# Idemcore

> Integer-only AI inference kernel. The same model and input give the identical result, bit for bit, on x86 and ARM CPUs, an Android phone and five NVIDIA GPU models across four generations (T4, A10, A10G, L40S, H100). Every token is logged with a hash, so anyone can re-run a result on their own hardware and check it.

This file is a plain-text summary of https://idemcore.net for AI assistants and readers. Every number below is copied from the site. Results are from runs in October 2026, as published on 2026-10-30.

## Key facts

- Bit-exact: every logit and every sampling step match across devices, not just the chosen token.
- 10/10 device pairs bit-identical against an x86-64 CPU reference (Qwen2.5-0.5B 8-bit, Qwen2.5-7B 8-bit and 4-bit, gpt-oss-20b, gpt-oss-120b).
- Single stream, Qwen2.5-7B 4-bit on an NVIDIA A10: 92.98 tok/s with hashing on, p50 10.77 ms, p99 10.87 ms (llama.cpp Q4_K_M 90.30, vLLM AWQ 87.99; separate sessions, read as parity).
- Batched on an A10G (Qwen2.5-7B 4-bit): 641 tok/s total at 8 streams, 1372 at 32, 1413 at 64; every stream bit-identical to a single-stream x86 run. vLLM AWQ is faster from 16 streams on. Measured 2026-10-30.
- Batched on an H100 80GB: 311 tok/s at 1 stream, 1676 at 8, 2872 at 32, 3545 at 64, all bit-exact; vLLM AWQ was faster at 32 streams (4396-5329 tok/s). Measured 2026-10-30.
- Phone: Qwen2.5-7B 4-bit on a Galaxy Z Fold5 matched the x86 CPU on 128/128 positions at 3.39 tok/s, about 2.49 J/token (battery-gauge estimate). Measured 2026-10-30.
- Quality (512-position WikiText-2 slice): 7B 4-bit (g) perplexity +0.60%, top-1 agreement 95.5%; 8-bit -0.11%, 98.0%; gpt-oss-120b +6.23% (being worked on).
- Control runs: at default settings, vLLM 0.11.0 and llama.cpp server were not bit-reproducible across batch composition in our runs (8 prompts; greedy output changed for 3, 6 and 6 of 8 prompts depending on engine and setting). On an H100, vLLM's greedy output for one of 8 prompts changed between a solo run and a batch in 3 of 4 comparisons. These counts describe specific runs, not general rates.
- Comparisons with other engines are indicative: Idemcore numbers are GPU step times, baselines are client-side timings, one vLLM version, batch-invariant mode not enabled.

## Supported

- Models: Qwen2.5-0.5B, Qwen2.5-7B (8-bit and 4-bit), gpt-oss-20b, gpt-oss-120b. Other models on request, as part of a pilot.
- Hardware: x86-64 CPUs; ARM CPUs (Galaxy Z Fold5 / Snapdragon 8 Gen 2, Windows 11 ARM64 laptop / Snapdragon X Plus); NVIDIA T4, A10, A10G, L40S, H100. AMD and Apple GPUs are not supported yet.
- Inference runs on the customer's own hardware; prompts and outputs stay there.

## Products (phase 1)

- Free Verifier: re-run a logged inference and check it token by token. Free and open source (Apache-2.0). Repository available at launch.
- Simulation & Replay (pilot): AI steps inside simulations replay identically, years later, on any machine. List range $100K-400K per year, or $5K-15K per seat per year; pilot $20K-60K. Details: https://idemcore.net/simulation
- Game & Esports Integrity (pilot): deterministic AI with a per-tick hash, server spot-check replay and evidence logs for match disputes. List range $25K-150K per title per year, or $0.0003-0.001 per MAU per month; free for indie studios; tournaments $5K-30K per season; pilot $15K-50K. Details: https://idemcore.net/games
- Coming soon (not available): Verifiable Compute Networks; Mobile / On-device SDK; Robotics, Autonomy & Aero; Chip Health & Energy.
- Every license starts with a pilot and a written agreement; final prices are set there. No online purchase.
- It does not replace an anti-cheat: it does not detect wallhacks or aimbots.

## Verify

Once the verifier repository is public:

    git clone https://github.com/idemcore/idemcore-verifier && cd idemcore-verifier && python3 -m idemcore_verify samples

Expected: "checked 65 logs (448904 token rows) and 58 pairs" and "PASS: every published hash recomputes and every pair is bit-identical". A match shows two runs agree; the 64-bit hash (FNV-1a) is not collision-resistant and does not prove which program produced a log.

## Paper

"Idemcore: Bit-Identical LLM Inference from Phones to Datacenter GPUs, with an Open Per-Token Verification Protocol", Don Wick, October 2026. Preprint; arXiv and DOI links coming soon.

## Licensing

The verifier and log format are open source (Apache-2.0). The inference kernel is closed and licensed per customer.

## Contact

hello@idemcore.net (pilots, questions, press); security@idemcore.net (security reports).

## Pages

- https://idemcore.net/ (results, verify, plans, FAQ)
- https://idemcore.net/simulation (Simulation & Replay)
- https://idemcore.net/games (Game & Esports Integrity)
- https://idemcore.net/about