Circuit and proof structure within the transparent boundary, the schematic form verification only covers the clear object and the assumption

The cover is a concept illustration generated by AI and is not an actual device, system architecture or experimental results.

  • Interpretation date: 2026-10-24
  • Data deadline: 2026-10-24 20:35:34 (Beijing time)

On October 22, Anthropic and Theorem announced a study of AI-assisted formal verification. The team reported that Claude generated proofs of safety properties for 239 test entry points across 123 open-source libraries. The results include proof files that others can check. To understand what it means to “write a proof for software,” we need to look at the precise proposition each file proves. This article examines the announcement from two days ago. research announcement

From trying a lot of input, to proving a proposition

Fuzzing tests will continuously generate input to find the case of program crashes; formal verification will write the requirements as mathematical propositions, and then check whether the derivation is true. The former is good at finding counter-examples, and the latter can make an overall demonstration of an input scope under the clear premise.

This work compiles the existing test entry into RISC-V binary, and then uses the machine and operating system model to define the nature to be proven, and Claude constructs proof and Rocq check. The scope of the announcement includes inputs less than 64 MB, with conditions such as single-threaded, static programs and fixed memory budgets. Methods and Scope

For scientific readers, the important change is that the evaluation object becomes specific: not only read the "this code should be safe" written by the model, but also check whether the proof of delivery corresponds to the specified program and specified requirements. The language persuasion of the model is therefore no longer the final referee. However, this division of labor still needs someone to make a serious decision: what is it to prove?

The scale of the announcement and the public sample must be read separately

The pinned version of the OSS Verify that this article checks is exposed 25 libraries, 60 entrances proof. This is a subset of the results of the announcement, and the large total number of research cannot be directly regarded as the number of materials that the public has already obtained; let alone the "number of entrances" can be written as "the number of fully verified libraries". Public Release Notes

There is also an easy-to-missing detail: the proof of this version is more common and added to runtime error detection builds, indicating that they do not have the defined errors within the specified scope, and have the same observable behavior. README clearly pointed out that these public certificates do not prove termination, so It is not ruled out that the program keeps running without ending. Propositions and Uncovered Content

This difference does not weaken the proof itself, but it will change the news headline. "The specified error is not triggered" "It will definitely end" and "The calculation result meets the user's needs" are different requirements. If they are merged into a sentence "the software will not have a problem", the reader will not know which level of conclusion the evidence actually supports.

How does the inspector avoid proving another thing?

Check the document description, the process will check the SHA-256 of the binary file, compile the certificate, and use the additional package definition to prove the type of the target proposition; then review the assumptions used, and reject the assumptions outside the allow list. These steps constrain the "what file was checked" and "what the proposition" and "what you believe extra". Proof check mechanism

The most valuable inspiration here is to separate generation and acceptance. If the model can modify the questions, answers and scoring methods at the same time, passing the inspection does not necessarily mean that the original problem is solved. Fixed objects and acceptance rules can leave failures for the reasons that can be positioned, and also let latecomers know where to re-check.

At the same time, the document takes additional rocqchk Check the column as optional steps and indicate that the default CI does not execute it. Seeing that "the machine has been checked", you should continue to ask which process is checked, and you cannot make up for "all inspection methods have passed". This article reads the instructions and does not actually run these checks. Default and optional checks

Mathematical derivation is correct, reality mapping still needs to be reviewed

The trusted base is the basic components and assumptions that the whole set of arguments still rely on. The dependencies listed in the project are not only the Rocq kernel, but also include machine semantics, file conversion tools, checking package logic and proposition text. It clearly does not guarantee the relationship between binary and source code; the proof is pointing to the byte specified by the hash. Trusted base description

This means that the reader needs to separate whether the logic is true and whether the model is close to the real machine. LPS is a formal model of the Linux process running RISC-V programs. The author clearly marks that it is still incomplete and records the scope and assumptions. The conclusion in a version of the model cannot be automatically transferred to another operating environment. Machine Semantics and Restrictions

An intuitive analogy is: it can strictly prove that the route on a map is connected, but it is still necessary to check whether the map is missing the broken bridge in reality. For software, the version, execution environment and usage method jointly determine the scope of application of this "map". Pointing out that the boundary is not a proof of formal form, but makes its guarantee can be used correctly.

Next, what results are more weighty?

This article also expects to see three follow-up signals: whether the external inspection can be reworked stably, whether the assumption of the semantic model can continue to be audited, and how much investment is required to maintain the certificate after the code update. These issues are related to checkability, practical applicability, and continuous use costs, respectively.

Comparison of scale also requires a complete denominator. In the follow-up study, if you can simultaneously report how many entrances you have tried, which failures, which are limited by the scope of the model, and manual intervention and total calculation consumption, it is easier to judge where the progress comes from. This is the evaluation method suggested in this paper, not claiming that the team has completed such a control experiment.

The available material makes AI-assisted construction of software proofs a workflow worth studying systematically. The public sample is too limited to support a security guarantee across the software ecosystem. This account draws on a research announcement and the authors’ technical materials. It reports neither peer review nor multiple independent replications. I have not rerun the proofs or audited the implementation. The question to follow is how the scope of checkable guarantees can grow while their limits remain explicit.

References and reading range

  1. Anthropic:LLMs are making formal software verification feasible at a much larger scale, 2026-10-22. Methods, scales, limitations and related footnotes of the announcements that have been read; the complete appendix of its link has not been reviewed, and the additional experimental conclusions in the appendix are not used. The announcement does not give a specific release time zone, and the experiment is only described as a continuous period of weeks.
  2. Theorem: OSS Verify README (fixed version). The public quantity, proposition description, uncovered content and version information have been checked; the pinned commit is 8f5f467, this round has not independently confirmed its submission time.
  3. Theorem:How a proof is checked. Read check steps, hypothesis review, optional checks and CI descriptions, no commands are executed.
  4. Theorem:What you have to trust. Read full text to define trusted substrates and binary objects.
  5. Theorem:LPS: the Linux process semantics. Key reading status, design principles and version scope; no independent verification form model. These repository materials and announcements are evidence of the same cooperative project.

Apps by the author

The following is the author's own product promotion. In addition to writing a technical blog, I have also independently developed two apps to help organize daily life. Welcome to understand:

  • Daily Expense Journal: Record daily income and expenditure, arrange budget and savings goals, and make each account and plan more clear. View on the App Store.
  • One Small Action: Choose small actions according to the current time and energy, record the feelings after completion, and leave a clue to the daily review. View on the App Store.

Find more articles and projects on my website: zffjzz.top. Your feedback helps me improve these tools.