From Claims to Discriminating Tests: A Computational Framework for Measurement-Explicit Analysis of Anomalous Capability Reports

Unusual capability reports pose a methodological problem: appearance, mechanism, and observations needed to distinguish alternatives are often conflated. We introduce the Anomalous Capability Simulator (ACS), a measurement-explicit framework that converts reports into 21 canonical experiment classes, typed parameters, synthetic models, competing explanations, and discriminating tests. Its Claim-to-Test Compiler formalizes candidate mechanisms and admissible measurements, computes minimum-cost separating sets, compares greedy selection, constructs minimum worst-case adaptive decision trees, and extends pairwise discrimination to noisy Bernoulli observations using finite-sample Bayes accuracy. A validation benchmark contains 6,300 rows across 21 experiments, three stipulated worlds, and 100 seeded trials per experiment-world cell. All 21 canonical fixtures are deterministically resolvable at fixed cost 6; synthetic noise raises the maximum pair cost at 95% target accuracy from 9 in the easy regime to 27 in the moderate regime and 807 in the hard regime. Stress tests recover a strict greedy failure and an admissibility obstruction. The Local Emergence experiment shows that an apparent outcome can remain compatible with competing mechanisms until mass balance, boundary integrity, and provenance are examined. ACS shifts analysis from whether a report seems extraordinary to which observations make rival explanations distinguishable, while preserving unresolved cases and separating reproducibility from empirical confirmation.