Benchmarking as Information Design: Disclosure-Conditioned Generalized Accuracy for Strategic LLM Evaluation

Benchmark scores for modern language models are no longer properties of models in isolation: once tasks, protocols, and leaderboard feedback are public, they become functions of the model, the adaptation budget, and the evaluator's disclosure policy. We formalize this view with \emph{disclosure-conditioned generalized accuracy} (\DCGA), the expected hidden-family performance of a model after bounded best-response adaptation under specified pre- and post-disclosure policies. We then propose \emph{capability-aligned disclosure} (CAD), an evaluator-side mechanism that learns transfer-aligned feedback directions from pilot data, penalizes benchmark-local excess, and releases a thresholded public signal at a fixed bandwidth budget. We evaluate CAD on a family-based benchmark with explicit semantic and protocol variation, matched attacker budgets, held-out families, downstream external tasks, and a frozen analysis plan with calibrated clustered-bootstrap uncertainty. Across confirmatory experiments, exact-score release yields the largest visible gains but also the largest exploit gap, while CAD significantly reduces exploit gap relative to exact-score release and significantly improves overall and held-out hidden gain relative to the strongest matched-bandwidth baseline. These results suggest that modern benchmark design is not only a problem of test-set construction, but also a problem of information design: what a benchmark measures depends materially on what the evaluator reveals.