What a genome corpus can and cannot certify about CRISPR antiviral escape: resolution floors, joint coverage, and a worked audit of an HSV-1 guide pair

Multiplex CRISPR antivirals rest on an escape argument: cut a viral genome at several conserved sites at once, and no single repair event can restore an uncuttable, viable genome. That argument is almost always supported by per-guide conservation percentages computed against a public sequence corpus. We show that this evidentiary chain has two structural weaknesses, both properties of the method rather than of any design. First, the corpus imposes a resolution floor: a site absent in 0 of 183 complete HSV-1 genomes is still consistent with a population absence frequency of 0.0162 (Clopper–Pearson 95% upper limit). Sampling resolution, not the repair model, sets how low an escape probability can be *demonstrated* — a single site cannot be certified below ~1.6 × 10⁻², a pair below ~2.6 × 10⁻⁴, and the certifiable minimum guide count for a 10⁻⁶ per-genome threshold is four, not the three a point estimate implies. Second, joint intactness is not the product of marginal conservations and cannot be recovered from a per-guide table; it requires a per-genome presence matrix. A published clinical lead pair with marginals 0.847 and 0.978 has joint intactness 0.825 — below either marginal — because its guides fail in different isolates. As a case study we audit the EBT-104 guide set (Amrani et al., 2024): their ICP0g2 site carries a recurrent G→A substitution at spacer position 9 in 25 of 183 genomes, and 96% of the pair's modelled escape traces to isolates that have already lost it. Changing one guide raises joint intactness to 0.978 and improves every off-target measure. Absolute escape probabilities span 3.1 orders of magnitude while guide-set rankings hold in 27 of 37 swept settings — including a spectrum measured in post-mitotic human neurons, which lowers escape without changing any selected set: the model ranks, it does not measure.