This is a Synthena Medical update, the sister effort to the artificial-life notebook. The full state-of-the-engine write-up lives on its own page: Synthena Medical.
Most AI for drug discovery sells you the green light. We build the machine that refuses to fake one, and this update is about what that refusal bought us in a single stretch of work: three new validated predictions the body model can make that no single-organ model can, one dangerous bug caught before it could ever lie, and a hard, honest no to the one question everyone wants a yes to. Every number below comes from a leakage-controlled run, and every claim is capped in code at exactly the evidence it earns. The system cannot say cure, kill, or efficacy: the words raise an error.
The bug that could have made us liars
Our whole pitch rests on a trust gate, so we attacked our own gate and found a real crack: a throwaway phrase, “with no adverse effects observed, this candidate cures the cancer,” could poison it into waving the cure claim through. We fixed it, re-attacked it with fresh poison, and confirmed it now rejects the poison while still passing a genuine disclaimer. 195 tests green, the five protected honesty modules never touched. The most valuable thing we did all week was find the single way our own system could have over-claimed, before it did.
The body started predicting across organs, three times
Each prediction was hashed before we looked at the clinical number, and each was validated against published, adjusted clinical data:
- diabetes to arterial stiffness: about 0.26 m/s per mmol/L, matched on two independent cohorts.
- diabetes to stroke via pulse pressure: inside the blood-pressure-adjusted clinical interval, p approximately 0.00002.
- diabetes to tumour proliferation: p = 0.0085, direction agreeing with three meta-analyses, RR 1.28 to 1.38.
The body doing the one thing it exists for, computing a true cross-organ fact, and it tripled.
The test designed to catch us fooling ourselves
Trained on only pre-2018 chemistry, would the engine recover the novel molecules discovered after? It recovers known-like molecules near-perfectly (0.987) but ranks genuinely novel scaffolds below chance (0.415). We measured our own discovery ceiling and published the number. Similarity recovers the known; it cannot invent past it.
The no that taught us a rule
A fourth claim, body-fat to blood-pressure, self-falsified on its own pre-registered test, because the coupling it leaned on is contested in the literature. From that null we learned the rule that now guides what we build: a cross-organ link is only trustworthy on an independent, measured, uncontested anchor.
Where it honestly stands
As an instrument, stronger: about 9 out of 10. As a source of real-world cures: about 2.5 out of 10, and unmoved. We just ran ten diseases and packaged them for a lab: seven actionable, each with ranked candidates, a decisive assay, and a stop rule, and three honest blanks, Alzheimer’s, ALS, and pancreatic cancer, with zero candidates because that is the truth. Not one claims a cure.
Software cannot bootstrap ground truth from a model of itself. That last mile is a wet lab, by design. One confirmed prediction from a partner lab is worth more than everything above. We did not build a cure. We built the most honest instrument we know how to build for finding one. Simulated is not measured is not cure, and the day we stop saying it is the day we have become what we set out to replace.
The full write-up is on the Synthena Medical page.