Show HN: Salt – a systems language with Z3 theorem proving in the compiler | Hacker News Reader

Show HN: Salt – a systems language with Z3 theorem proving in the compiler

salt-lang.dev

44 pointsbneb-dev54 comments

This looks pretty impressive but it’s all AI-generated (or written in a similar style) and therefore the documentation is lacking.

There is a language specification [1][2] but it lacks coherence.

I think the way to improve it would be to try to teach this language to people and get feedback from them. That is, it needs beta testers. It looks like there’s no community of users yet?

[1] https://github.com/bneb/lattice/blob/main/docs/SPEC.md

[2] https://github.com/bneb/lattice/blob/main/SYNTAX.md

For complete transparency: AI augmented engineering is the core workflow for this project.

I have been pretty diligent about trying to de-slop the project after long RALPH loops and `/goal` prompts, and I review and edit documentation. Based on your feedback, I just made another pass.

Please feel free to let me know if there is anything specific lacking from the docs, and I will update them in the future.

If you compare say, the Go language specification for an assignment statement [1] to Salt's explanation of its let statement, there is a big difference. Salt's docs lean heavily on "it works like you would expect from other languages" so you have to reason by analogy. A specification shouldn't work that way; it should actually explain the rules for each language construct, without assuming knowledge of other languages.

[1] https://go.dev/ref/spec#Assignment_statements

Thanks, I restructured that doc and will work on it moving forward.

For people looking for other languages with statically checked contracts, you might want to check out SPARK, which has been around in some form since the late 1980s. It is a subset of the Ada language and had been used for safety critical code in aerospace and defense projects, as well as for some Nvidia firmware.

It also uses Z3 to discharge proof obligations generated by the contract annotations, and it lets you use swap out different theorem provers as backends.

The GNAT Ada compiler (which is part of GCC) allows you to turn off the dynamic safety checks that are usually inserted into Ada programs at build time so you can optionally remove them if they are proven unnecessary.

Here are some resources for comparison:

- https://www.adacore.com/languages/spark

- https://learn.adacore.com/courses/intro-to-spark/chapters/01...

SPARK seems interesting. Any ideas how it compares to Salt?

- C performance? - Generics? - Syntax ergonomics?

Thanks for sharing!

Yes, it has generics. The syntax is Pascal-like, which some people used to C-family languages dislike but I personally find nice. Performance can definitely match performant C code. (intended use cases include real-time systems and embedded devices)

The language overview can be found here: https://docs.adacore.com/spark2014-docs/html/ug/en/spark_201...

It is a way to formally verify Ada, it meets all those criterias.

It looks like verification is distinct from the compiler.

> [int overflows, etc.] No runtime cost when Z3 can prove it. Otherwise, the compiler emits a safe runtime check as fallback.

Super interesting approach. I see this eventually be integrated into future mainstream languages, though that may take a while. I suspect that the game programming crowd will try to use it first, due to the possibility to prove certain edge cases at compile time and skip the runtime cost. But perhaps this optimization drive is no longer the case because we've got bazillions of cores nowadays. I may be too old for these predictions. Cool nonetheless.

Thanks for reading and sharing your thoughts!

You should also look at Liquid Haskell.

LH seems cool but the syntax would definitely take some getting used to...

> I suspect that the game programming crowd will try to use it first

> perhaps this optimization drive is no longer the case because we've got bazillions of cores nowadays

gaming is one of the industries where developers have to care about performance. you can see it with all the backlash on unreal engine, players just dont want slow games even if you ignore the fps chasing pcmr crowd. and getting to 60 or even 30 fps with complex game logic and high res graphics is not easy.

one of the ways you can see this is actually language choice. almost all engines are written in c++ (or sometimes rust). vm languages like java are rare because you dont want gc and extra allocations on the hot path. even for scripting theres a lot of focus on fast and low overhead interpreters, thats why lua is way more popular for games than python or js.

I was confused by the matrix multiply example since it doesn't show requires clauses. So I looked at what is presumably the source I'm supposed to look at: https://github.com/bneb/lattice/blob/main/basalt/src/kernels...

There's a comment saying every function in that file has requires clauses, but... They don't?

Also, the actual implementation of matrix multiply is manually tiled, while the website boasts about automatic loop tiling. It's hard to know what to trust here.

Looking. I did a huge purge/rewrite of the benchmarks and this might have regressed. Happy to track here or GitHub.

For here (and the website), I would be more interested in trustworthy information about what is actually verified at compile time, and what happens when verification fails.

