Is the number your decision rests on actually real?

Teams at the moment a computational result turns into a spend:

• A founder about to commit wet-lab or CRO budget to a hit list.
• An investor at term sheet, looking at validation statistics they cannot independently check.
• A team whose in-house model reports numbers that look, honestly, a little too good.

If your result has already survived a blinded holdout and a negative-control panel, you probably do not need us. Say so on the call and we will tell you the same.

We run the target and the compound set through seven adversarial gates. Each one is a way results like yours have failed before:

• Receptor provenance — is the structure actually the protein you think it is? Mislabelled PDB entries are more common than anyone admits.
• Calibration — do predicted values track measured ones, with a confidence interval and an honest n?
• Discrimination — does the model beat a plain chemical-fingerprint baseline? Many do not, which means the expensive model is buying nothing.
• Negative controls — what is the false-positive rate on compounds known not to bind?
• Enrichment — does the ranking concentrate real actives at the top, or just look tidy?
• Blinded holdout — compounds the pipeline has never seen, scored against a threshold fixed in advance.
• Replication — does the result survive being run again?

Alongside that, your compound set is triaged for PAINS and BRENK liabilities and checked against a lipophilicity null model — the single most common way a “novel chemotype” turns out to be an assay artifact.

• A target card: every gate, its result, its evidence, and its confidence interval.
• A triage table for your compounds, with liability flags and calibrated binder probabilities.
• A go / no-go read in plain language, with the reasoning shown.
• An explicit list of what the data does not support — usually the most valuable page.

Written so you can forward it to an investor, a board, or a skeptical colleague without translating it first.

Because we have made these mistakes and published them. We retracted our own CYP result when it fell from 0.94 to roughly 0.47 on novel chemistry. We retracted a knockout result that passed at n=89 after finding the panel contaminated. Two of our seven targets clear our own gates, and we say so on the front page.

That is the skill you are hiring: not optimism about computational methods, but a practised eye for how they quietly fail. The full ledger is public.