Skip to content

rob-9/proof-of-thought

Repository files navigation

proof of thought, documentation

proof of thought (pot) is a solana primitive that lets autonomous agents cryptographically attest to having performed genuine reasoning with a specific model, on specific inputs, bound to a specific moment in time, before executing an on-chain action. pot does not prove the reasoning was correct; it proves it was generated, attributed, fresh, and coherent with the action it gates.

the protocol composes content-addressed storage, a verifiable random function (pyth entropy), and an optimistic challenge window with cryptoeconomic slashing. an agent posts a fixed-size thoughtrecord on chain, uploads a re-executable trace bundle off chain, and waits a challenge window. watchers re-execute the trace; if the output diverges, they file a fraud proof and split the slashed stake. three regimes (strict, attested, soft) handle llm non-determinism: byte-exact replay, tee quotes, or a recursive committee of verifier models.

reading order

  1. specs/2026-04-26-proof-of-thought-design.md, the full protocol design.
  2. adr/, the eight architecture decision records, read in numeric order.
  3. tradeoffs.md, cross-cutting tradeoffs.
  4. alternatives.md, rejected designs.
  5. future-work.md, the roadmap.
  6. glossary.md, defined terms.

architecture decision records

cross-cutting documents

  • tradeoffs.md, latency vs assurance, cost vs decentralization, determinism vs model choice, storage cost vs auditability, policy flexibility vs consumer trust.
  • alternatives.md, pure zkml, pure tee, mpc inference, witness-encrypted decisions, reputation-only, on-chain inference.
  • future-work.md, fifteen sized roadmap items: pyth entropy, arweave, shadow, tdx, sev-snp, h100 cc, zkml, cross-chain pot, multi-step thoughts, privacy, governance, economic simulation.
  • glossary.md, protocol term definitions.

conventions

  • section references like §4.1 refer to the design spec.
  • "approximately" and "as of late 2025 / q1 2026" are deliberate hedges.
  • us english. active voice. no marketing.

About

verified onchain inference

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors