Lila Sciences

Nothing here exists without the models. The company describes its system as an operating system for science in which models generate the hypotheses, proprietary instruments it calls AI Science Factories design and run the experiments, and results feed back in real time. Its own framing puts the model at the center and the robotics in service of it: the laboratories are described as the world's largest verifier for science, existing to supply the reward signal a reasoning model needs to learn.

There is no consulting business, no contract research offering and no product that would survive the removal of the machine learning. Both commercial routes are the model: Catalyst applies it as an operating layer to a partner's existing research, and Creation uses it to originate discoveries and found companies. Whatever else is uncertain about this company, that the artificial intelligence is the whole proposition is not in question.

The oversight structure is described and, unusually, the degree of autonomy is quantified. The September 2026 preprint characterizes the platform as an artificial intelligence guided, human supervised closed loop at over 90 percent automation, which tells a reader both that a person is in the loop and roughly how much of the work is not theirs. The loop's components are named: combinatorial synthesis, high throughput screening, machine learning composition property models, adaptive multi objective optimization and language model reasoning, with lead candidates advanced to long term validation under stated conditions.

The human role is described concretely rather than asserted. The company's own account of the catalyst campaign has two senior scientific leads reviewing the model's output, judging it wrong, and being overruled by the experimental result, which is a real description of where authority sits: the model proposes, people review, and the experiment decides.

What is missing is everything a buyer could hold the company to. Nothing published states at what confidence a proposal is not run, what a contradicting result does to the model, which decisions require a person rather than allowing one, or what the remaining share of non-automated work consists of. The platform is described elsewhere as running the scientific method at scale and as reaching insights humans would not reach on their own, and no stop condition is published anywhere alongside that.

The company describes its architecture in general terms and identifies nothing. The technology page sets out the approach at the level of a thesis: a scientific reasoning model trained on what the company calls evergreen tokens generated by its own experiments, a generalizable reinforcement learning engine powered by verifiers and tools, autonomous laboratory workflows supplying real world training rewards, and continuous policy optimization. The tools the model chains are named by category, including molecular dynamics simulators, protein structure predictors, quantum chemistry solvers, gene editors and robotic workflows, and the domains are listed from DNA and proteins through to catalysts and coatings.

None of it identifies anything. No model is named, no architecture or model class is given, no parameter count, no training corpus size, no benchmark and no version or update practice. Scientific Superintelligence and AI Science Factory are presented as trademarks rather than as descriptions. The September 2026 preprint names the components of one campaign system, including machine learning composition property models, adaptive multi objective optimization and context aware language model reasoning, which is more specific than anything on the company's website. It describes a campaign rather than the production platform, though, and it does not say which language model was used.

The model provider goes unnamed, though the architecture itself is described. The June 2026 collaboration with Nvidia identifies the BioNeMo agent toolkit as an external framework the company builds on, and the Department of Energy work names the national laboratory and university institutions involved, so some parties are on the record.

The model layer is not. The company says it is training its own scientific reasoning model, and its September 2026 preprint describes context aware large language model reasoning inside the campaign loop without naming which language model, so an unnamed third party model sits in the chain rather than the chain being a short in house one. No foundation model base is named for the large language models the company says it has built, no training data provenance is given, there is no open source inventory, and nothing states whether the scientific models are trained from scratch or adapted. No compute or hosting provider is stated beyond what the Nvidia relationship implies.

Nvidia appears both as a technology partner and, on press reporting, as an investor, so it is not an arm's length supplier.

The evidence and the claims sit in different domains, and the open question is whether the platform has been shown to work on medicines.

On therapeutics, the work that brings it into health care, the company publishes claims and no evidence. It states that its platform has produced thousands of discoveries, genetic medicine constructs performing better than commercially available therapeutics, hundreds of antibodies, peptides and binders, and, in an August 2026 blog, an ultra stable messenger RNA and a CAR-T therapy that outperformed a leading therapy used in patients today in non human primates, designed by three scientists in a matter of months. No paper, preprint, dataset or trial supports any of it, and the account circulated through the company blog and social media rather than a scientific venue. These are vendor reported results with no method, denominator or reference standard. Unpublished primate data is not unusual at this stage, but the claim is large and nothing checkable stands behind it. No molecule from this platform is in a clinical study, and no customer or commercial partner is named in any of the seven industries.

