What a Behavioral Score Can Certify: An Identification Calculus for Prediction Under Partial Observation and Consequential Use

This paper is not a new behavioral score, a psychometric model of latent behavior, or a measure of human predictability. It is an identification calculus for determining what claims a predictive result can support, developed in the domain where the failure modes are sharpest. Behavioral-prediction studies often report a predictive score and infer a stronger conclusion: that behavior is regular, that a latent disposition has been recovered, that little relevant information remains unobserved, or that deployment changed behavior. These conclusions require different evidence and do not follow from predictive performance alone. The calculus separates observer-relative forecastability, model deficit, representation-relative residual uncertainty, legibility loss, and post-deployment causal effects. Under logarithmic loss it derives an exact budget in which a screening-defect term prices the failure of a declared representation to screen the action from the observer, so the accounting survives mis-declaration rather than assuming it away. Three results follow. Trace-law non-identification: one observable law is compatible with representations spanning an exact segment, so improving a model cannot convert a trace result into a representation diagnosis. Exactness: the sensitivity set traced by declared caps is attained at both endpoints, not merely contained. And a dichotomy: a representation measurable with respect to the audited observer identifies its own decomposition, while any latent content beyond it leaves the decomposition unidentified — the content that makes a representation worth declaring is the content the trace cannot price. A further result shows that before-after drift does not identify a deployment effect. A separate gate governs the noun rather than the number. Calling a result behavioral prediction requires certifying that the recorded target is the focal actor's action rather than a status, a jointly produced event, or an instrument reading. Targets are classified before they are scored, and the same certificate carries a different noun in each class. Applied to a labour-market panel, the procedure yields a calibrated incremental floor under a declared bootstrap convention and zero under a valid distribution-free construction. The honest output at these sample sizes is a diagnostic rather than a certified magnitude: it tells a study whether it has the data to license one, and the rows required are of order 3.7 x 10^5 for an eighteen-cell observer against the 1,635 available. Target certification precedes score certification: a behavioral score cannot certify behavior until the study has certified that its target is behavior. All numerical results are produced by the companion package (https://doi.org/10.5281/zenodo.21599025). A full run exits zero if and only if every asserted invariant holds; test counts, dependency versions and the SHA-256 of the archived results are generated rather than stated by hand.