The Governed Discoverer
Why we are built differently from everything else calling itself AI for drug discovery, and why the next medicine worth believing in should be the one we can prove
Matt Hardy Founder and Chief Executive Officer, NomosLogic
The last decade proved that artificial intelligence can generate almost anything: a sentence, an image, a protein, a plausible molecule. The next decade will be decided by a far harder question. When the output is a drug and the stakes are a human life, can anyone prove the machine was right? That question, not raw generative capability, is the entire basis of our company. We did not set out to build a cleverer way to guess. We set out to build the first discovery platform whose every claim can be defended, step by step, to a scientist, a regulator, a board, and a patient. We call the result governed discovery, and it is a genuinely different kind of system.
1. The trust problem is the only problem that matters
A large probabilistic model is brilliant and unreliable in the same breath. Ask it a question and it returns a fluent, confident answer whether or not the answer is true, and it returns it in exactly the same voice either way. In a consumer product, a wrong answer is a minor annoyance you learn to shrug off. In drug discovery, a wrong answer that looks right is one of the most expensive mistakes a company can make. It is a compound that advances because a model liked it, a year of medicinal chemistry spent on a dead end, a trial that reads out negative, and, at the end of the chain, a patient who never received the medicine they were waiting for because the industry spent its finite resources chasing a confident illusion.
The field has largely responded to this by wrapping a chat window around a public database, pointing it at some open chemistry, and calling the result discovery. That is an interface, not a discipline. The scarce thing in this market was never generation; models that generate candidate molecules and sequences are becoming abundant and cheap. The scarce thing is a system you can actually trust with a decision that costs real money and real time, a system that knows the difference between a hypothesis and a fact and never lets the first masquerade as the second. We built that system first, deliberately, when it was the harder and less glamorous thing to build, because it is the only part of this that does not commoditize.
Capability is commoditizing. Governance is not. In high-stakes science, being trustworthy and being right are the same thing, and trust is the only durable moat.
2. What governed, gated probabilism actually means
Our architecture rests on one idea that most of the field has backwards. Probabilistic intelligence and deterministic guarantees are not opposites you must choose between. They are two halves of one machine, and the engineering is in deciding precisely where each belongs.
We are probabilistic where it creates value: in hypothesis, in molecular and sequence design, in search across an enormous space of possibilities. There we want the system to be bold, imaginative, and fast, and we accept that most of what it proposes will be wrong, because that is the nature of discovery. We are deterministic where it protects value: in governance, in provenance, in tenant isolation, and in the gates that decide whether a bold proposal is allowed to count as evidence. A proposal is free to be as speculative as the model can make it. It is not free to be promoted, exported, or believed until it has survived a fixed, auditable set of checks that the model itself cannot talk its way past.
This is the difference between a system that produces answers and a system that produces evidence. An answer you cannot trace is an opinion with good production values. An answer that arrives with its provenance, its calibration, its falsification condition, and its honest probability of being right is something you can build a program on. Everything below is how we make that second thing the default rather than the exception.
3. The engine: one governed discovery loop
At the center of the platform is a single governed loop, and the fact that it is a single loop is itself a strategic choice. The loop proposes a candidate, scores it against a rigorous oracle, kills it or keeps it against an explicit numeric condition, and then optimizes what survives, carrying provenance and an unvalidated-hypothesis label through every stage from start to finish.
What makes the loop powerful is that only two parts of it ever change between problems: the representation of the candidate, and the oracle that scores it. Everything else, the governance, the falsification machinery, the provenance ledger, the honest labeling, is shared and reused unchanged. For small-molecule therapeutics, the candidate is a molecule and the oracle is physics-based docking that returns a binding affinity, killed or kept against a numeric threshold. For personalized cancer vaccines, the candidate is a peptide epitope derived from a tumor variant and the oracle is an offline immune-binding predictor that returns a presentation likelihood, killed or kept against its own numeric floor. The candidate and the score are different sciences. The governed machinery between them is identical.
We did not build a tool for one disease. We built one governed loop and proved it generalizes across modalities, from small molecules to neoantigen vaccines to drug repurposing. Each new modality inherits the trust spine instead of rebuilding it.
