PARS: A Constraint-Driven Reasoning and Verification Architecture for AI Freehand Exact-Binary Generation

PARS/Poole Adapative Reasoning Stack (Parse, Branch, Transform, Perturb, Test, Reconstruct, Build) is a constraint-driven reasoning and verification architecture applied in this paper to AI freehand exact-binary generation: experiments in which a generative model authors the complete byte-level representation of a machine-consumable binary while specified conventional generation paths such as compilation, assembly, linking, and prebuilt executable substitution are excluded from the claimed path.

The paper distinguishes executable correctness from executable provenance and defines separate evidence classes for exact artifact identity, reconstruction from recorded bytes, audited generation paths, and procedural visual generation. It introduces the PARS Freehand Exact-Binary Protocol (PARS-FEBP), explicit hard-invariant replay, independent reconstruction, cryptographic identity verification, adversarial provenance testing, and PASS / FAIL / INCONCLUSIVE classification.

Historical evidence includes directly represented Linux ELF executables, a WebAssembly module, procedural visual and audiovisual systems, Binary Brains, a purpose-built virtual computer, byte-level failure-and-repair lineages, and an independently audited experiment that remained INCONCLUSIVE despite successful execution because the stronger upstream provenance claim was not established.

The paper also defines a prospective evaluation protocol for comparing PARS-FEBP with simpler prompting approaches. It does not claim that PARS has already been prospectively demonstrated to outperform those approaches, that exact-binary generation generalizes to arbitrary software, or that prompt/context adaptation constitutes model-parameter learning.