A test for intent

This paper develops a reusable protocol for testing whether a candidate internal-state readout captures a model's strategy rather than merely reflecting its topic or the immediate question. The protocol requires held-out prediction of the realized public response, performance exceeding a private-note text baseline, failure under controls that remove or relocate the question, and four additional statistical criteria. On the 114-row evaluation bank the tested probes showed predictive advantages but both representations had already incorporated the question; the four additional criteria remain unmet. Consequently the prototype bank remains empty and no readout is validated. The paper presents the evaluation protocol; it does not demonstrate that the proposed regularizer reduces deception.