This is why the platform is not a point solution. The same disciplined engine that discovers a small-molecule binder today extends to immunogen design and to rescuing shelved compounds for new indications tomorrow, without relaxing a single governance guarantee, because the guarantees do not live in the science-specific parts. They live in the loop.
4. Four mechanisms that keep the loop from being confidently wrong
Honesty under uncertainty is not a value statement on a wall. In our platform it is four concrete mechanisms, each of which is a deliberate product decision, and together they are the reason a result from this system means something.
4.1 Calibration-gated trust and honest abstention
An oracle in our loop is not trusted simply because it produced a number. A predictor earns the right to have its output counted, for a specific context, only after it has demonstrably discriminated real signal from noise on a known control set for that context. Until a predictor is calibrated for the exact situation in front of it, its score does not promote anything. The candidate is not quietly killed and it is not quietly kept; it is held in an explicit undetermined state, visible and auditable, waiting for either calibration or evidence. And when a predictor cannot score at all, because a model or a reference is unavailable, the organ abstains and says so rather than fabricating a confident binding it cannot justify. In a field where a false positive costs a year, a system that is willing to say it does not yet know is not weak. That willingness is the entire value.
4.2 Typed, numeric kill-conditions: falsification as a contract
Every claim the platform makes ships with the specific, numeric condition that would prove it wrong. These kill-conditions are not prose a model generates and a human squints at later; they are typed, structured, validated slots, a metric, a comparator, and a magnitude, attached at the moment the claim is born. A predicted binder carries, in machine-checkable form, the threshold beyond which it is considered falsified. This inverts the usual incentive of generative systems. Instead of accumulating reasons a candidate might work, the platform states, up front and precisely, what would make it fail, and then actively looks for that failure. A claim that cannot be falsified is not allowed to pose as a discovery.
4.3 Measured truth overrules predicted truth: the learning loop
A prediction is a starting point, never a conclusion. When a real measurement arrives, from a wet-lab assay on the organization’s own compound or epitope, the platform treats it as ground truth that overrules the model. The measured result flows back into the same governed evaluation and can falsify a finding the model was confident about, and, critically, the correction does not evaporate once it has been applied. It grounds back into a shared knowledge graph and reshapes what the platform proposes and trusts next. The system is built to get better every time the real world corrects it, so that the thousandth prediction is made by a platform that has internalized the outcome of the first nine hundred and ninety-nine. Closing this predicted-to-measured loop on every modality is where we are investing most deeply right now, because it is the mechanism by which a governed system stops merely being careful and starts becoming wise.
4.4 Tamper-evident provenance
Every output the platform commits to is sealed into a per-organization, tamper-evident, cryptographically hash-chained ledger. The seal is content-addressed, so the same claim produced twice is recognized as the same claim, and it is fail-closed, so if the system cannot sign an output it refuses to release it rather than releasing it unsealed. The practical consequence is that you can prove, after the fact, exactly what the platform produced, what inputs produced it, and that nothing was altered in between. This is what turns an interesting result into an asset you can take to a patent examiner, an auditor, or a partner. It is the difference between telling someone your discovery is real and showing them an unbroken chain that proves it.
5. The governance moat is hardest exactly where it matters most
It is easy to talk about governance on public chemistry, where the data is not sensitive and the stakes of a leak are low. We designed the platform so that the governance holds at the opposite extreme, on the most sensitive input it will ever touch: a named patient’s genome, consumed to design a personalized cancer vaccine. If the discipline survives there, it survives everywhere.
In that setting, governance is not overhead layered on top of the science. It is the enabling condition without which the science is not permissible at all. Concretely, the platform enforces this as architecture, not as policy:
No sensitive egress. Patient-derived data and anything designed from it cannot leave the trust boundary unsealed. The egress lock is fail-closed: when in doubt, nothing leaves. Scoring runs inside the boundary on self-hosted predictors, so a patient-derived sequence is never shipped to a third-party prediction service.
Encrypted and recoverable at rest. Reversible sensitive data is encrypted with authenticated encryption and looked up through a blind index, never stored in the clear, and never reduced to a one-way hash when the data legitimately has to be recovered.
