Source · Tier B · PreprintSource · EigenAI: Deterministic Inference, Verifiable Results
EigenAI: Deterministic Inference, Verifiable Results
D. Ribeiro Alves, V. Patankar, M. Pereira, J. Stephens, N. Vaziri, S. Kannan. 2026. arXiv.
| Original | https://arxiv.org/abs/2602.00182 |
|---|---|
| arXiv | 2602.00182 |
| Version | v1 (30 Jan 2026), the only version, read in HTML. |
| Accessed | 2026-09-25 |
| Note | Developer's paper on its own system (all authors give eigenlabs.org addresses); statements about EigenAI are provider-reported. Describes a deterministic inference engine built on llama.cpp with custom GEMM and reduction kernels, and an optimistic re-execution protocol in which a stake-weighted committee re-executes challenged outputs inside TEEs. Table 5 reports bitwise-identical outputs on the same host and GPU and across hosts with the same GPU SKU, and a 0.0% match between A100 and H100. |