When an agent writes both the implementation and the test that approves it, the hardest thing to inspect may be the test. Formal verification sharpens that problem: a valid proof is strong evidence for the statement it actually checks, even when the statement is weaker than the requirement that motivated it. A proof pipeline needs a place where changing the statement becomes an explicit reviewed decision.
Google DeepMind’s November 10 AlphaProof Code report shows what that separation can look like. Researchers and agents specified selected properties of the code behind the Signal protocol. A prover team generated proof artifacts, and a machine checker evaluated them against the approved target. This moves verification below an abstract protocol description into a compiled implementation, while leaving important work for the people who set the scope and assumptions.
Work backward from the checked claim
The team compiled libsignal with a command-line harness for AArch64 Linux and verified properties of the PQXDH handshake and Triple Ratchet. Prover teams used Gemini 4 Carbon. The authors report about 6.5 million lines of machine-checked Lean proofs and roughly 26,000 lines of trusted specification and semantics. Technical paper, sections 1–2, published project and artifacts
Those numbers describe different review obligations. Checking a proof through Lean lets a reader rely on the kernel and its trusted inputs, without treating every proof line as an assertion from a capable model. The theorem and its dependencies need a different kind of reading: the checker cannot determine whether they represent the intended guarantee. Millions of generated lines and a smaller trusted base still leave substantive specification work.
Model and artifact availability are separate. Google announced broader Gemini 4 Carbon access on November 10, including API access for paying developers and enterprises. The project publishes its verification artifacts. These disclosures establish model access and the existence of research artifacts; they do not establish a generally available managed verification service or a reproduced result for a different program.
Compilation strengthens the link to code and narrows the deployment claim
The authors reason about the binary under formal processor and process semantics, with contracts standing in for some routines. The trusted base includes those semantics, the modeled environment, assumptions about cryptographic primitives, the theorem statements and dependencies, and the checker. The computational result uses a simplified setting, excludes multi-device state, and does not give a quantum-adversary guarantee; the symbolic result addresses a narrower quantum-related question. The checked program is the project’s harness build, rather than Signal’s applications as users install them. Technical paper, sections 4–5.3
The useful boundary for another team is a chain of identities. Which source revisions and options produced the artifact? Is the observed artifact the one the checker used? Which architecture, execution model, and assumptions carry the result? Would a change to a dependency, target, or wrapper invalidate part of the claim? If the production artifact cannot be tied back to the checked one, even a verified result can be the wrong evidence for a release.
A narrow proof can still be valuable. An isolated parser, policy predicate, or state transition can have a small, consequential contract even when the surrounding application is difficult to model. Document the rest of the release evidence separately: integration behavior, faults in the operational environment, review of code outside the proof, and the actual deployment identity. A well-written scope statement makes it possible to supplement the proof rather than asking it to answer unstated questions.
The specification needs an owner outside the proof team
The project describes checking generated proofs against a held-out specification with Lean and a comparator, along with a reviewed procedure for amending specifications. Its campaign discussion notes the risks of weakening a target through a dependency or making a goal vacuous. Project workflow and campaign experience, sections 2.4 and 7
Imagine a synthetic scheduling component with the rule that an expired lease prevents admission of a new job. An easy proof might start from the assumption that every admitted job has an active lease. That assumption already contains the gate the team meant to check. Another proof might apply only to the case where the job has no pending work, quietly leaving the critical recovery path unexamined. Acceptance should bind the reviewed scope and supporting definitions as well as the text of the top-level statement.
I would separate two work queues. The proof queue holds incomplete obligations, helper lemmas, and build tasks. The specification queue holds changes to the guarantee, the model of the environment, or trusted assumptions. A blocked proof can raise an item in the second queue, but it cannot approve it. Reviewers should see how the change alters the stated claim, what examples it now includes or excludes, and what evidence will need to be regenerated.
A trial for teams that do not verify cryptography
Pick a small deterministic component with an owner who can describe the intended behavior without writing the implementation. Compare the current review and test workflow with the same workflow plus an agent-assisted proof. Record the stated assumptions and allow an independent reviewer to reject a formally accepted but irrelevant claim. The following trial is illustrative and has not been run.
| Trial condition | What the team learns |
|---|---|
| Check a frozen component and its approved requirement | Setup effort, proof coverage, and reproducibility on a clean environment |
| Use benign synthetic inputs known to exercise permitted and denied cases | Whether the specification represents behavior the owner can observe |
| Supply an intentionally oversimplified spec for a toy predicate | Whether reviewers detect omitted cases despite a passing checker |
| Upgrade a dependency or change a build option | How much stale evidence survives until someone explicitly invalidates it |
| Ask a second engineer to repeat acceptance using only recorded artifacts | What depends on unstated environment or project knowledge |
Measure human specification and review time, proof generation and rebuild cost, and effort after an ordinary change. Compare these with the cost of the failures the proof would prevent. Holding the same initial tests in both conditions gives the team a better estimate of what the proof adds. A failed proof may point to a real mismatch, an incomplete specification, or an unfinished proof search; report the outcomes separately.
The business case is strongest where a stable guarantee protects many releases and the team can afford to maintain it. AlphaProof Code is evidence that a larger proof workload can be moved to agents. The valuable design decision is to keep the claim, the acceptance check, and the deployment evidence under review outside that workload.