The Deterministic AI Runtime wraps any inference pipeline and produces a cryptographic Merkle proof for each run. Input, output, parameters, and timestamp are all committed. The same input always produces the same proof — fully auditable, fully replayable.
Each inference run builds a Merkle tree: leaves are the hashes of individual pipeline steps (tokenization, embedding, inference, output). The root commits to the entire computation. Any single bit change in any step changes the root — tamper-evident by construction.
The CGK uses Banach's fixed-point theorem: a contractive operator on a complete metric space always has a unique fixed point, reachable by iteration. This guarantees distributed convergence without Byzantine agreement protocols — the math proves it, not the trust model.
The Counterfactual Sovereignty Engine generates proofs of exclusion — cryptographically proving what the AI did NOT consider in a given decision. This is the inverse of standard explainability: it bounds the decision space by proving which inputs were excluded, not just which were included.
Three specialized agents with parameterized bias profiles debate a decision and reach weighted consensus. No human approval required. Transparent pricing algorithm. The consensus result is hash-committed — auditable.