Canonical Execution Semantics for Stochastic Program Generators

AI systems increasingly generate computation---through large language models, program synthesizers, and agent frameworks---rather than having humans write it directly. When the generator is stochastic, independently generated programs that represent the same intended computation arrive in different surface forms, producing different hashes, different provenance records, and failed replay comparisons. We argue that execution systems for stochastic generators require a canonical execution boundary: an architectural invariant that partitions the pipeline into a stochastic upstream (generation, admission) and a deterministic downstream (planning, execution, provenance). Stochasticity does not propagate beyond this boundary. We report on BLISP, a deterministic execution system that enforces this boundary through four mechanisms: (1) typed specifications that constrain stochastic generators to propose structured parameter tuples rather than executable expressions, eliminating surface-form divergence by construction; (2) a canonicalization pipeline whose traced normalization algebra collapses 278 surface forms to 235 canonical operations; (3) 8-layer execution hashing that decomposes provenance into distinct semantic layers for per-layer fault localization; and (4) description/identity separation in a content-addressed capability registry, where discovery metadata evolves independently of execution identity. We evaluate canonicalization collapse under 1,200 stochastic LLM generations, replay determinism across 50 independent runs, and provenance stability across registry evolution. The results support the central thesis: AI-generated computation should not execute stochastic outputs directly; stochastic proposals should first compile into deterministic semantic objects before execution begins.