Evidence Below the Conformal Rank Floor

With $m$ calibration observations, split conformal prediction cannot certify any miscoverage level below $1/(m+1)$; below that floor the distribution-free interval is unbounded. E-values are unbounded above and multiplicative, which suggests they might evade the limit. Under the same exchangeability assumption they cannot. Specializing a known permutation-orbit result, we show that every calibration-symmetric distribution-free e-value is at most $m+1$. This bound is sharp. It does not arise from retaining only ranks: an e-value may use the complete calibration sample, its spacings and any fitted structure, and the ceiling still applies. The levels demanded in risk and safety applications therefore require a tail-model assumption; e-values do not remove that assumption, and they add no distribution-free resolution. What they add is that the model's evidence becomes an explicit betting factor that later observations can settle. Simulations quantify that trade-off. A favorable tail-class match does not restore the declared level: at $10^{-5}$ with $m = 999$, a plug-in bound is conditionally conservative in only about a third of fits, and its marginal exceedance probability is eleven times the declared level. One-sided $95\%$ upper bounds approach nominal only where the excess law is exactly generalized Pareto, at substantial cost in width; see Section 4.