One cancer at a time. Fresh data from ClinicalTrials.gov, PubMed, and Open Targets. A model reads the trials, reasons about the mechanism, builds a living dossier. Runs unattended. Streams in public.
The loop fetches one fresh pack per cancer: active neoantigen trials from ClinicalTrials.gov, recent papers from PubMed, target scores from Open Targets. Every request is logged. Every source is public.
The model receives the evidence pack plus everything it already believes about that cancer — open hypotheses, active leads, what it retired and why. Memory is the dossier. Reasoning compounds across visits.
Each lead gets strengthened, weakened, retired, or opened. Citations are verified against the pack. Anything unsupported gets flagged UNVERIFIED. One attempt. No retries. The trace is what it is.
A lead is cheap to write. Promotion is mechanical. The model does not get to decide.
Corroboration from independent literature on a later visit is the only path up. A claimed level without the evidence behind it gets clamped back down, and the clamp prints to the public terminal.
A coherent therapeutic hypothesis tied to a verified trial or paper. Cheap to create, cheap to kill. The starting point.
Corroborated by new, independent literature on a later visit. The loop did not go looking for it — it arrived.
Corroborated repeatedly. Should be rare. The desk tried to retire it and failed on the evidence.
Contradiction outranks confirmation. A negative trial is recorded so the loop cannot re-propose it. Retiring a lead on good evidence is a success.