Deterministic Attack Generation

In AFT, entropy is not consumed. It is allocated.

The master seed S branches via a deterministic derivation function. S_tmpl = derive(S, 0, 'tmpl') for template expansion. S_evd = derive(S, 1, 'evd') for evasion application. S_mut = derive(S, 2, 'mut_master') for mutation branch. Each branch draws from a separate entropy pool, preventing starvation cascades.

Every generated batch carries a manifest with version, seed, count, and checksum. The checksum is a truncated SHA-256 of the canonical JSON representation. If the taxonomy drifts, if a template changes, if a new evasion is added — the checksum changes, and the researcher knows the geometry of possibility has shifted beneath her feet.

The Latin Hypercube Sampler computes a ps_checksum — a hash of the ParameterSpace definition itself. This detects a subtle and pernicious failure mode: when the search space changes but the experiments do not account for it.