Three tasks that must remain separate
Search identifies studies by disease, stage, molecular target and location. Preliminary matching explains which public criteria may fit, conflict or need clarification. Screening formally evaluates source documents and is performed by the research team.
An agent can assist with the first two tasks. A match to a public protocol never guarantees an open slot, eligibility or benefit for an individual.
- search produces a list of potentially relevant studies;
- matching shows the rationale and missing data;
- the research team confirms or rejects eligibility;
- enrolment requires the applicable procedures and informed consent.
Use only the data the task needs
Preliminary navigation usually needs a structured minimum: cancer type, stage, key biomarkers, treatment line and dates, and accessible geography. A full medical record should not be collected by default.
The public project uses synthetic scenarios only. Do not send medical records through the proposal form or general email; sensitive-data workflows require separate protected design and legal review.
What an auditable result looks like
A useful output is not a long link list. Each candidate should show the registry identifier, current status, location, quoted criterion, check date and questions for the coordinator. ‘Potential match’, ‘conflict’ and ‘unknown’ are different states.
TrialGPT illustrates a useful pattern: search, explainable matching and ranking generate material for expert review, not an autonomous enrolment decision.
What to measure in a pilot
Define a reference set and stopping criteria before use. Practical measures include recall of candidate studies, false exclusions, criterion-citation accuracy, coordinator time, missing-data rate and handoff quality.