The console can place a predicted pass probability beside the validated engine result, forecast the next operational bottleneck from real events, and read a volunteer's cross-study history before enrollment. Every figure the model shows is cited to real records. The bioequivalence verdict still comes only from the validated TOST engine.
The predictor sits next to Module 05, never in front of it. It reads the study features and comparable recorded history, returns a probability, and cites the studies it leaned on. Any citation the model cannot match to a real local record is dropped before you see it. The model proposes. A qualified person decides.
A gut-feel go or no-go, blind site selection, and over-participation that slips through between studies.
A cited probability beside the validated verdict, a forecast from the live event stream, and a volunteer's real history read before the needle.
The flag reads a volunteer's real apparent clearance across prior studies and raises a review when it trends toward enzyme induction, cited to the exact studies and values. It never blocks enrollment on its own. A reviewer accepts or dismisses it, and that decision is recorded with who and when. The model may read the trend in prose. The slope and the values stay engine-computed.
A slow shift in a repeat volunteer's clearance that no single study is placed to notice.
A cited, deterministic trend surfaced before the needle, for a qualified person to act on.
The predictor is back-tested against every archived study that has a recorded outcome, with that outcome provably held out of each prediction. Accuracy, sensitivity, specificity, ROC AUC, and calibration come from validated code, with a deterministic held-out split and an exportable evidence pack in JSON and CSV. Deploy-stage under the hosted app, and honestly labelled offline.
A model that reports its own accuracy, with no held-out test and nothing to hand a regulator.
Leave-one-out back-testing with provable exclusion, and an evidence pack QA can read.
Run one of your own studies through the console and watch the prediction sit beside the validated engine, cited and human-gated.