The note said "no evidence of diabetes." A similarity matcher read the word diabetes, scored it 1.0, and excluded the patient. She was eligible. The trial lost her anyway, and every patient dropped this way is one a coordinator has to go find again. The failure is mechanical: a vector similarity score has no way to represent the word "not." It sees "diabetes" in the note, sees "diabetes" in the exclusion criterion, and matches them, negation and all. The same blind spot pins a family member's cancer on the patient and cannot read "unless completed more than 12 months before randomization." Those aren't language problems a bigger model fixes. They're logic problems. So we built TrialProof to compute eligibility instead of predicting it. The model does the one thing it is good at, reading the messy note and proposing facts with the span it read them from. An adversarial verifier then strips the negated mention, and a deterministic engine over a medical ontology returns ELIGIBLE, EXCLUDED, or NEEDS-REVIEW. The verdict is code you can audit, not a number you have to believe. 🧬 On a fixed 13-case labeled gold set, TrialProof scored 100% against the fair lexical baseline's 53.8%, and lost 0 eligible patients where the baseline lost 3. Every decision it emits carries a reproducible reasoning trace, the baseline carries none, and re-running the whole set returns byte-identical output. All patients and protocols here are synthetic. If your team is matching patients to protocols and wrestling with how to trust what an AI pulled from a clinical note, we would genuinely like to hear how you are thinking about it. The screen failures that cost the most are the eligible patients you never knew you dropped. #ClinicalTrials #ClinicalOps #PatientRecruitment #HealthAI #NeuroSymbolicAI