Outside medicine the evidence is real. A September 2026 preprint reports an AI guided closed loop campaign that proposed, synthesized and screened 2,942 catalysts across 53 material systems and 26 elements in three months, identified six high performing material families, and carried leads through more than 1,000 hours of stability testing against a named industry standard, with degradation times published for each. The method is set out and the company's search agent is benchmarked against fixed policy Bayesian optimization and in context language model selection. Twenty two authors are named, including scientific leads with Caltech and MIT affiliations. In July 2026 the company also received Phase I awards under the United States Department of Energy Genesis Mission for three projects alongside Caltech, Lawrence Berkeley National Laboratory, Northwestern and Argonne, a competitive federal selection rather than a self assessment. Both are materials and energy work: they show the platform's closed loop working with a stated method in another field, and they say nothing about whether its medicines work.

How information moves through this system is not published. The only privacy document is a website notice which states that it applies to information collected on the website or in electronic messages with it, and that it does not apply to information collected offline or through any other means. It reaches nothing the platform processes.

No protected health information is evidently at stake, though the company's own claims leave that open: it states that the platform has produced genetic medicine constructs and a CAR-T therapy tested in non human primates, and it names therapeutics and biotech as two of its seven industries. Nothing published says what data those programs run on, where it came from, or how it is governed. The commercial model sharpens the question further: the Catalyst route is described as an operating layer applied to a partner's existing research programs, so partner data would flow into a platform whose entire design is that every experiment becomes training signal for the whole network, and nothing states whether a partner's data is retained, used to train models serving other partners, or separated at all.

Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, is not published.

No status is stated and the only privacy document rules itself out of everything but the website. It says it applies to information collected on the website or in electronic messages with it, and that it does not apply to information collected offline or through any other means. Protected health information is not mentioned and no business associate position is taken.

The notice does address the European and United Kingdom data protection regimes, so the company knows how to write a regime specific position and has written one for visitors rather than for the platform. Therapeutics and biotech are two of the seven industries it sells into and it claims a CAR-T therapy tested in primates, so the question of whether human derived data reaches this platform is live and unanswered. No business associate status is stated, and no plain statement scopes the platform out of protected health information with a service level document behind it.

Nothing is published. There is no security page, no trust center, no certification and no attestation anywhere across the site or its footer, which carries a privacy policy and terms of service and nothing else. No SOC 2 report, ISO/IEC 27001 certificate or independent penetration testing on a stated cadence is named, and the privacy notice offers no security commitment reaching anything but the website.

The absence is conspicuous given the company Lila keeps. It holds Phase I awards under a United States Department of Energy program alongside national laboratories, markets an aerospace and defense vertical, and sells a Catalyst route described as an operating layer applied to a partner's existing research programs, which would place partner science inside its systems. Federal and defense adjacent work normally comes with a published security posture as a matter of course. Whatever exists internally, none of it has been made visible.

The platform makes no device claim and is scoped accordingly. The models and the autonomous laboratories are not offered as a medical device and carry no clearance, which is the appropriate position for what they are.

There is nothing to clear. No candidate is named, no investigational application is disclosed, no clinical program exists, and the therapeutic work is preclinical and, on the published record, not assigned to a named indication. The company's strongest therapeutic claim, a CAR-T therapy said to outperform a leading therapy used in patients today in non human primates, carries no regulatory position, no program name and no filing.

One further exposure is not about medicines at all. The Department of Energy Genesis Mission awards and the aerospace and defense vertical bring this company within federal contracting and export control regimes that sit entirely outside the medicines framework, and nothing published addresses its posture under either.

Something checkable now exists, and it arrived in September 2026. A preprint describes an AI guided, human supervised closed loop platform at over 90 percent automation and reports what it did with denominators attached: 2,942 catalysts evaluated across 53 material systems and 26 elements, surfacing complex oxides that the paper states conventional design logic and off the shelf language models would not predict. It includes a retrospective benchmark of the company's own sequential learning agent against fixed policy Bayesian optimization and against in context language model selection, which is an evaluation of the model's behavior against named alternatives rather than a claim about it.

Where candidates fail is published rather than omitted. Two of the three lead materials eventually exceeded the failure threshold during long term testing, at roughly 200 hours and roughly 470 hours, while the third held below it for 1,000 hours. Reporting the hours at which its own candidates degrade is a candid disclosure.

