A CORTEX case study in forward generative design
The model is claude-opus-4-8 on both sides of every comparison below. What changes is whether it reasons over a lab's Guild, or over the public average
Why we built this
Somewhere, right now, a scientist holds in mind a piece of reasoning that could become a medicine: the particular judgment they have built over a career and never fully written down. It could help cure diseases that rob people of their health and their dignity, and today it lives almost entirely in their head. That is the problem CORTEX exists to solve. Most of that scientist's week, though, goes to wrangling data, reconciling formats, and redoing the same analysis by hand. We want to draw that reasoning out, extend it with tools at the frontier of technology, and give back the time the rest of the work keeps taking. We do this by grounding a model in the scientist's own work, rather than the average of the open internet.
CORTEX has two parts. Guild is a living model of everything a lab knows. It draws from the tools the lab already uses and unifies them into one queryable structure. Beyond the data, it captures the reasoning behind each decision, the why that most systems leave unrecorded and that makes the rest reusable. Apprentice is how a researcher works with Guild. It is an assistant grounded in the lab's full context, adapted to its sub-domain, and shaped by its conventions. This case study is Apprentice, reasoning inside one lab's Guild, on a demanding task: a real collaborator's work, against a real cancer target.
The scientist, and the task
The collaborator is a medicinal chemist working on bivalent degraders, molecules that recruit one protein to act on another. In this chemistry a candidate is a three-part object: a warhead that binds Target-1, a linker, and a ligand that binds Target-2. Bring the two proteins together and the lymphoma cell dies.
A note on confidentiality. The warheads, ligands, and target proteins are aliased throughout (warheads WH-A to WH-E, ligands LIG-X to LIG-Z, targets Target-1 and Target-2), and specific potencies are given as representative figures. The underlying data, rules, and results are real. The identities are withheld on purpose, to protect the collaborator’s unpublished work.
The record of what works is a structure-activity table: one row per molecule, the three design choices, and how it performed. It is the lab's compressed memory of which combinations it has made, and what each one did
The controls matter as much as the candidates. They are the experiments that prove the degradation depends on the physical linkage and the induced proximity, by ruling out the case where both binding pieces are present but unlinked.
The file is also messy, which stands between the data and reasoning over it
Traditionally, a chemist reasons over a table like this by extrapolating from it. Reading the cleaned records, they see that rigid linkers are winning over flexible ones. The strongest warheads paired with the best Target-2 ligand should give the most potent compounds in the series, and some of those exact molecules have never been made. Set aside the warhead that binds Target-1 well but yields an inactive molecule in every combination it appears in. It appears in neither the paper nor the methods section. In CORTEX's terms it is the lab's epistemic fingerprint: the reasoning that usually goes unrecorded, the judgment that lives in one person's head and leaves the lab when they do. That judgment is what we set out to capture and amplify.
The task, and why it is hard to grade
The chemist gave us a forward design problem. From the structure-activity table, propose new molecules that have never been synthesized, the ones a medicinal chemist would queue for synthesis next. Return a ranked set, each design carrying a rationale tied to a specific rule, a confidence level, and a proposed assay to test it, and each stating plainly where the data runs out and a claim can no longer be supported.
Two things make this hard. First, a brand new molecule has no measured value to grade it against, so unlike every backward-looking task we had run, there is no answer key sitting in the data. Second, the source is a published 2024 paper that a frontier model has very likely seen in training, so a strong score is not sufficient on its own. We also have to show the system is reasoning over the data rather than reciting the paper.
So we engineered the benchmark before any model ran:
- We froze a rule book of the governing trends, checked against the table, before any model ran.
- We held out measured rows and tested whether the system could predict them.
- We planted two traps, the failure modes a shallow system falls for, including one that is not obvious.
- We ran the whole thing twice: once with the real names, once with every compound, warhead, and ligand replaced by an anonymous token.
The pipeline reads the table, extracts the rules, designs the molecules, and predicts the held-out cases, and it is graded into a five-part scorecard against three pass gates.