What if you write a function that takes an array (as an array type if you have one, or as a pointer, or a slice, or whatever). It returns `array[0] + array[1]`, but it has no bounds checks in the code, neither as actual `if` statements nor as `requires` clauses.

Verification of such a function must fail at compile time, right? Does it? What does the error message look like? Someone else found some quote somewhere that seems to suggest that integer overflow checks are implemented as runtime checks if you can't eliminate them at compile time. Does that also happen for bounds checks? If yes, does the user get a warning saying that M out of N checks were moved to run time?

It's just really unclear if you are really doing "Verified safety without the runtime cost." or if you have reinvented Java, which will be just as good (actually better) at eliminating these bounds checks as compile time.

Bounds and div-by-zero are the easy wins; the real tax with contract systems is the loop invariant Z3 can't infer on its own, where you end up hand-writing an `ensures` longer than the function body. What's the gnarliest invariant you had to spell out by hand to get one of those kernels to verify?

For loops already worked but might not have been documented correctly. I just shipped the functionality for `while` loops now, auto-infer works for common cases of invariants:

  let mut i = 0;
  while i < 5 {
    arr[i] = val;
    i += 1;
  }

So in that case, the compiler detects `i < N` with monotonic increment and synthesizes `i >= 0 && i < N` automatically.

(Works with `<`, `<=`, `i += 1`, and `i = i + 1` via the same mechanism as for-loop induction variables.)

Complex loops (`while i + j < n`, `while p.addr() != 0`) require syntax that I have slopped together an explicit `invariant` keyword that I am not 100% happy with, but will refine later. Z3 checks base case + inductive step via Hoare logic and reports counterexamples on failure.

---

RE: "gnarliest invariant"

`ensures(result != 0)` on the SYN cookie in the kernel contract for the OS project in the monorepo.

The more interesting loop invariants aren't the ones about the loop counter but about the contents of the array. If you require `forall i, array[i] == value` after that loop, will the system come up with the loop invariant about the prefix of the array for any given `i`? What about something like bubblesort or insertion sort?

Agree with your overall point. I made some changes to extend the current paradigm a bit: https://github.com/bneb/lattice/blob/main/docs/tutorial/09-c...

---

Invariants aren't always inferred automatically. When they aren't, they can be made explicit for Z3 to check.

At the moment, bubble sort with array-content invariants is fully provable:

  fn bubble_sort(arr: Ptr<i32>, n: i64)
      requires n > 0
      ensures forall i in 0..(n-1) => arr[i] <= arr[i+1]
  {
      for i in 0..n {
          invariant forall k in 0..(i-1) => arr[k] <= arr[k+1];
          // bubble pass
      }
  }

Z3 proves the base case (vacuously true at i=0) and the inductive step for each iteration. Both loops have fixed trip counts regardless of data, so the frame axioms are concrete. At concrete sizes the compiler unrolls the loop: for i in 0..4 produces 8 Z3 checks (4 iterations × 2), all proven at compile time, outputting Z3: 8/8 checks proven (100%), 0 deferred to runtime.

And insertion sort is partially provable:

This starts with the same `forall` invariant structure. After the base case, the outer for-loop inductive step is wired but the inner while-loop's trip count depends on the data which can't be determined statically for which indices were modified, so the frame axioms can't fully constrain the array state. This would require splitting on the case for the while condition, which is future work, and wasn't part of my v1.0.0 plans.

Made updates directly influenced by your input. Please feel free to follow up in Github!

See https://github.com/bneb/lattice/blob/main/docs/deep-dives/z3...

Moving theorem proving upstream into compilers, sandboxes, and the browser seems like the future when we're dealing with increasingly sophisticated AIs. I'm working on similar formal methods but applied to agent sandboxing; do you see Z3 as a better fit than lean? https://github.com/coproduct-opensource/nucleus

Salt chose Z3 because it felt right for a compiler. The 100ms timeout means it's not sound, but it's useful. Lean could be the right choice when a proof is a hard requirement?

> Moving theorem proving upstream into compilers, sandboxes, and the browser seems like the future

The way I saw this proposed, back in the 2000s: "proof-carrying code."[0]

The idea is: the compiler compiles the program, and simultaneously generates a proof that the program doesn't violate X, Y, or Z safety rules.

Later, the end-user downloads the program and the proof together, and the local execution environment (the browser, the OS itself, etc) verifies that proof.

The idea being: constructing proofs is hard and sometimes involves manual steps. But verifying a proof is easy and automatic. So do the slow, manual thing once, and the fast thing each time the program is downloaded.

