NomosLogic Inc

09/04/2026

Announcing From Deterministic Convergence to Discovery: Designing the Next Generation of Therapeutics
A new book by Matthew L. Hardy, Founder and CEO, NomosLogic Inc.

First edition. Twelve chapters, three parts, four appendices.

Drug discovery does not fail because it lacks capability. It fails because it has no shared standard for what evidence licenses a resource commitment, and in the absence of that standard the evidence assembled to justify a commitment is systematically weaker than the decision requires.

That is the argument of From Deterministic Convergence to Discovery, and the book makes it without naming a single piece of software. No system is described anywhere in it, and nothing in it requires any. The rules are written to be adopted by any laboratory, using any tooling, by a reader who disagrees with most of the conclusions.
The Diagnosis
Roughly one in seven programs that reach first-in-human testing produces an approved medicine. The most careful recent estimate, drawn from more than 185,000 development paths between 2000 and 2015, puts approval likelihood from Phase I at 13.8 percent. The largest single loss occurs in Phase II, where efficacy rather than safety is the endpoint, and where insufficient efficacy dominates the recorded causes of termination.

The more important observation is the stability of the curve. Approvals per inflation-adjusted research dollar have halved approximately every nine years since 1950, and the series passes without inflection through high-throughput screening, combinatorial chemistry, the human genome, systematic RNA interference screening, and the solution of protein structure prediction. Each of those was a real capability gain, several of them by orders of magnitude. None produced a discontinuity.

The book draws the inference directly. If capability rises by orders of magnitude and outcomes do not respond, capability was not the binding constraint. Worse, each of those gains occurred downstream of the decision that determines the outcome, so their effect was to relocate failure later, at higher confidence, having spent more. The decline of pharmacokinetic attrition into efficacy attrition is the clearest case: a cheap early failure was converted into an expensive late one, while the metric the work was designed to improve improved.
The Regime
Part I establishes five disciplines, and none of them depends on any particular method being correct.

Decision grade against publication grade. The literature is a proposal mechanism that admits claims which will not survive. A literature claim imported at its published confidence is overconfident by a predictable amount. A decision-grade claim requires provenance, a pre-recorded falsification condition, a calibrated confidence, and narrow reproducibility, as gates rather than trade-offs.
Typed kill conditions. A condition works by spending the interpreter's discretion over a specific future result in advance. A condition written after the data are seen has no function despite being textually identical to one written before, which is why the externally verifiable timestamp is its entire evidentiary content.
Calibrated or abstain. A system required to answer every question will answer the ones it cannot, and those errors will be formally indistinguishable from its valid outputs. The classification must admit a third outcome: not decidable at the confidence this decision requires, paired with a named resolving experiment.
Never launder. Qualifications are lost through chains of individually faithful compressions, so that a claim arrives at a decision more confident than any claim beneath it. This is more damaging than fabrication because no single step is defective and the intermediate states that would prove what was lost are destroyed as housekeeping.
Audit by outcome, not by document. The artifacts of a governance regime are cheap to produce and the discipline is not. Ask for the kill rate, the list of programs stopped, and the instances in which a provenance condition fired on a favorable result.
Two Acts the Book Refuses to Conflate
Part II separates constraint inference, which determines what is true about a biological system, from governed search, which determines what to build in response. Inference is constrained by data and, given a specified model, produces a determinate answer. Search is constrained by an objective and produces a distribution of candidates from which one must be chosen.

Conflating them makes both failure modes illegible. A program can characterize a system correctly and fail on the molecule. A program can generate an excellent molecule and fail because the target was substitutable. From the outside both look identical, and governance that cannot distinguish them cannot learn from either.

The quantitative treatment publishes its thresholds in full: concentration of fitness contribution, concordance under removal, fitness retention under partial perturbation, and reconfiguration pathway identification, with classification boundaries stated and identified as calibration parameters fitted on a development cohort. A method with withheld thresholds cannot be evaluated or refuted. The data and the fitted parameters remain the protected asset.

Across twenty-six studies in four domains and 1,515 samples, every domain median fell in the distributed range. A study framed on BRCA1 returned BRCA2 inside the same load-bearing constraint family, showing that gene-unit analysis is blind to constraint families. A hepatic pair individually redundant proved jointly load bearing, which no component-independent analysis can represent at any statistical power.
The Chapter Written Against Itself
Chapter 7 reports that validation under the rules of Part I, and the report is weaker than the original analysis concluded.

