On six development anchors, August 2025 through January 2026, six-month AUROC was 0.819 for enrollment shortfall, 0.832 for termination and 0.806 for completion. These rank later positive outcomes above negatives; they are not per-trial probabilities.

The twelve-month comparison also improved 12 of 14 targets. Withdrawal and results submission declined at both horizons. One seed; development anchors. The served arm includes fresher facts, and a new facts-only rival was not recomputed, so this does not isolate encoder value.

The serving tools provide outcome probabilities, attention lists and expected counts. Calibration maps are fit on an earlier eligible block and checked on a later reporting block.

The current six-month reporting read has expected calibration error 0.002 for shortfall and 0.001 for termination; positive Brier skill is 0.019 and 0.014 respectively. Small rare-outcome calibration error does not imply strong per-trial certainty.

Registry records do not supply a complete live enrollment feed. Forecasts concern what the registry later records, subject to its publication cadence and source quality.

The current scenario path composes rolled trial state with stepped facts and clock, while holding sponsor context at the history cutoff. It serves the twelve-month production artifact.

A supplied move is compared with the same future date with nothing filed. The difference is a conditional model response, not a causal estimate. Real-event-chain evaluation tests outcome ranking; imagined trajectories remain descriptive.

Open Trial Dynamics