[0]: https://en.wikipedia.org/wiki/Proof-carrying_code

Sounds a bit like Microsoft's Singularity project, though I don't know if they even used a proof system to ensure a program couldn't do bad stuff(TM). After verification they just let the program run in Ring 0 (kernel space) of the CPU to skip even the performance hit of the cpu's isolation.

I thought Singularity was just building an OS on a memory-safe language and runtime, C# and .NET.

The same thing Java tried to do with applets at the browser level.

I listened to an interview with one of the researchers. They had a component to verify "this binary will never allocate mem outside it's allowed area" and by statically verifying that they could enable high performance IPC. Also, that was the source for the decision to allow it to run on Ring 0. That's what stuck in my mind.

Information is sparse about Singularity, which i understand is the same as Midori. Can you find that interview again?

I know about OpenGenera; what left-field OSes might i have missed?

Can be that I remembered it incorrectly, as it was in 2008, but here it is:

https://se-radio.net/2008/03/episode-88-the-singularity-rese...

Thanks, that was fascinating! I also see that Galen Hunt contributed to Windows Phone, which I adored. (Maybe I'm quixotic, maybe I'm a poor picker, but i love a thing that's beautiful and doomed.)

Why no GC? Guaranteed memory safety would be nice

It was a design decision. I chose Arenas instead.

The designed arena allocation is both faster and more predictable than GC. The trade-off is that you have to think about memory regions, but you never pay a GC pause.

Why not use Verus?

It augments Rust with Z3 and is not a pile of unverified slop.

Haven't heard of Venus (the programming language), but it sounds interesting. I'm curious to know what you mean by verified?

Do you happen to know if Venus has a Result type with canonical error codes, pipes, and other ergonomics?

Does it have C parity performance?

Verus, not Venus: https://verus-lang.github.io/verus/guide/

Ah, the issue is that I don't like that syntax: Salt is macro free, and doesn't need assert statements, etc... These are stylistic but I am assuming that under the hood things look quite similar.

I put an effort in to de-slop the project wherever I can. If you've noticed anything specific, I'd kindly request that you open an issue on Github or respond here with your findings so that I might correct them in the future.

Thanks for taking an in-depth look at Salt.

Thanks for trying to de-slop an ai-created project.

Not sure what tone you're going for, but can you expound on what you mean by 'ai-created'?

I'm not that commenter, but I feel like you wouldn't have to de-slop the project if it wasn't slop to begin with.

My workflow is to design, then use goal and loop prompts to RALPH a feature. At that point, usually there is a bunch of AI generated code that is generally correct but needs refinement, editing, testing, benchmarking, etc... That human-in-the-loop iterative step is the 'de-slop'. Would you prefer different phrasing, or is the trigger here the use of coding agents?

The biggest problem with obviously AI driven projects right now is that nobody knows how much expertise you have in this area (seems low given you don't know about systems that do very similar things), nor how carefully you thought about the problem and the implementation.

If I look at a project like Verus I know that experts thought about how to structure the system and it's concrete guarantees and semantics as well as the actual implementation.

So the result is that I trust Verus and I don't trust this.

You description of "design, then use goal and loop prompts to RALPH a feature" makes me even more convinced that this is slop because it sounds like you haven't thought deeply about every line of code and have delegated it to an AI that we all know makes mistakes. Thinking about the design is not really a substitute for thinking about the implementation.

People keep making AI OSS projects and expecting the same reaction as people had to OSS projects before AI, but pre-AI OSS came with a bunch of implicit promises about quality and effort and care that AI projects do not have.

I've interviewed several "I just read design docs and delegate all the actual thinking/analysis to the AI" engineers recently, and they are not grounded enough in reality to know when the AI is telling them something true or false and so confidently show me huge piles of code that purports to do things that I know are just nonsense. At some point I know these folks were good engineers, but they just drank too much of the kool aid.

To be clear, I have some of my own 100% slop projects where I have only vague ideas of what is going on beyond the high level design. But I have that currently put in a containment zone where I can easily verify the outputs or don't care about reliability (it's mostly UIs) and the scale of the damage they can do is minimal. Everything else, while basically still 100% AI generated at this point, is still reviewed and often repeatedly re-prompted to do precisely what I want the code to do, because if you let the coding agents make decisions for you, they still make bad decisions.

I'm not sure what the point of bringing up Versus is. Is it https://verus.io/ ? Is it a plug for your own project?

