Conditional scenarios
Conditional trial forecasts
The composed roll advances the learned trial state and updates trial facts and clock through supplied moves. Sponsor state and aggregates stay at the history cutoff.
- 01
Start from the dated trial and sponsor record.
- 02
Roll a supplied action and a matched path with nothing filed to the same date.
- 03
Read both through the same outcome heads; report the difference as a model sensitivity.
Actions in training were chosen by sponsors. These are conditional forecasts, not measured intervention effects. The composed scenario artifact currently serves the twelve-month production path.
Held-out real chains
Outcome evaluation
The September 27 composed read used 59,769 held-out chains with eight real events per trial and twelve-month labels. At depth eight, composed rolled AUROC was 0.775 for shortfall, 0.743 for termination and 0.834 for completion.
Composition improved every reported target over latent-only rolling. Facts and clock carry much of the signal. Timeline-slip ranking remained below the never-moved comparison: 0.736 versus 0.738.
This test supplies real moves; it does not establish that the model can predict the particular future action sequence. Tracking error in the model’s own representation is a diagnostic, not evidence of outcome forecasting.
What the result supports
Scenario assumptions
Heads fitted on rolled inputs improved some targets and hurt others, so they were not adopted. The current composed heads are used on both sides of each scenario.
Sampled trajectories are descriptive. The old requirement that sampled futures reproduce today’s forecast was withdrawn: that read mixed outcome windows and payloads drawn from other trials. Historical numerical action deltas from the previous model family are not current AACT 2.0 claims.