Idemcore: Bit-Identical LLM Inference from Phones to Datacenter GPUs, with an Open Per-Token Verification Protocol

Two correct floating-point LLM engines rarely return the same bits, so a third party cannot check an inference result except by trusting a tolerance. This preprint presents Idemcore, an integer-only engine for transformer inference, together with an open verification protocol: every decoding step emits a 64-bit hash of the full-vocabulary integer logits and a hash of the integer cumulative distribution used for sampling. Anyone can re-run a request on hardware they control and compare logs line by line. We report full-model, end-to-end bit identity across x86-64 CPUs, a Windows 11 ARM64 laptop, an Android phone and five NVIDIA GPU models (T4, A10, A10G, L40S, H100) for Qwen2.5-0.5B, Qwen2.5-7B (8- and 4-bit) and gpt-oss-20b/120b. On a same-harness H100 serving benchmark of Qwen2.5-7B (1 November 2026), Idemcore serves 2729–4008 tok/s aggregate at 32–256 concurrent requests; vLLM and SGLang are faster today. Idemcore trades throughput for bit-identical output and per-token hash receipts at every batch size. The verifier (Apache-2.0), log specification, methodology and sample logs will be released at https://idemcore.net; the inference engine itself is closed.