There are several points you made where I think we just fundamentally disagree so I'll decline to engage further at this point.

Hope you have better luck with your candidates and your own projects. I look forward do seeing them on HN down the road!

It's https://github.com/verus-lang/verus

It's not my own project, but it's doing basically the same thing you are doing which is why it is relevant.

With 92 contributors, I'd bet that The Vera Project is much more mature.

Salt has some different ergonomics and performance goals though, which makes this project unique.

So you're just going to gloss over all the rest of the very valid points he's making?

Salt is a systems programming language that embeds the Z3 SMT solver in the compiler.

You add `requires` and `ensures` clauses to functions, and the compiler proves them at compile time. When Z3 succeeds, the check is elided (zero instructions emitted).

When it fails, you get a counterexample. When it times out (100ms limit per obligation), the check is skipped and counted.

It compiles through MLIR to LLVM and targets KeuOS, a microkernel with an ECS (Entity Component System) architecture. Both are MIT-licensed.

---

How it works

Call `safe_div(x, 7)` and Z3 proves `7 != 0`. Check elided.

Call `safe_div(x, 0)` and the compiler stops.

The key difference from Rust/Zig/C: the compiler calls Z3 during normal compilation. No separate verification tool, no annotation language, no proof assistant. The contract syntax is part of the language.

---

What's real

- Compiler: 1,752 unit tests passing, clippy clean. Compiles through MLIR to LLVM IR. x86-64 and ARM64 backends. - Kernel: 14/14 QEMU e2e tests pass. TCP stack (connect/send/recv/close), ICMP, deterministic builds. NetD (network daemon) runs as a Ring 3 process on SPSC shared memory rings.

- ECS architecture: 13 entity syscalls (402-413). Entity lifecycle (spawn/exit/wait), memory regions as entities (map/brk/alloc), I/O routing via capabilities, socket entity tracking, performance counters, world persistence diagnostics.

- Shell: Inline `ecs`, `ps_ecs`, `free_ecs` commands query ECS World without spawning child processes.

- Benchmarks: Salt vs C (`clang -O3`) on 21 algorithm benchmarks. Salt at parity or faster on 19/21. Allocation-heavy workloads (hashmap, LRU, buffered writer) see 2-10x wins from arena allocation. Compute-bound (matmul, sieve, fib) at 0.9-1.0x of C.

- LSP: VS Code extension ships with semantic tokens, go-to-def, find-refs, Z3 hover.

---

What's not done (research-quality, not production)

- The standard library is incomplete. Many things you'd expect are missing.

- Z3 handles integer arithmetic, bit-vectors, and reals. String and quantifier support is partial. Contracts outside Z3's reach are compile-time checked where possible, silently skipped otherwise.

- Error messages from the Z3 pass can be opaque.

- The kernel targets QEMU (x86-64). Tested on AWS bare metal instances, not local 'bare metal' yet.

- One nights-and-weekends developer.

---

Why this exists

The goal was to find out whether formal verification could be a compiler feature rather than a separate toolchain. The benchmarks say the compiler is fast enough (Lettuce compiles in under a second with contracts enabled). The kernel contracts catch real bugs. But the language hasn't been used by anyone outside the project, and that's the test that matters.

---

Links

- Source: https://github.com/bneb/lattice)

- Tutorial: https://github.com/bneb/lattice/blob/main/docs/tutorial/your...

- Architecture: https://github.com/bneb/lattice/blob/main/docs/ARCH.md

- Benchmarks: https://github.com/bneb/lattice/blob/main/benchmarks/BENCHMA...

>One nights-and-weekends developer.

What?

This is being made by one person as a side project (not done during core working hours)

Why is that a bullet point under "What's not done"? It's confusing in that context, almost as if an AI wrote it and then it wasn't looked over.

You could nitpick the subsection title, and from that perspective "What's not done" is a bit of a catch-all for caveats.

With that said, I think it's pretty easy to infer the meaning there.

I'm not really sure how else to be more transparent about things, but I am using an AI-augmented engineering as the core workflow for the project, and that includes drafting docs and this post.

I read and edit the work. For a side-project with no current community support or users, I need to make a judicious decision when to spend time and effort. IMO it is reasonable _not_ to try to trick anyone into thinking that AI is not used. My proof reading and editing will miss things from time-to-time. I'm OK with it at the current scale of the project.

Next time curate what Claude/Codex tells you, this text seems targeted towards you, not a potential user

100%

What gives you that impression?

Reply on news.ycombinator.com