The limits are scope and independence. The evaluation covers one campaign in materials chemistry; nothing comparable is published for the therapeutic work, where the claims are the largest and the evidence is a company blog. There is no model card, no governance framework, no responsible use policy and no independent audit of model behavior anywhere, and the preprint has not been through peer review.

Mechanisms exist that let an output be challenged, and one of them is now visible in the published record. The closed loop is human supervised at over 90 percent automation, every proposal is physically synthesized and screened before it counts, lead candidates go through long term validation under stated conditions, and the September 2026 preprint publishes the campaign's method, its denominators and the hours at which its own lead materials degraded. A reader can check the work rather than take it.

Nothing stands behind the output. No accuracy or error rate is published for the models themselves with a method and a denominator; the figures that exist describe a campaign's yield, 2,942 candidates screened to six material families, rather than how often the platform is right. No guarantee, indemnity or remediation obligation is published, and with no customer or commercial partner named in any of the seven industries there is no counterparty relationship on the public record for a commitment to run toward. On the therapeutic side, where the claims are largest, there is nothing published at all to challenge.

No clinical record integration exists. There is no electronic health record touchpoint, no clinical workflow surface and no provider connection, because the platform runs experiments across seven industries rather than serving health systems.

On the research systems of record a partner would run, nothing is named either. No electronic laboratory notebook, laboratory information management system, research data platform or partner cloud environment is named as connected, and no connector or application programming interface is offered. The company does describe standardizing the way its agents talk to instruments and data as more of its autonomous facilities come online, which is an internal interoperability program rather than a buyer integration, and no standard, specification or third party adoption is named for it. The Catalyst route is described as an operating layer applied to a partner's existing research programs, which would require reaching into a partner's systems, and no partner, system or method of doing so is disclosed.

The platform runs for Lila rather than being deployed to anyone, so there is no tenancy model, hosting option or isolation commitment to evaluate. Location is implied rather than committed: the company operates from Cambridge, Massachusetts, San Francisco and London, and describes its autonomous laboratories as an extensible network it is bringing online facility by facility.

Nothing published states where data rests. No cloud or compute provider is named, no residency or cross border position is given despite sites in two jurisdictions, and no segregation statement covers how one engagement's data is kept from another's. The Catalyst route makes that the live question: it is sold as an operating layer applied to a partner's existing research programs, on a platform whose stated design is that every experiment anywhere in the network becomes training signal for the whole system. Nothing reconciles those two, and no retention position exists for anything a partner contributes.

Well disclosed on capital and silent on customers. Investors are named at length across both rounds, including Flagship Pioneering, General Catalyst, March Capital, the ARK Venture Fund, a subsidiary of the Abu Dhabi Investment Authority, the State of Michigan Retirement System, Braidwell and Collective Global, with the rounds dated and the leads identified. Two commercial routes are named and described in outline: Catalyst as an operating layer applied to a partner's existing research programs, and Creation as a longer horizon partnership intended to originate discoveries, launch products and found new companies.

Against that, not one customer or commercial partner is named in any of the seven industries, no contract, deal value, upfront, milestone, royalty, equity or revenue share arrangement is described, and no figure attaches to either route. The Creation route in particular implies a founding or equity stake model that is never stated.

The disclosed figures also do not add up. A $200 million seed plus a $235 million Series A is $435 million against a widely reported total near $550 million, and the valuation above $1.3 billion and the Nvidia investment come from press reporting rather than from the company, so the two most quoted facts about this company's scale are not its own disclosures. Some income arrives as competitive federal award funding under the United States Department of Energy Genesis Mission.

Within healthcare the modality coverage claimed is unusually wide. Under Therapeutics the company names messenger RNA, proteins, antibodies, cell therapies and small molecules, and it carries a separate Biotech vertical for turning biological data into research decisions. That breadth is the direct consequence of selling a general scientific method rather than a domain tool, and the same logic carries the platform into five further industries: advanced materials, chemicals, oil and gas, energy and environment, and aerospace and defense, all given equal billing in the site navigation.

Validation reaches one corner of it. The published catalyst campaign is in materials and energy, with a preprint, denominators and long term testing behind it, and the Department of Energy awards cover three projects in that direction. Nothing comparable exists on the therapeutic side.

Breadth of claim is not depth of practice. No program, indication, target, candidate or partner is named in any therapeutic modality, the one named therapeutic result is a blog account of primate data, and the same platform is simultaneously offered to five industries outside life science, which spreads the evidence rather than concentrating it.