Completed real models ranked by strict overall accuracy. Mock and oracle runs are excluded. A model appears only after evaluate finishes every question and the run is scored.
Strict 0/1 accuracy per task for the models checked above. Enumeration-style tasks (applicability, action reachability) collapse first.
Selected models overlaid across the five deal archetypes. Pick a task (or overall) to compare familiarity by world type.
Questions emitted per domain, stacked by task. Quotas are uniform; shortfalls mean the emitter skipped rather than guessed.
Mean and max applicable actions across plan states — the size of the enumeration burden the applicability task imposes.
Mean reachable states (log scale) and mean optimal plan length per domain instance.
Strict set-match is soundness-first; Jaccard tolerates near-misses. A wide gap means answers are close but unsound.