blueVigil
Follows one research subject's treatment exposure, toxicity, and therapeutic effect across a full trial course, computing dose-limiting-toxicity occurrence and evaluability as two separately computed answers and printing a fixed action envelope — disposition, evaluation completeness, and constraint conflict — before any proposed next step. A named clinician adjudicates; blueVigil records what was run and by whom.
Clinical trial investigators, study coordinators, and sponsor safety and pharmacovigilance teams following one research subject through a treatment course.
blueVigil does not decide anything about a subject's treatment. It records what was run and what a named human decided, and nothing in this package may originate a clinical definition, grading rule, dose-limiting-toxicity criterion, or threshold — every one of those arrives as versioned, dated rule-pack data. No LLM is anywhere in the engine that computes them.
It never adjudicates attribution. Whether a toxicity relates to the study treatment, the disease, or something else is recorded only when a named person reviews and signs it; an assessment nobody has made yet reads as not-yet-assessed, never as a computed "unrelated."
It never confirms a dose-limiting toxicity or computes an overall therapeutic-effect verdict. The engine can flag a dose-limiting-toxicity candidate and, on a separate and independent axis, whether the observation window was even evaluable — but only a named human's own adjudication can confirm one. The five therapeutic-effect domains are never merged into a single score; a sponsor export counts how much evidence exists per domain, never how much benefit occurred.
A signed export proves what blueVigil ran and recorded — it does not prove an underlying clinical claim is true. A change detected against a prior signed export is surfaced for a named person to reconcile; it is never applied automatically.
blueVigil is worked one research subject at a time: pick them from the roster, follow their course on one timeline, work the toxicity episode and any readings that disagree, check what the dose-limiting-toxicity record actually decided and who decided it, then see what a sponsor safety export would show. The five screens below are captured from a running build, not a mockup — against JY-SYNTH-SUBJ-001, the one synthetic subject in this build with a full course. Every one of them still carries the same illustrative-data notice, because the clinical definitions behind them have not been reviewed end to end.
The subject course for JY-SYNTH-SUBJ-001 — one timeline blending four tracks (exposure, toxicity grade, events, baselines) against the same axis instead of four separate charts. The peak grade-3 toxicity carries its own note: not confirmed, no signed assessment. Underneath the timeline, the Subject safety summary lists every computed fact still waiting on a human decision, starting with a concordance assessment already marked Discordant. (Not shown on a phone: the chart's own labels are too small to read once the timeline is squeezed into a 390px width — see the note below.)
Safety evidence expands into the episode board, where that discordant reading is worked in full.
Episode board — the same subject's toxicity episode with its full grade history, not just the latest grade, each entry traced to a named source fact. Underneath it, the concordance table: on the same day, the biomarker reading improved and the imaging reading worsened against the same baseline. blueVigil keeps both readings in full and computes no combined value — Discordant is the only verdict this screen prints, and a person decides what the disagreement means.
Decisions carries the same subject into the dose-limiting-toxicity record, where occurrence and evaluability are answered separately.
The decision record answers two questions separately rather than collapsing them into one verdict: whether a dose-limiting toxicity occurred (No DLT observed) and whether this subject is evaluable at all (Evaluable). Each of the seven criteria is shown on its own — four pass, three fail — instead of being rolled into a score that would hide which ones failed. The engine records itself as computed, no candidate, and that is the furthest it can go: the cohort panel counts one subject at DLT (confirmed), a state blueVigil's engine cannot reach on its own and that only a named investigator can enter.
Sponsor & study turns the same subject's course into what a sponsor actually receives.
Snapshots — the five-domain therapeutic effect ledger for the same subject. Two domains read Improved against baseline (objective disease, from imaging and biomarker evidence; biologic/molecular, from biomarker evidence alone), one reads Stable, one reads Worsened with no recorded basis for that call yet, and Patient-reported has not been measured at all. blueVigil never merges these into one score — the ledger only counts how many domains have evidence, never how much benefit occurred.
Zooming back out, Subjects → Roster is where a reviewer actually starts: every subject in the study, before choosing whose course to open.
Roster — every subject in the study, the screen a reviewer opens first to choose who to look at. The same DLT bucket counts from Decision record appear again here as a study-wide summary, and the table names what it leaves out in plain text: no patient name, age, sex, MRN, allergies, dose level, or cycle/day sits on this screen at all — only what each subject's dose-limiting-toxicity determination required.
In active build. The engine computes treatment-exposure course, longitudinal toxicity episodes with full grade history, dose-limiting-toxicity occurrence and evaluability as two independently computed answers, explicit multimodal concordance, a five-domain therapeutic-effect ledger, and signed, offline-verifiable safety export snapshots — each behind its own falsifying test, none reachable unless explicitly enabled. The five screens above are captured from that running build against one synthetic research subject; every one of them still carries an illustrative-data notice, because the clinical definitions behind them have not been reviewed end to end. Dr. Jean-Yves Mazière reviewed the design these screens were built to.
Not yet true: Not included in the free edition. No signed pilot. No live or real patient data has ever been processed.