Passive, timestamped evidence accumulates until an evidence cutoff; everything up to the cutoff is what a model may see. The constructor then issues one operational question about a future state of the target that is still unresolved when predictions are sealed. A deterministic, pre-registered rule later maps future evidence to exactly one answer — never self-report, never post-hoc judgment.
Two weeks of passive events for the generated persona P-SYN-01, across five evidence streams. This is the shared substrate every instance below is cut from.
At the evidence cutoff, the target’s next state is not yet a fact — it is a possibility space: a small, operational set of mutually exclusive options the future must resolve into. The benchmark scores how well a system’s sealed probability distribution over exactly these options predicts how the instance actually resolves. The ten demo instances span all four personal families: attention shift, focus recovery, priority change, and commitment follow-through.
Greyed events after the dashed line are unresolved at seal time: they do not exist yet when predictions are sealed. They are consumed later by the deterministic resolution rule.