is not addressed. The core product works on literature and research datasets; a partner's statement about learning from "experimental and clinical data" means clinical material can enter enterprise deployments.

A privacy document reaches the service, and no HIPAA position is taken. The privacy policy (effective 30 October 2025) covers what users upload to the platform, including personal data in inputs, with US hosting, de-identified analytics and deletion rights. It does not mention HIPAA, protected health information or business associate status, and nowhere does the company say the product does not process protected health information.

The buyer is usually a research organization rather than a covered entity, and the core product works on literature and research data, but a partner's chief scientific officer describes the system learning from "experimental and clinical data", so identifiable clinical data can reach enterprise deployments.

A recognized attestation is named, with the evidence held in a trust portal rather than on the site. The home page says "We are SOC2 Type II compliant", and the footer links a trust center at trust.edisonscientific.com, run on Vanta, which may hold the SOC 2 report and other documents for download. The privacy policy describes commercially reasonable administrative, organizational and technical safeguards.

Running agents that execute arbitrary analysis code on customer data is a real security question for a pharma buyer protecting unpublished targets and assay results.

No device claim is made and the product is scoped as research software. Kosmos generates hypotheses and analyses for scientists and makes no diagnostic or treatment recommendation, so no clearance is expected and none is claimed.

The home page now says Kosmos supports the drug development lifecycle "from target identification through clinical development and filing". Output used to support a regulatory submission falls under FDA's approach to AI used in drug development, where the sponsor must establish the model's credibility for its context of use; that obligation sits with the sponsor, and the company states no position on it.

Evaluation work and candid limitation statements, with no governance program behind them. The company publishes its own evaluation of Kosmos: a technical report (arXiv, November 2025) in which independent scientists judged 79.4 percent of report statements accurate, blinded evaluation of shallower runs, and a benchmark it built for measuring AI in biology research (LABBench2, February 2026). Its founders state plainly that Kosmos "often goes down rabbit holes or chases statistically significant yet scientifically irrelevant findings", that the risk grows with run length, and that they run it several times on the same objective.

None of this is a stated governance framework with monitoring commitments, and nothing evaluates systematic bias in hypothesis generation, such as over-weighting well studied genes, heavily published fields or the literature the underlying language models saw in training; the report notes that one reproduced finding predated the models' training cutoff. Safety engineering is not addressed beyond these caveats.

Mechanisms to check an output exist, with nothing standing behind it. Every statement in a Kosmos report is cited to the code that produced it or to the primary literature, and the company says each conclusion can be traced through the platform to specific lines of code or passages, so a scientist can challenge any claim. The company also publishes an accuracy figure, 79.4 percent of report statements judged accurate by independent scientists in its November 2025 technical report, but neither the announcement nor the report's abstract states how many statements were judged, so it does not stand as an error rate with its denominator.

No warranty, indemnity, remediation commitment or credit for a failed or misleading run is published, and the company's own caution is that runs can chase statistically significant but scientifically irrelevant findings. For a sponsor paying per run and acting on the findings, the cost of a wrong lead falls on the sponsor.

No clinical record integration exists: Kosmos is a research platform, and no clinical record system is involved. For a pharma or biotech buyer the record systems that matter are its research systems: electronic lab notebooks, laboratory information systems, research data platforms and cloud data environments.

Kosmos works on research literature and on datasets the user supplies, and the enterprise offer builds "custom world models" on the customer's experimental history and proprietary scientific data, "integrating information across your organization".

That is integration claimed with no system named: no electronic lab notebook, laboratory information system, research data platform or cloud data environment is named as connected, and nothing describes how customer data reaches the platform beyond upload and enterprise deployment. Scientists interact with Kosmos through Slack, Teams and email, which carry results to people rather than connecting a system of record. The direction of any data flow, and where it runs in production, are not described either.

Options and residency are stated, with the processing path left open. Two options are described: the hosted platform, whose privacy policy (effective 30 October 2025) says data is transferred to servers in the United States, with standard contractual clauses for European users; and an enterprise deployment "in your environment", where the customer retains governance and control over model training.

What is not documented is where the language models run. The home page says Kosmos is "model agnostic" and uses "the best frontier models", without naming them or saying whether calls to them leave the customer's environment in the enterprise option or leave the United States in the hosted one. Tenant separation on the hosted platform is not described.

A price and a pricing basis are published, without full tiers. The Kosmos announcement of 5 November 2025 launched it at $200 a run (200 credits at $1 a credit), with free usage for academics and founding subscriptions that lock in the $1 credit price, and says the price "ultimately will probably be higher"; users sign up on the platform without contacting sales, and enterprise deployments are quoted separately. That is a launch price the company expects to raise, not a current tier list with figures. The company no longer keeps a pricing page, and today's tiers and credit allocations are not published, though the Kosmos signup page on the platform could carry them. Even so, the cost of a run is clear before any conversation with sales.

The enterprise offer, with deployment in the customer's environment, forward deployed scientists and custom world models, carries no published price.

Coverage is named, with validation behind part of it. The technical report's discoveries span metabolomics, materials science, neuroscience and statistical genetics, and the home page positions Kosmos across the drug development lifecycle from target identification through clinical development and filing, for biopharma organizations and academic researchers.

Validation sits with the research end: the discoveries are target and mechanism level findings, and the named enterprise partnerships (Incyte, Population Health Partners) are recent, with no published results. Nothing shows the platform in clinical development or regulatory filing work yet, so those stages are claimed rather than validated.