Per-Tick Certification of Spike-Exact Fixed-Point Spiking Neural Networks

A per-tick certificate for bit-exact spike-train preservation under fixed-point quantization. Under hypotheses H1-H5, R_cert > 1 implies that no firing decision differs between the two arithmetics. The proof is a hand-written sketch, not machine-checked. Observed corpus result. On 18,375 in-domain LIF configurations, the gate has zero observed false positives at 74.17 % coverage. A further 460 adversarial configurations inside H1-H5 show no observed violation. Sixty saturating cases are outside H3 and are reported only as out-of-domain observations. Statistical scope. Ranking AUCs are descriptive. Configuration-level intervals and DeLong statistics do not account for architecture-by-seed clustering and are not used for population inference. The contribution is conditional soundness with usable empirical coverage, not superior ranking power. Status. Preprint, not peer-reviewed. Bilingual (EN + FR).