Glossary / Deterministic Replay

Definition

What is Deterministic Replay?

Reproducible AI inference

Deterministic replay is the ability to re-run a past AI decision and reproduce the exact output bytes. Because GPU inference is non-deterministic by default, Trinitite runs a determinism-fixed kernel so the same prompt, seed, and weights reproduce the same response — and signs a hash-chained receipt an auditor can independently verify.

On modern GPUs, floating-point reductions re-order across batches and CUDA streams, so the same prompt can drift silently between calls. A batch-invariant kernel locks that accumulation so output is bit-for-bit reproducible months later.

Every governed call mints a signed Deterministic LLM Inference Receipt — model digest, seed, input and output digests, policy hash, and kernel fingerprint — chained append-only and anchored to an RFC 3161 timestamp authority and the Sigstore Rekor transparency log. An auditor replays a decision from the public verifier with no login and no NDA.