How it works

Four steps, two of them never touch a model

1. 01

   Harvest

   Every trial still open to new patients is pulled from ClinicalTrials.gov, EU CTIS and ISRCTN using their public interfaces. EU CTIS publishes no recruitment flag, so its authorised trials are listed as approved to run with the recruitment question left open, never as recruiting on CTIS's word alone — unless another registry for the same trial does publish one, in which case that registry's answer and its wording are shown. Where two registries disagree, the trial says so rather than picking a side quietly. The same trial registered in two places is merged by its cross-referenced registry numbers, so a UK trial that only exists on ISRCTN and CTIS is not missed.

2. 02

   Split

   A language model splits each trial's eligibility text into single criteria and, where it can, a structured rule (an age range, an ECOG limit, a prior-lines cap, a drug that must or must not have been given). Each criterion carries a confidence and the exact words it came from. A quote that is not a character-for-character substring of the registry text is rejected by code.

3. 03

   Grade

   Ordinary code compares each rule with the details you entered and marks it met, unmet or unknown. Unknown produces the question to ask the trial team. Unmet carries a note on whether that kind of criterion is ever waived. The trial-level grade is a fixed function of that table. No model is involved, so the same inputs always give the same grade, and your details never leave the browser.

4. 04

   Routes in

   A curated, versioned set of notes on how a trial abroad is paid for and reached, attached by rule: an EU resident looking at a trial in another EU country sees the planned-treatment route; a site at the NIH Clinical Center shows its free-care note. Every note names who to verify it with.