The geometry of AI validation: From structural blindness to reusable audits

This work studies how evidence collected at one AI search budget can support validation at other deployment budgets. In best-of-N systems, increasing search changes which candidate errors matter, so precise evaluation at a small search width may leave larger searches structurally unresolved. The paper derives an exact ambiguity frontier and then develops a constructive solution: retaining candidate ranks and truth labels allows a shared audit to estimate reliability across many search widths. Minimax theory separates the costs of acquiring expensive truth labels from observing additional candidates and provides matching upper and lower bounds, including adaptive acquisition. Experiments on mathematical reasoning and code generation show that reusable, search-aware audits can substantially reduce worst-case validation error across deployment widths.