Kenshiki Labs develops a runtime control plane and proof layer for consequential AI decisions, recording model inputs, claims, retrieved evidence, checks, and outcomes to enable reconstruction in audits and discovery. Its bounded-synthesis pipeline processes governed AI requests in three phases—Prepare, Propose, Prove—binding evidence beforehand, constraining reasoning within boundaries, and producing replayable records afterward. Key components include Prompt Sanitizer, which authenticates callers, binds identity via REBAC authorization, validates inputs, and enforces policies before retrieval or generation, ensuring evidence scoping and audit trails. The system maintains a per-decision evidence boundary and claim ledger with provenance, governing execution without altering the model.