A tie correction to complete separation is withdrawn in favor of raw values of 0.980 and 0.997. A discordant run at 0.710 was excluded by an outcome-triggered rescue rather than a blind validity screen, so the study mean lies somewhere between 0.896 and 0.989. The null used was an unstructured random matrix, which can support only the claim that the model responds to interaction structure, not the claim that it identifies the correct one. Three stronger nulls are specified, requiring no new data, along with a blind re-run. The strong form of the book's central empirical claim is classified not decidable pending that work.

Five specific pre-declared conditions that would refute the approach are stated. None has been evaluated at the time of writing, which is the accurate state of the work.

The reason for including this is stated plainly: a book arguing that discovery fails for want of falsification discipline, which then presents its own results without applying it, refutes itself in the presentation. The three defects that appeared are among the most common in computational biology reporting, and they appeared together, in a chapter arguing for rigor, written by someone who knew all three principles. That is the argument for structural governance rather than individual care.
The Proposal
Part III is written to be used. It closes with a two-page record, lodged before analysis and published with the result, stating the claim and its type, the decision it licenses and that decision's reversibility, kill conditions with the response each firing triggers, the number of analyses planned, blind-applied validity criteria, the null construction with an argument for its severity, the required confidence and abstention boundary, the orthogonal verification, and provenance by content hash.

Voluntary adoption fails, and the book says so. The costs are local, immediate, and visible; the benefits are delayed and partly external. That is a collective action structure no better argument can change, which is why the pressure point sits with funders, regulators, and acquirers rather than with laboratories. Clinical trial registration is the exact precedent, in a neighboring field, within living memory.

Appendix A converts Part III into checklists. Appendix B holds the mathematical reference. Appendix C is a form. Appendix D is the glossary. Every chapter closes with a summary that states its claims compactly, and reading the twelve summaries in sequence takes twenty minutes.
Who It Is For
The intended reader signs off on program advancement decisions: a research director, a translational lead, a chief scientific officer, a reviewer, a regulator, or an investor who has grown tired of asking what the model predicted and being told that the model is proprietary.

A secondary reader is the computational scientist who has built something that works and cannot get it believed. The instinct in that position is to publish more results. The more effective move is usually to publish the conditions under which you would have withdrawn the result, and to have published them first.

A method that cannot survive being reported honestly was not worth defending, and one that can does not need the other kind of presentation.

Matthew L. Hardy is the Founder and CEO of NomosLogic Inc. and the author of Deterministic Convergence and The Adaptation Paradox.

Copyright 2026 Matthew L. Hardy. Availability:

FROM DETERMINISTIC CONVERGENCE TO DISCOVERY: Designing the Next Generation of Therapeutics

09/04/2026

Drug discovery has multiplied its computational capability by orders of magnitude and has not improved its success rate. That sentence is not a complaint about the difficulty of biology. It is a measurement, and it carries a conclusion most of the field has not accepted.

08/28/2026

AI in drug discovery can generate countless novel molecules. It can also generate countless molecules that are physically impossible, violating fundamental laws of stereochemistry, geometry, and quantum mechanics. This is a profound waste of resources.

Our PROTEUS engine acts as a deterministic filter against this class of error. It runs evolutionary simulations grounded in biophysics to ensure that proposed structures are not just computationally plausible, but physically viable.

Before a single dollar is spent on synthesis, the hypothesis must pass the test of reality. Imagination must be constrained by the laws of physics.

08/22/2026

Nine out of ten candidates entering Phase I never reach approval, and roughly half of those failures happen in Phase II, where efficacy is the endpoint. The molecules were safe. They bound their target. They did not change the disease.

The field calls this translational failure. That is a description, not a diagnosis.
The mechanism is architectural. Biology is selected against fragility, so critical functions sit on partially redundant components with reserve capacity. Remove one node and the phenotype reroutes. The target was real. The system that uses the target was never modeled. We selected for visibility when the discipline required necessity.

Fixing target selection is necessary and it is not sufficient, because of something the field has not fully absorbed:
An architecturally validated target does not specify a molecule.
A constraint defines a surface, a geometry, a set of contacts. It does not define atomic composition, or even modality. Drug-like chemical space is estimated above ten to the sixtieth. Applying the constraint partitions that space, it does not enumerate it. What survives is a set of structures that all engage the right interface and differ enormously in selectivity, exposure, synthesizability, and safety, and none of those properties are encoded in the constraint.

So there are two scientific acts here, not one. Deterministic inference establishes what is true. Stochastic search establishes what should be built. Different failure modes, different validation standards, different relationships to uncertainty.

Collapse them and both failures become invisible. You can get the architecture exactly right and fail at candidate generation. You can generate a beautiful molecule against a redundant node and fail in the clinic with a clean chemistry package.

