Preface
On 2 November 2026, Anthropic published a research post on “provable inference”: a way for a provider to prove that a response came from a specific model, without revealing the model’s weights. Japanese press picked it up within days. The work is real and carefully scoped. The coverage is a little wider than the scope.
A proof is for those who agree to be checked.
This post reads the work against what it claims, asks who it can bind, and ends with a guess about where it might lead. The guess is marked as a guess.
1. The Specimen
The primary source is Anthropic’s post, “Progress on provable inference for large language models”, published on 2 November 2026 and modified on 3 November. Its authors come from Anthropic and from several universities and companies. A code repository, FlashProve, accompanies it. A technical report is linked from the post; we did not read it.
The problem the post names is simple. A user cannot tell whether the model that answered is the model the provider advertises. A routing service could substitute a cheaper one. An attacker could tamper with the token stream. The post proposes that a provider commit to a model once, then prove, for a given response, that the committed model produced it.
The post calls itself intermediate progress rather than a finished agenda. The repository describes itself as a research artifact: not maintained, and not accepting contributions.
2. What Is Proved, and What Is Not
How it works. The provider publishes a commitment tied to one set of weights. For each response, the provider’s GPU produces a proof. Anyone can check the proof on an ordinary CPU, without the weights. The method is a software-only zero-knowledge proof, not trusted hardware.
What was demonstrated. The demonstration used Llama 4 Maverick, a mixture-of-experts model of about 400 billion parameters, with roughly 17 billion active per token, on a single H100 GPU. It was not a Claude model.
Costs reported. The one-time commitment step took about 39 minutes. A batch of 24 requests took about 223 seconds to prove, with each proof around 19 MB. Compared with ordinary inference on eight H100s, the post reports proving to take roughly 70 to 95 times longer. Checking a proof is cheap: 17 seconds on 24 CPU threads in the post’s account, including checks of the commitment, and under two seconds per proof according to the repository. The post says earlier open-source systems were orders of magnitude slower.
What is hidden. The weights, the intermediate values, and which experts ran at each layer. The model’s architecture is not hidden.
What is proved. That each checked token is the committed model’s greedy choice, meaning the most probable token at each step, computed on an integer version of the weights. Sampled decoding is not covered. Contexts are limited to 1,024 tokens.
What is outside the proof. The post says the protocol does not establish that hardware is not also doing something else, such as training on the side. Side channels such as timing are out of scope.
3. How the Result Travelled
We read a Japanese press report of 5 November 2026 (the outlet’s name is not in the text we received). Its facts are mostly right: the zero-knowledge method, the hiding of weights, the cheap check on a CPU, the public prototype. Against the primary source, four points need care.
- “Lower cost than existing methods.” True against earlier proving systems. Against ordinary inference, the post reports a large overhead.
- The limits are not mentioned. Greedy decoding only, a 1,024-token context, and a demonstration on Llama 4 Maverick are not in the report we read.
- “Prevent fraud and attacks in advance.” The post describes proof that a response came from a committed model. That is evidence after the fact; it is not a mechanism that stops anything.
- “Seconds to verify.” Plausible per proof, though the two Anthropic sources give different figures depending on what is counted.
None of this makes the work smaller. It makes it narrower than the headline.
4. A Tool for the Willing
A proof helps only if the party being checked takes part. Anyone running their own model, whether a state, a company, or an individual with open weights, has no reason to produce one. For a system that is already unconstrained, nothing here constrains it.
What remains is a role closer to a condition of trust than a barrier. If a bank, a regulator, or a hospital decides it will accept only outputs that come with a valid proof, then those who cannot produce one are not blocked, but they fall outside the circle of outputs that institution will rely on. This is our analogy, not a claim from the post. Web certificates work somewhat this way: they do not stop anyone from running an insecure site, but they decide which sites visitors will trust.
A second point concerns where trust goes. The repository says the commitment identifier must come from a party the verifier trusts, not from the server being checked or from files sent with the proof. So the proof does not remove trust. It moves it to the question of who publishes the commitment, and who vouches that the committed weights are the ones that were audited. Our reading is that the proof is most useful for closed models, since an open model can simply be rerun.
5. A Speculative Note: If the Target Moves Faster Than the Proof
This section is a guess, not a finding.
The system binds a provider to one set of weights. The commitment is made once per model, and the post says it holds for about sixteen million requests. A proof then says an output matches the committed model. All of this presumes a model that stays still long enough to be committed to, audited and approved.
Suppose, for the sake of argument, that development moves toward systems that take part in improving themselves, so that weights change often. We can imagine a few consequences, none of which we can show. Each new version would need its own commitment and its own audit. The proof would keep saying, correctly, that this output came from that version, while the question people care about, what this version does and whether it is safe, would change faster than anyone could examine it. The proof would stay exact, and the exactness might stop mattering. In that picture, a mechanism for verifying identity could become background noise next to the change it was meant to follow.
There is a smaller worry as well. The system proves one decoding mode, short contexts, and one model. A fast-changing system could drift outside what the proof covers before the proof is extended.
We hold this loosely. Faster proving, shorter audits, or proofs tied to individual training steps could change the picture, and the post itself calls this intermediate progress. It may also turn out that verification becomes part of the improvement loop rather than noise beside it.
6. Checkpoints
We will check on 31 December 2026 and again on 30 June 2027. Whatever we find, we will report it here, including if the picture turns out better than this post suggests.
Checkable by anyone
- Has the repository or the post been updated, for example with sampled decoding, longer contexts, or a smaller proving overhead?
- Has any provider publicly committed to publishing proofs for its production models?
- Has any regulator or procurement rule referred to proofs of this kind?
Checkable only by a provider
- For how many model versions has a commitment been published, and how often do the weights change between them?
Checkable in the world
- Has anyone outside a provider independently verified a proof for a model they did not build?
7. What This Does Not Show
- It does not show that provable inference is useless. It shows a narrow, well-scoped result that makes one specific claim checkable.
- It does not show that self-improving systems are coming, or what they would do to verification. Section 5 is a guess.
- It does not show how the proof would be used in practice. The use cases in Section 4 are our reading, not the post’s.
- We read several things only in part. We read the post and the repository’s description through page text. We did not read the technical report, run the code, or check the proofs. We read the Japanese report as supplied, without knowing its outlet. The figures are as the post and repository state them.
8. Conclusion
Provable inference makes one claim checkable: this output came from that model. For the narrow range it covers today, that is a useful thing to have. It also leaves several questions open: who vouches for the commitment, what the model does, and whether anyone unwilling to be checked is affected at all.
If the systems being verified begin to change faster than they can be examined, the proof might keep its precision and lose its place. We do not know that this is where things go. We think it is worth asking.
A proof is for those who agree to be checked.
We will look again on 31 December 2026.
Yoshimichi Kumon Organizer, LSI Inventor, ARDS/ARKS (PCT GA26P001WO) Visiting Researcher, Waseda BFC MIT Sloan + CSAIL
References
- Anthropic, “Progress on provable inference for large language models” (published 2 November 2026, modified 3 November 2026). https://www.anthropic.com/research/provable-inference
- anthropics/FlashProve, GitHub repository (research artifact; Apache 2.0). https://github.com/anthropics/FlashProve
- Technical report, “FlashProve: an open-source library for provable inference” (linked from the post; not read by us). https://assets.claude.com/8e1d181a83015078581c92d896757041713f0160.pdf
- Japanese news report on the work, 5 November 2026 (secondary; outlet name not shown in the text we received)
- LSI, “The Switch Within Reach.” https://logos-sovereign.space/?p=522




Ⅽomment