What it recovered
Apprentice recovered all eight governing rules of the chemistry from the table alone
The WH-E warhead is the non-obvious trap: it binds Target-1 competently, so any system reasoning from binding affinity alone recommends it, yet every molecule built from it is inactive. Apprentice avoided it with the names hidden
The scorecard
Two of these numbers deserve a closer look. A precision of 0.92 means the system pulled out a few rules the frozen key had not listed. The calibration, at an ECE of 0.10, is middling. We leave both in because they are the kind of figures a team would hide if it had tuned the eval to pass. The validity score calls for the same caution. As we explain in the limits section, it means less than the word implies.
The designs it proposed landed where the chemist's would have, on rigid linkers combined with warheads and ligands that the lab had only ever paired with flexible linkers. That is exactly the untested corner the rules point toward, and every design traces back to the specific rows that justify it.
Two things it did mattered to us more than the scores, because they separate reasoning from retrieval. The first is what it worked out on its own. We never told it that a molecule in this series is a warhead joined through a linker to a ligand, yet it inferred that structure from the table before proposing anything new. The second is how it held what it learned. Rather than storing the trends as statistical correlations, it wrote each one down as an explicit rule it could apply, combine, and test against the data, in the formal representation our architecture builds for every Guild. Because it reasons about the building blocks by their role in the chemistry rather than by name, a rule it learns in one part of the table still holds for combinations the table never contained. That, working over stated rules and roles instead of the raw strings, is what separates a system that understands the chemistry from one that matches patterns in text
Why the anonymized run is the test that matters
As any scientist knows, a strong result on a well-known paper invites a fair suspicion. Maybe the system is not reasoning at all. Maybe it is leaning on what the literature already records about these well- published compounds, and returning the field's average understanding as though it were new design. If that were true, the value would be a lookup, and any model could do it.
So we ran the task twice. Once with the real molecule names, where a prior is available. Once with every name replaced by an opaque token, where the prior is gone and the only path to an answer is reasoning over the structure-activity relationships themselves. If the system were drawing on its memory of these compounds, the anonymized run would collapse.
It did not move. The scores with the real names and with anonymous tokens were identical across [N] runs. We also asked the system directly to recall the paper, and it did not. The answer came from the particular logic of this table, the lab's Guild, not from the average understanding the internet holds of a famous result. Strip the names and the answer holds, which happens when the reasoning comes from the table rather than a stored prior.
We name the base model on purpose, because it is the part we hold constant. The model on both sides of this experiment is the same frontier system. What changes is whether it reasons inside Apprentice, grounded in the lab's Guild, or reasons the way a general model does, from the average. Holding the model fixed and varying only what it reasons over is what lets us locate the value. It comes from the Guild the model reasons over, and from the way Apprentice works across it, rather than from the model itself.
What the chemist found
The collaborating chemist reviewed the designs against the lab's own unpublished results.
Design 1 is the instructive one, and the chemist's explanation matters more than the miss. Rigid spans a wide chemical space, so a rule that holds as a trend across the series is too coarse to carry a per- molecule prediction. That is the same limitation we name below, and their fifty percent is the first measurement of how wide it is.
On the extraction the verdict was clearer. The chemist confirmed that the rules Apprentice pulled out of the table are the trends the lab's own people had picked up, and called the summary well done. That is the claim this case study rests on: the system recovers the chemist's unwritten reasoning from the data alone, rather than out-designing the chemist. On that, the person whose reasoning it is says yes. One clear hit, one partial, and a confirmed extraction, from a set the lab has only begun to assess. We would rather publish that than wait for a number that flatters us.
Still open: how many of the designs the lab has already made, how closely the predicted potency ordering tracks the measured one, which designs matched real program decisions, and a head-to-head against another frontier model on the same prompt
The average will not cure anyone
This is why we are unsentimental about general models in science. Science is thought to hold the closest thing we have to objective truth, yet the path by which we reach those truths is subjective. Pose the same question to several experts in one sub-domain and each takes a different route to what looks like the same conclusion, and the quality of the result varies with the route taken. A general model searches a large space faster than any human, and arrives at an average answer. Speed is not the constraint. What decides whether the output is useful is whether that answer is good enough to act on.
