The Benchmark Contamination–Capability Separation Problem in Foundation AI

Benchmark scores are often treated as direct evidence of model capability. In foundation AIthis assumption is unsafe. Public tasks can enter pretraining corpora, benchmark formats canbecome familiar through prompt engineering, leaderboard optimization can tune models to visibleevaluation patterns, and benchmark-specific heuristics can imitate ability without stable transfer. Thispaper introduces the Benchmark Contamination–Capability Separation (BCCS) model: a standalonerelease-gated framework for separating apparent benchmark success from clean capability evidence.A benchmark claim is represented as a scored object with public-score, private-score, paraphrase,transfer, negative-control, exposure, canary, prompt-familiarity, and repeated-run stability channels.The central quantity is the contamination-capability gap, defined as the public score minus a cleancapability core computed from private, paraphrased, and transfer channels. A model claim is releasedonly if the gap, exposure risk, negative-control leakage, transfer fragility, and run variance all remainbelow explicit thresholds. The contribution is not another leaderboard. It is a causal evidence contractfor when a benchmark score may be treated as evidence of capability. A 3000-record diagnostic ledgeris generated and audited by a reproducible Python harness. In this ledger, only 88 of 3000 recordspass release gates, while 786 become high-value holds and 2126 are blocked. The most frequent failedgate is contamination gap, followed by capability-core insufficiency and transfer instability. The resultsupports a clear thesis: foundation-AI evaluation must report not only how high a model scores, buthow much of that score survives contamination-resistant separation