CIS · Checkable inference
The baseline that doesn't move.
When the AI computation itself is pinned to the last bit, any change you see comes from the thing you changed, not from the machine.
- AVX2 x86same
- Scalar x86same
- AVX-512 cloud CPUsame
- aarch64 phonesame
- Tesla P100 GPUsame
Apache-2.0
CIS-2 · open standard
CIS-2 is an open specification with verifiers in C and Rust for bit-exact fp32 inference, with public conformance vectors. In our own runs, one digest reproduced on AVX2 and scalar x86, an AVX-512 cloud CPU, an aarch64 phone and a Tesla P100 GPU.
The labs behind it
CIS-3 · LLM edition · Proprietary
Checkable inference, taken to large language models.
Available to prospective licensees and collaborators under NDA, on request.
Licensing use cases
What a fixed baseline makes possible
The baseline for every AI test harness
A non-changing, per-hardware reference that evaluations, safety tests, red-teaming and regression suites run against, so drift in results means drift in the model, not noise.
Hardware and cloud qualification
Prove a new server, GPU, driver or cloud instance produces identical results before it goes into service.
Model supply-chain integrity
Detect silent changes to model weights.
L06 · One flipped bit in the weights: which checks notice? L07 · Realistic edits to the embedding table, including one hidden row
Audit and dispute evidence
Replay a run and show the record matches, bit for bit.
Vendor verification
Check that an AI provider ran the model and settings it says it ran.
Reproducible research and evaluation
Results anyone can re-run exactly.
Book a briefing
Put a baseline that doesn't move under your test harness.
CIS-2 is open under Apache-2.0. Commercial support, integration into your test harness, and CIS-3 licensing are available.