A stochastic generator with no architectural constraint is a treatment with no diagnosis. A deterministic architecture with no search downstream is a diagnosis with no treatment.
That is why we run two engines with an explicit handoff rather than one system with a broad mandate.



https://www.nomoslogic.com/blog/two-engines-not-one-a-perspective-on-drug-discovery

08/16/2026

The human mind is brilliant at abstraction, but it struggles to hold more than a handful of variables at once. A single human cell operates on a system of millions of variables. The scale of the biological problem has long since outgrown our native cognitive toolset. This is the core challenge of modern drug discovery.

To reason effectively about a system this complex requires a new kind of cognitive infrastructure. It requires systems that can manage and compute across vast networks of rules and evidence. Our own knowledge base contains over 1.2 million clinical rules applied to more than 16 million clinical assets. This is not about "big data" for the sake of it. It is about building a logical framework that allows a researcher to ask a precise question and get a deterministic answer from a sea of complexity.

We are not building a bigger boat. We are building a better compass.

08/16/2026

Can we model neurodegenerative disease as a materials science problem? Pathological protein aggregation has striking parallels to processes like uncontrolled crystallization or material fatigue in engineered systems. Viewing these diseases through a materials science lens, instead of a purely biochemical one, forces us to ask different questions.

What are the initiation conditions for this phase change? What are the propagation dynamics? Are there ways to introduce impurities or alter the energetic landscape to disrupt this process? This cross-domain approach does not discard biology. It augments it, creating a new set of testable hypotheses. The goal is to find novel ways to frame the problem that lead to novel avenues of investigation.

A new question is often more powerful than a new answer.

08/14/2026

It’s time we start thinking about disease and biology from a multi-specialty perspective. For too long, scientists and researchers have operated in tunnel-visioned silos. Biology and disease are both architectural. The laws of physics apply. The laws of thermodynamics apply. Chemistry, mathematics, and systems architecture all apply. So why have these disciplines rarely worked together, until now.

That is exactly what NomosLogic is building: infrastructure to bring these fields together and move medicine into its next evolutionary phase.

I don’t believe curing and preventing disease is actually a hard problem. I believe we’ve made it hard by refusing to look past our own specialty, chasing prestige inside a black box instead of the cure outside it. No two patients share the same biology, so why do we keep treating them like they do. No disease was ever purely biological to begin with, so why do we keep looking at it like it is.

08/12/2026

The most profound questions in medicine are often hidden at the intersection of disciplines that rarely speak to each other. A geneticist, a radiologist, and a pharmacologist may all hold a piece of the same puzzle, but their data exists in mutually unintelligible formats.

Unlocking the next wave of discovery requires a translational layer, a Rosetta Stone for biology. We built a system with over 496,000 nomenclature mappings across more than 40 systems to serve this exact purpose. It creates a coherent logical space where multi-omic data, from DNA to blood to medical imaging, can be fused and interrogated.

The tool doesn't provide the answer. It creates the conditions where a new, more powerful type of question can finally be asked. Judgment remains the essential human element.

Innovation happens at the boundaries.

08/12/2026

Does borrowing a concept from physics and applying it to biology loosen the standards of rigor? Quite the opposite. It tightens them. A hypothesis must now survive scrutiny from two distinct intellectual traditions. It has to make sense chemically *and* physically. It must be plausible evolutionarily *and* computationally.

This is the central discipline of cross-domain research. You are not importing a loose analogy, you are importing a set of hard constraints. A model for protein aggregation inspired by polymer physics cannot violate the laws of thermodynamics. An information theory approach to genetic regulation must still account for the known biochemical binding affinities. The burden of proof multiplies.

Before we run a single wet lab experiment, our computational models must face a gauntlet of pre-registered kill conditions. If a hypothesis cannot survive this theoretical and logical cross-examination, it does not earn the right to consume precious lab resources. Rigor is not about adherence to one method, it is about surviving all of them.

Discipline is the shortest path from hypothesis to truth.

08/11/2026

Patience and urgency are not opposing forces in deep technology; they are partners. The timeline for understanding complex biology is long, measured in years or decades. But the timeline for testing a single, well-formed hypothesis should be brutally short.

The industry standard for resolving a complex genomic variant can be two to four weeks. This cadence dictates the pace of research. Condensing that same resolution to under 130 seconds does not rush the scientific process. It radically increases the number of learning cycles possible within the same window of time.

This is patient urgency. We remain patient for the ultimate answer while being relentlessly urgent about completing each iterative step. The goal is to accelerate the rate of inquiry, not the speed of conclusion.

The right tools turn waiting time into working time.