mfornet

There is a lot to unpack in your comment, thanks for commenting.

We are heavily using LLMs and agents for writing and verifying all the code. We have some safeguard inplace, such as not breaking the wasm testsuite, and being able to run wasm code and produce the correct results (even if that is not the goal of this interpreter). There is an ongoing effort (not by us) about creating a formal WASM spec in lean, generated from SpecTec, when that lands our plan is to prove that our interpreter follow the specification.

> E.g. are you able to do an "in-place Store" like WasmRef-Isabelle, and can you represent memories and tables as plain vectors of bytes/refs in memory, can you grow them in-place, etc.? Or any other optimizations/lessons learned?

Even if the interpreter can be use to run wasm code, we don't really care about efficiency, it is not intended to run WASM, but instead to be able to verify code, so we are intentionally not building an optimized interpreter, or rather we are optimizing toward ease of demonstration.

> How are you modeling the explicit sources of nondeterminism in the Wasm execution semantics? E.g. NaN representation, {memory., table.}grow, host calls, stack exhaustion, relaxed SIMD instructions, etc.,

Going over your list, some functions that might fail, like {memory., table.}grow and stack exhaustion can't fail on our interpreter, so that transition is not represented here. We need to revisit this hypothesis, but part of the reasoning is that we all properties we prove about the WASM bytecode, are properties held by the original code, and usually (for example in lean) you can prove that a function has some property for all values of a natural number, even if you can't really run the function for all values due to some kind of "system" failure (stack overflow).

NaN, we certainly need to revisit floats, we are delegaing its behaviour to Lean floats. Again part of why we think this might work is because we are not planning to run code with this interpreter, but just to write proofs.

Host calls is properly modeled and being developed as we speak, since we really care about being able to proofs properties of a code that runs on some particular host. When you write a statement about some code, you do it on the presence of some particular host. I'm exploring the possibility of formalizing NEAR smart contracts, which are WASM binaries.

AI has been great so far filling in most of the proofs, and I'm trying to avoid SMT-style proofs early on, to make sure we have a solid API that can be scaled to arbitrary complex code without increasing too much the cost of the verification. I'm sure nonetheless that SMT solvers will play a role going forward in filling up some proofs.

The reason AI has been so good at filling most of the proofs in my opinion, is that proofs are actually "simple", but very tedious and long. Part of our work right now is make sure that the tedious part can be solved mechanically as easy and efficient as possible, so both human or AI can focus on the interesting parts.

> Doesn't that put the Rust compiler (and its assert lowering) in the trusted base?

Yes, but I would argue the are already in the trusted base before this project, we are not removing that. We want instead remove "your code" from the trusted base, and just keep the compiler and the specs.

> How do you know the asserts you wrote are the traps you're reasoning about?

You just do. The asserts and the specs have a similar role, they are both consumer facing, and consumers need to make sure they are correct and cover what it is intended.

> Is the plan to build a new separation logic framework, or use e.g. iris-lean or splean as a base?

We plan to build our own, though we are currently evaluating what is the current state of iris-lean. Thanks for pointing to splean, I wasn't aware of it before.

> I’d imagine you’d still want step indexing to allow reasoning around cyclic heap structures

We have some APIs around loops and function calls that hides the "fuel-detail" and instead asks you to provide a well-founded relation that is used to prove that the behavior of the program is correct for "some" fuel.

The GCD example make use of this primitive to provide the invariants of the loop plus the argument that loop will terminate.

> Fun project in any case! I look forward to seeing how it develops :)

Feel free to join the telegram dev chat (link in github)

Both.

You can write "annotate" your rust code using asserts. On the wasm side asserts are converted to trap instructions, so the Lean spec will simply be: For every input this code never traps.

Part of our focus is making sure that specs are both easy to write and read, since they are human facing. Eventually you could imagine how writing code will mostly be writing specs, and both the code and the proofs will be handled by AI agents. In this scenario it is very important that humans can easily audit and modify the specs.

Initially we considered formalizing rust code, aeneas is a very promising project that would unlock a lot of features right way by transpiling to lean. However, we didn't want to lock ourselves to rust, so we decided to use a lower level target such that we could verify code from "any" language.

We considered LLVM-IR, and RISC-V.

Ultimately WASM felt like the right decisions. More importantly WASM spec is very well done in details, and it is written with formal verification in mind early on, and there are are plans from to include Lean as one of the targets for generating the spec automatically from SpecTec. Once this exist, we will formalize that our interpreter is correct under the definition generated from the official Wasm-Lean-Spec so we remove it from the Trusted-Base going forward.

> what if I want to reason about larger programs that dynamically allocate, where the addresses may not be known statically? How can I make sure these do not overlap?

We are actively working on this, as it is a pre-condition :P to reason about the simplest of useful programs. The idea is to develop an API around separation logic that allows you to reason about logic that manipulate non-overlapping regions of memory.

It won't be relevant if address are not known statically since API theorems will be parametrized over non-relevant constants such as addresses, function indices, etc...

> And since this is a shallow embedding into lean, what’s the approach for verifying properties of non-terminating programs?

To use the interpreter there is the concept of fuel, which we explicitly hide from the reasoning layer. Using fuel you can write statements of the form, this function returns out of fuel for any value of fuel passed to the interpreter, which is equivalent to prove that your program doesn't terminate.

As I see it, this model will be able to predict “easy” to derive tokens but will no chance on “hard” tokens.

For example doing a sum of random numbers. If the token you are trying to predict is not in the training data, even if similar patterns exist, this model defaults to the Neural Model.

I guess then it is an aide to the neural model on filling the easy patterns.