Information Location under Matched Predictive Opportunity

Model rankings are difficult to interpret when benchmark tasks differ both in how predictable the target is and in where the predictive signal is located. We construct a finite-support synthetic testbed with a known binary posterior and the same public semantic facts available to every learner. In three core conditions, an identical two-point posterior marginal is determined by either a current warning, dependency topology, or the order of earlier warnings; a fourth topology-history interaction condition matches Bayes Brier opportunity but not the full posterior marginal. Across 2,304 development runs of specified tabular, static-graph, size-control, and graph-plus-GRU systems, rankings are not uniform: the temporal system has better mean size-integrated normalized regret than the static system in both ordered-history conditions, while the topology and interaction contrasts reverse direction across opportunity levels. Large tabular-static gaps also appear in the current-state control, where topology and ordered history do not determine the posterior, so rankings cannot be read as direct measures of structural-information use. Exact expected-risk evaluation removes sampled evaluation-label noise. In selected post-hoc replays, learned content pathways lose prediction-relevant distinctions although the underlying input information remains available; one selected failure is rescued by a persistent learning-rate reduction, whereas a one-step reduction and an analogous intervention in another selected trajectory are not. A planned confirmatory stage was not run because its precision target was infeasible within the replication cap. The results motivate separating predictive opportunity, posterior-determining information location, and stable learnability.