Certification Is Not Evidence: Evidence Classes and the Assessment of Safety Arguments in Battery Energy Storage Procurement

On 13 August 2026, Clearstone Energy's appeal for a 150 MW storage system in Axminster, England, was dismissed because the inspector was not persuaded that deflagration mitigation strategies would work effectively for cabinets spaced 10 cm apart. Separation distance is among the parameters that UL 9540A installation-scale testing allows to be assessed. The refusal did not follow from any breach of standards: it followed from available evidence failing to cover the deployed configuration. A previous paper by this author argued that the capacity to sustain a safety argument is a property of the control architecture, and translated that argument into technical specification items. This paper addresses the next question, which Brazil's LRCAP 2026 technical qualification and the December auctions place before developers, insurers, lenders and licensing authorities: once the vendor has answered, how does one assess whether the answer demonstrates anything? The central proposition is that there is no strong evidence in the abstract — there is evidence adequate or inadequate to a given claim, configuration and operating regime. Three theses follow. Epistemically, each claim admits only certain classes of evidence, and most procurement disputes arise from accepting a class below what the claim requires. Structurally, the absence of a higher-level test is in itself neither favourable nor unfavourable: its meaning depends on why testing stopped where it did. Temporally, evidence has a scope of validity bounded by the edition of the method, the identity of the tested configuration, and the presumed operating regime. The paper proposes seven evidence classes ordered by recoverable information and differentiated, at the upper end, along two independent axes; an explicit sufficiency criterion; a correspondence rule between claim and minimum required class; a claim class concerning authority arbitration among controllers; and a taxonomy of five modes of non-response. It applies these instruments to a constructed dossier and states explicitly what it does not demonstrate.