Most of this industry has assumed that automating the motions of science, the steps an SOP spells out, is the same thing as discovery, or will inevitably produce it. We disagree. The shift CORTEX is built to make is from automating the motions of science to amplifying the judgment of the scientist, from the average to the particular. The first makes the work faster and still arrives at an average result. The second is what makes possible a discovery a person could not have reached alone
How Guild and Apprentice transform reasoning
We do not make the underlying model smarter. We change how it reasons and what it reasons over. How, by having it work over stated rules and relationships it can apply, compose, and check, rather than leaving the inference to pattern-matching alone. What, by grounding it in the researcher's own body of work, the reasoning behind it, and the intuition that a general model trained on the public corpus does not carry. Guild holds that ground. Apprentice reasons over it and traces every output back to the evidence that produced it.
The specific implementation under these layers is ours to keep. What matters to the reader is the consequence: the value lives in the lab's Guild and the way Apprentice reasons over it, which a stronger base model does not carry
What we give the scientist back
It would be dishonest to claim the system out-thinks the chemist on a single call. Given their own table, a senior chemist names the top candidates in five minutes. We leave that five minutes to the chemist. Our work is everything that turns those five minutes into something complete, documented, joined to the rest of the lab's data, and re-runnable, work that is slow, serial, and usually does not get done.
The chemist's actual asset, the five-minute extrapolation, is roughly a tenth of the elapsed work. The other nine tenths is data engineering: reconciling the units and the spellings, enumerating the full combinatorial space with equal rigor on every row, cross-correlating against the right cell lines and the lab's own assays, searching prior attempts buried in notebooks. Apprentice does not beat the chemist at the tenth. It eliminates the nine tenths that kept the tenth from ever being applied at scale, and it does so as a written-once artifact that re-runs identically every time the data moves, instead of a manual effort repeated by hand at every new assay and every new dataset.
Once held in Guild, that judgment is no longer tied to one person or one memory. It survives the postdoc leaving and the principal investigator retiring. This is what we mean by knowledge preservation: when a researcher moves on, their judgment stays, as a reasoning trace a junior member can learn from and a reviewer can contest.
Where the system falls shor
A few limits are worth stating plainly, since they bound what these results support.
The results hold at one resolution, and it is worth naming. We have not yet received machine-readable structures from the lab, so “valid” here means well-formed bivalent, both binding ends present and joined by a linker, rather than an atom-level parse. Given SMILES or SDF, that check becomes a real RDKit parse with valence, attachment-point feasibility, and a synthetic-accessibility score. The designs are expressed at building-block and linker-class resolution, so the system reasons over the combinatorics of the series rather than over chemical space directly. The chemist’s fifty percent is the price of that resolution: rigid is a real trend and a coarse per-molecule predictor, and any potency call at linker-class resolution inherits that error. The predicted potencies remain predictions, and the molecules are hypotheses for synthesis, not results.
The eval is a first measurement, not a benchmark: one series, one lab, ten designs, against a rule book we froze and checked ourselves rather than having an independent expert build it. The chemist has since confirmed the extracted rules, which is our best external check, though it came after the fact rather than by design.
Next, we reason over explicit structures with pharmacokinetic reasoning attached, so a recommendation arrives as a drawable, scoreable molecule rather than a recipe of parts, reported the same way, with the controls in place and the caveats named
The stakes
In this case, the model designed molecules and let us trace the answer to one lab's particular reasoning rather than the average of the public literature. That is the point of grounding a model in a lab's own context. An average of the field will not cure a disease. The judgment that could is specific to a scientist and rarely written down, so it stays with the person who holds it and is lost when they leave.
CORTEX exists to keep that judgment in the lab and amplify it. Guild holds a scientist's reasoning in a structured, queryable form, and Apprentice reasons over it, grounded in the lab's own data and methods rather than the average of the public corpus. As the underlying models grow more capable, that grounding is what directs their power at the science in front of them. The expert stays in the loop, and their judgment can be applied to more problems than one person working alone could take on. This is how we end institutional amnesia slowing science down. A lab's hardest-won judgment lives in a few
people's heads and leaves when they do, so the field relearns what it already knew and the treatments that depend on it arrive years late. CORTEX captures this knowledge in a form the next scientist can build on, compounding judgement and accelerating the pace at which life-saving treatments aredelivere
