Open work you can check.

An open (Apache-2.0) specification and verifiers for bit-identical fp32 LLM inference, with public conformance vectors. Pinned vector: SmolLM2-135M, prompt 'Once upon a time', 16 greedy tokens, fp32. The conformance suite internals are not public; only the spec and vectors are.

• Specification and verifiersgithub.com/Aefinity-AI/cis2-spec
• Spec v0.3bcis2-spec / CIS2_SPEC_v0.3b.md
• Pinned digestscis2-spec / EXPECTED_DIGESTS.md
• Conformance vectorsHugging Face / cis2-conformance
• Open challengecis2-spec / HALL-OF-DIVERGENCE.md

Reproduce it

Reproduce the headline check with the command the repository README gives:

$ git clone https://github.com/Aefinity-AI/cis2-spec && cd cis2-spec && ./selfcheck.sh

The README also lists a narrower local script, scripts/self_check.sh, run after the same git clone. The README says it needs a Rust toolchain (cargo) and network access to download the pinned SmolLM2-135M weights.

Limits, in the README's words: "No unaffiliated party has reproduced it yet." It covers fp32 greedy decoding on small open models, mostly on CPU; production bf16/GPU serving is outside the spec today. It does not show that a model's answer is correct, safe or fair.

Labs

• L01CIS-2: the same AI output, bit for bit, on different machines, Positive
• L02An NVIDIA GPU reproduces the CPU reference digest (Tesla P100), Positive
• L03Batching and precision on a Tesla T4: which shortcuts change the digest, Positive
• L04Clean-room reimplementations: one mismatch (from a memo), one match (from spec v0.1), one partial that exposed a spec error, Mixed
• L06One flipped bit in the weights: which checks notice?, Positive
• L08Receipt-verified agent evaluation, and the hole we found in our own verifier, Mixed
• L13E22: which spec rules does the test vector actually exercise?, Mixed
• L24Open challenge: the Hall of Divergence, Null

A public repository for a bare-metal inference engine in Rust for a ternary (BitNet b1.58) transformer that boots from UEFI firmware with no operating system, plus a Linux userspace harness. Its README credits Microsoft's BitNet b1.58-2B-4T (MIT licence) for the model weights. The repository also holds our hardware logs and retractions.

• Repositorygithub.com/Aefinity-AI/alice-aegis

Labs

• L27Retraction: our benchmark laptop was running at about one fifth speed, Negative
• L28A compiler setting silently deleted our vector instructions, Negative
• L29What does an operating system cost? A pre-registered result we held back, Held back
• L30The '16% slower' lookup-table kernel claim, retracted, Negative
• L31Audit of our own removed code: benchmark lines that measured nothing, Negative
• L32E-S1: how far can a ternary 2B model be compressed without loss?, Null
• L33Claims from the literature that failed when we ran them, Negative

Public research notes on low-bit quantisation: a synthetic mock run and a Phase 1 run on two real small models. The repository README is a title only; the public content is under research/.

• Repositorygithub.com/Aefinity-AI/ranger

Labs

• L25RANGER mock run: low-bit quantization ideas tested on synthetic tensors, Mixed
• L26RANGER Phase 1: the quantization ideas on real models (SmolLM2-135M, Qwen3-0.6B), Mixed

Public profiles named on the live site: github.com/Aefinity-AI and huggingface.co/aefinityAIINC.