Content-free observability. The platform is heavily instrumented so that any run can be validated, but the telemetry records counts, configuration, and metrics only. It never logs the genome, the variant, the peptide sequence, or the generated content. You can prove what happened without the record itself becoming a liability.
Strict tenant isolation. Every read and write is scoped to the acting organization at the data layer, so one tenant’s work is structurally invisible to another, enforced identically whether the platform runs on a laptop database or a production cluster.
Governance is not the tax we pay to do the interesting work. On the highest-stakes problems, governance is what makes the interesting work defensible in the first place.
This is the compliance-by-design argument in its strongest form. For a regulated buyer, the governance trail is not paperwork that slows the science down. It is the evidence that the science can be trusted, generated automatically as a byproduct of doing the work correctly.
6. A platform that compounds
Most AI tools forget your expertise the moment the window closes. Ours is built to absorb it. When a domain expert reviews a result and corrects it, that correction is grounded into a shared knowledge graph and becomes part of what the platform knows, so it reshapes future proposals and future scoring rather than being lost. A claim that a reviewer refutes is struck at the source, which makes it structurally impossible to quietly resurface later as if it had never been challenged.
The effect accumulates. A governed platform that learns from every correction its best people make, and that can never silently un-learn a refutation, becomes more valuable with use in a way a static model never does. The organization’s hardest-won judgment stops living only in the heads of a few senior scientists and starts living, durably and auditably, in the system itself.
7. We are the discoverer, not the toolmaker
This is the distinction most easily missed and the one that matters most commercially. We do not sell a smarter search box and wish our customers luck with it. We run the governed discovery loop end to end, from a biological target, through designed and scored candidates, to a wet-lab-ready dossier that carries an honest probability of success rather than a marketing number. The molecule, and the intellectual property around it, is the asset we create.
That reframes the entire business. In the prevailing model, the AI vendor rents shovels to the people digging for gold and is indifferent to whether they strike anything. In ours, we are running the mine, and we have engineered it so that every claim about what came out of the ground can be proven. A discovered molecule is a candidate we can move to protect, with the full provenance chain standing behind the discovery, and the honest labeling that scopes exactly what has and has not been established about it. Value in this company is not a subscription to a feature. It is the compounding portfolio of defensible discoveries the governed loop produces.
8. What we deliberately do not claim
Part of being trustworthy is being disciplined about the limits of what has been shown, and this is itself a differentiator. A predicted binder is an unvalidated hypothesis, never a validated drug and never a clinical claim. The leap from a model that likes a molecule or an epitope to a safe, effective therapy in a real patient is exactly the kind of overreaching, cross-disciplinary claim our verification is hardest on. The platform labels that bridge, scopes it, and attaches the condition that would falsify it. It does not suppress the hypothesis, and it does not inflate it into a promise. In a field littered with confident overstatement, a system engineered to state precisely what it has not yet earned the right to claim is doing something rare, and something a serious partner will recognize immediately.
9. Why now, and why us
The easy half of this revolution is ending and the hard half is beginning. Raw generative capability is becoming something any team can rent by the token, and when a capability is abundant it stops being a differentiator. What will not commoditize is the ability to make probabilistic intelligence accountable: calibrated, falsifiable, sealed, isolated, corrected by reality, and willing to say no. That is a different discipline from building an impressive demo, and it is precisely the discipline that the highest-stakes markets, medicine first among them, are about to insist on as the price of entry.
We made the expensive architectural choices early, while much of the field was optimizing for applause. Those choices, the governed loop, calibration-gated trust, typed falsification, tamper-evident provenance, fail-closed data governance, and a platform that learns from being corrected, are not a slide in a deck. They are wired into the foundation, and they are very difficult to retrofit onto a system that was not designed to be governed from its first line. A competitor can add a chat window in a week. Rebuilding on a trust spine is a different kind of project, and it is the one we already finished.
Bet on the team that decided, before it was fashionable and when it was harder, that in this field being trustworthy is not a feature of the product. It is the product.
That is what NomosLogic is, and that is what Heuresis does. We turn probabilistic intelligence into governable infrastructure, and we point it at the problem where trust matters more than anywhere else on earth: discovering medicines that work. The next medicine worth believing in should not be the one with the most confident demo behind it. It should be the one we can prove. That is the one we are building the machine to find.

