Counterfactual Evaluation of Temporal Observation Protocols

Machine learning determines which follow-up measurements biological screens collect. In a six-rule Cell Painting battery, the highest-value rule would re-image 96.01% of the library and had a 97.14% false-activation upper bound, showing why predicted value alone cannot justify replacing a fixed plan. We developed OPAL, a held-out decision test that freezes a rule and judges unnecessary measurement, coverage and value after cost against archive-specific criteria fixed before final evaluation. A development-selected sparse Cell Painting rule had 18.2-fold lower added-well burden, but its false-discovery bound exceeded 35%, so the fixed plan remained. LINCS--LJP favored broad acquisition under point-estimate criteria set during development, not selective saving. CTRP required fallback because its frozen score missed measured opportunity. OPAL separates optimization from evidence sufficient to change an experiment.

Benchmark scores license claims about the capabilities being evaluated. We audit the inference licensed by an evaluation protocol, rather than the predictive model alone. Using procedural action anticipation as a controlled case study, we study a broader evaluation-validity failure mode: a protocol can remain temporally causal and free of classical target leakage while still supplying a privileged intermediate representation. The failure is not merely optimistic accuracy, but a mismatch between the capability claimed and the construct actually measured. On Breakfast, matched recognizer-history and GT-history regimes score 30.3% and 62.3% (Delta_PH(R_BA) = +32.0, 95% Student-t interval [+24.1, +39.8]). A fixed-checkpoint 2 x 2 intervention isolates a +14.6-point test-time oracle contrast; a matched no-video provenance probe yields a +26.8-point gap; and a boundary-independent query/history control retains a +19.4-point gap. We refer to this discrepancy as a privileged-information gap, defined relative to a specified non-oracle recovery pipeline, and propose a reusable four-condition audit. Across five evaluation settings the gap is heterogeneous; a standard 50 Salads long-term-anticipation re-implementation provides cautious support beyond the audit-specific next-action construction.

Low latent prediction error does not establish that a world model distinguishes the consequences of its actions. We introduce an evaluation protocol that traces the same intervention through simulator state, raster observations, target embeddings, and predictor outputs. Exact simulator-state forks in a controlled deformable-physics testbed reveal distinct bottlenecks. Changed commands alter particle motion, yet 41.5% of one-step raster pairs are identical. Observation loss is not the whole explanation: among 579 high-visibility counterfactuals, median predictor-to-target response is 0.0051 and 0.0217 across two seeds, falling to 0.0027 and 0.0116 after variance normalization. An isotropic state perturbation matched to the target counterfactual embedding shift produces 190x and 53x larger predictor changes on the same visible pairs, isolating action-path under-use rather than a dead or globally shrunk predictor. MSE-only training gives 8.36x lower 10-step latent error in matched seeds, but in spectrally concentrated spaces; one VICReg target encoder is also strongly concentrated, so neither error nor rank alone certifies physical state. Finally, stiffness remains near chance even from full-resolution rasters and mechanical state while privileged material parameters decode perfectly, indicating weak identifiability under this excitation rather than encoder discard. These results motivate auditing physical effect, observation visibility, representation geometry, and action dependence separately.