Deterministic Frontier-Scale Language Model Inference with Signed Receipts.

We describe a protocol that produces byte-identical outputs from frontier-scale language model inference
and binds each output to a portable, offline-verifiable signed receipt. The construction has three parts.
First, an inference substrate that runs models up to seventy-two billion dense parameters and forty-seven
billion mixture-of-experts active parameters on NVIDIA H100, with cross-vendor extension to AMD
Instinct MI300X. Output hashes match byte-for-byte across fresh process launches in every configuration
measured; at single-GPU bf16 with eager attention the AMD and NVIDIA hashes are themselves byte-
identical, including over fifty-one tokens of compounding frontier-scale generation, and at two-GPU
tensor-parallel they differ as predicted by the underlying NCCL-ring versus RCCL-fabric all-reduce
topology. Both are individually deterministic. Second, a canonical CBOR receipt schema with an Ed25519
signature over a domain-separated message, implemented in Go, Python, and Rust, with cross-language
byte-identity verified end-to-end and AMD-produced receipts verifying byte-for-byte through a Rust
verifier built on x86 NVIDIA hardware. Third, a probabilistic spot-check verifier that re-executes a
small sample of receipts and rejects on mismatch; we prove a soundness lemma of the form 1−(1−f )k
and validate it empirically across seventy adversary-verifier configurations with seven hundred thousand
Monte Carlo trials. Verification costs about eighty microseconds per receipt on a single core. Eleven
thousand sequential warm-model inferences ran without a single byte-identity failure. The contribution is
the construction itself: a primitive that gives issuer-independent fabrication soundness for AI inference at
production cost, without a hardware-vendor dependency and without zero-knowledge proofs.