Aefinity AI

Aefinity AI · Verifiable, private, autonomous AI

AI you can verify, keep private, and put to work.

Private AI built to stay in your building. A computation baseline that never moves. AI-run operations held to a lab standard. Our labs, good and bad, are on the public record.

Book a briefing What we offer
Video · 4 minutesWatch the 4-minute explainer What CIS-2 checks, how anyone can verify it, and what it does not prove.

We are not another AI platform or cloud service. Our mission All 33 labs

The record so far

Each mark is one lab, grouped by outcome. Select a mark to open the lab.

9 Positive

14 Mixed

6 Negative

3 Null

1 Held back

What we offer

Four ways to put verifiable AI to work.

  • Private AI

    ALICE

    Private AI built to run on your own hardware, with no cloud connection. Pilot-stage, scoped with each customer.

  • Open standard

    CIS-2

    The same model and input give the same output, to the last bit, on every machine that follows the spec. Open source (Apache-2.0).

  • Proprietary

    LLM edition

    CIS-3

    Our proprietary edition for large language models. Briefings under NDA, on request.

  • Services

    Autonomous Operations

    AI-managed workflows, set up and run to our lab standard: every action logged, every risky step gated for a human, every result checked before it ships. We run our own lab this way.

On the public record

  • 33published labs
  • 84registered experiments on a public, hash-chained ledger
  • C + RustOpen spec with verifiers in C and Rust
  • One digestSame digest on AVX2, scalar x86, AVX-512, an aarch64 phone and a Tesla P100 GPU, in our own runs

Autonomous Operations

Your work, run by AI, to a lab standard

We design, set up and manage AI-run workflows for office operations, customer service and research. Every action logged. Every risky step gated for a human. Every result checked before it ships.

See how we run it Book a briefing

How we work

  1. 01

    Pass mark set in advance

    Pre-registration: a pass mark is written before the run, then a verdict is recorded.

  2. 02

    Every number sourced

    Every number comes with its public source and a short verbatim quote.

  3. 03

    Failures listed

    Failures, nulls, retractions and corrections of our own work are listed with the successes.

  4. 04

    Run by us

    Every result is run by us. No unaffiliated party has reproduced any of it yet.

The hard questions

Ten questions AI raises that we can test, each with what we found so far and what is still open.

  1. 01Can two different computers produce the exact same AI output, bit for bit?3 labs
  2. 02Does the same hold on graphics cards and under the usual speed shortcuts?2 labs
  3. 03If someone changes a model's weights without telling you, how would you know?3 labs
  4. 04Can you trust a record of what an AI agent did, and what does it cost?3 labs
  5. 05Can someone who has never seen our code rebuild the same bits from the written description alone?3 labs
  6. 06How do we know our own tests test anything?7 labs
  7. 07How can an outsider tell that a lab did not tune results after the fact, or quietly drop failures?2 labs
  8. 08Do speed numbers survive scrutiny?5 labs
  9. 09Do published tricks for shrinking AI models actually work?4 labs
  10. 10Can a tiny model be evaluated in a way anyone can repeat exactly, and what does it score?1 lab

Six labs, good and bad

Chosen for balance: two positive, two mixed, two negative. The outcome label on every lab is ours, and the other 27 are one click away.

  • L01
    Positive

    CIS-2: the same AI output, bit for bit, on different machines

    With the arithmetic pinned in writing, one fp32 digest reproduced across AVX2 and scalar x86, an aarch64 phone and an AVX-512 cloud CPU, in our own runs. Until an outsider runs it, the claim stays 'reproduced by us'.

    Q1

  • L06
    Positive

    One flipped bit in the weights: which checks notice?

    In the larger run, among the 247 trials where the output actually changed, the text-replay column missed 182 and the top-k/receipt-style check missed 82. Some flips (43) changed nothing in the computation. A check on the computation itself is by definition silent only on flips that changed nothing, but that 0/247 is true by construction, so the informative finding is how often the weaker checks miss. Single prompt and model, harness-chosen flips, no adversary.

    Q3

  • L08
    Mixed

    Receipt-verified agent evaluation, and the hole we found in our own verifier

    A receipt can verify and still be filed on the wrong item. We found that by trying to break our own verifier, and binding each receipt to its item closed the case we found.

    Q4

  • L13
    Mixed

    E22: which spec rules does the test vector actually exercise?

    Passing the published test does not prove every rule matters. 12 of 16 single-rule breaks were caught. The tie-break and the denormal-flush pin are protected only by inputs this vector never contains. One of our own null results turned out to be a vacuous edit that tested nothing, and we corrected it in place. Correcting it exposed a weak spot in the spec's own rsqrt check, which tests only a value where every implementation agrees.

    Q6

  • L27
    Negative

    Retraction: our benchmark laptop was running at about one fifth speed

    Measure the machine before measuring the code. We retracted every earlier number from that box instead of reporting a speed-up.

    Q8

  • L21
    Negative

    E41: the spec pins how text becomes tokens, but not how tokens become text

    Matching digests is not the same as matching meaning. The missing piece is a text-level definition; whether v0.4 specifies detokenization or declares it out of scope is a spec-owner decision.

    Q5

All 33 labs

What we have learned

Each lesson cites the labs it comes from. All are from our own runs, at the scopes stated in those labs.

  1. 01Check the computation, not just the text it produced.
  2. 03A spot check can give a comfortable wrong answer.
  3. 05Measure the machine and the shipped binary before measuring the code.
  4. 12Everything so far has been run by us.

All 12 lessons

Our test standard

This page describes the protocol as our public repository documents it. It is stated and applied by us and has not been audited by an outside party.

  • Pre-registration: a pass mark is written before the run, then a verdict is recorded.
  • A public, salted, hash-chained fingerprint ledger of the private register, with a verify_ledger.py script and an OpenTimestamps proof file (LEDGER.tsv.ots).
  • No timing or throughput numbers in the spec repository.
  • Provenance on every number.
  • Defects recorded as numbered errata in CHANGELOG.md; pinned digests do not move silently.
  • Mismatches are welcomed as results.

Open the PR whether your digests match or not. A mismatch is a result.

cis2-spec / CONTRIBUTING.md

The standard, and what it cannot prove

Projects

  • CIS-2

    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.

  • ALICE (alice-aegis)

    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.

  • RANGER

    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/.

Book a briefing

Tell us what you need to verify, protect or automate.

We reply personally. Running inference already? See the replay pilot.

Book a briefing justin@aefinityai.com 409-320-1536

Our aspiration

We hope to earn trust as a standards company that leads by example: honest about every result, and every result repeatable.

This is an aspiration, not a claim.