Inference programs on Solana: three draft SIMDs for discussion

[Verafyai]

Our intention is to propose, and have a discussion around, the potential of a new class of "smart contracts" which we are loosely calling inference programs, or "Actually Smart Contracts," and whether the technical and security building blocks realistically exist to justify or support them. The key motivator is Anthropic's recent result on zero-knowledge proofs of LLM inference, and we wanted to test whether it could mean a new class of programs on Solana. We have no strong opinion on whether this is technically feasible, but we see interesting potential if it is.

The drafts were written with help from AI tools, so please flag anything that's poorly reasoned, unclear or longer than it needs to be.

What an inference program is. A program written in natural language and stored on chain: instructions, plus the rules that shape them, such as a DAO constitution, a policy, a standard or a GitHub repo at a commit. Another program calls it with a question and the evidence. AI models from different labs run it off chain, each signs a vote, and the caller gets back a signed ruling (outcome, reasoning, sources, signatures) that it checks before it acts.

What it could be used for. Checking a passed proposal against a DAO constitution before it executes. Guarding treasury or exchange withdrawals with a written policy. Certifying a code commit against a published standard. Forecasting and resolving prediction markets. Enforcing written limits on what an AI agent may do.

Why three drafts. The parts can be judged separately, and we'd rather they fail separately:

1. Inference Programs. How the prompt and its rules are stored and versioned, and how a program requests a ruling and pays for it.
2. Evaluators and Attestations. Who runs the models and what a vote proves. Proofs of inference are far too large for the chain today (about 18.78 MB each), so this proposes bonds and challenges instead.
3. Rulings, Disputes and Challenges. The signed result programs act on, and how it can be challenged with new evidence and overturned later.

What we want to settle. Whether the building blocks are ready, whether a shared standard is the right vehicle, and what should change before any of this becomes a PR.

How it might benefit Solana, if it works. Programs could enforce rules that code alone can't express, with every decision, its reasoning and its sources on a public record. Wallets and explorers could show what any ruling relied on. And model operators would have one way to be paid for verifiable work on Solana. No validator or runtime change is proposed.

The full drafts, motivation and slides: Bringing Inference Programs To Solana. Microsite: asc.verafy.ai.

We'd most like to hear whether this belongs in a standard at all, and what failure modes we've missed. Blunt feedback is welcome.