Rule identifiability in elementary cellular automata under dynamical noise

Exact computation (no sampling, no seeds) of rule identifiability in elementary cellular automata on a ring of 12 cells under dynamical bit-flip noise: zero-noise consistency classes for all 256 rules x 4096 initial states, the identifiability rate rho(f; eps) = lambda(eps) min_c nu_c(f; eps) for all 88 rule classes on a noise grid with the best noise level eps* per class, finite-horizon evidence margins from every initial state, and a boundary check with radius-2 generators. Includes code, data files with a SHA-256 manifest, and the note as PDF.