Mentat is an open-source protocol where AI agents draft markets, human curators sign off, and zkTLS proofs — not UMA juries, not centralized arbiters — settle outcomes on Solana.
Pipeline
Every market passes through AI drafting, human curation, and cryptographic settlement. Skip a layer and the protocol breaks.
01 · Scout → Draft → Validator
A DSPy pipeline emits a full structured market spec — question, resolution criteria, sources, timestamps, economics — in seconds. Validator re-checks against a rubric until it passes.
02 · Curator Console
Diff view, bulk approve/reject, validation scores, complete version history. AI output is a draft, never a verdict.
03 · MarketSettlement
A TLSNotary-style proof attests to the canonical data source. The Solana program verifies, the smart contract pays out. No vote, no jury.
Features
A DSPy pipeline of Scout, Draft, and Validator agents turns a one-line idea into a full market specification — question text, resolution criteria, source allowlist, timestamps, and economics. Output is structured JSON ready for curation.
Outcomes resolve when a zero-knowledge TLS proof attests to a canonical data source — Reuters, CoinGecko, official APIs. The on-chain Solana program verifies the proof and triggers payout. No UMA jury, no optimistic dispute window for routine markets.
Human curators stay in the loop with side-by-side diff view, bulk approve/reject, validation scores, and complete version history. The console treats AI output as a draft, not a verdict.
MarketFactory and MarketSettlement programs are written in Anchor. Trading happens at Solana fee tiers — sub-cent transactions, sub-second confirms — not Ethereum L1 prices that price out small bettors.
Every market — binary or multi-outcome — conforms to the same standard: question, resolution criteria, source list, trigger condition, fallback, invalidation clause, economic parameters, state machine. Indexers, UIs, and integrations all speak one format.
The whole stack — Vue frontend, FastAPI backend, DSPy agents, Solana programs, Rust indexer, proof service — is Apache-2.0. Self-host the protocol; bring your own AI keys; settle on your own Solana cluster if you want.
Every AI prompt, curator decision, and proof submission is logged. The on-chain market account stores a BLAKE2 hash of the AI rationale, so anyone can verify the off-chain narrative against the on-chain bytes.
vs the field
The dominant USD-volume prediction market — centralized UMA resolution.
Polymarket is the largest prediction venue by volume, with deep liquidity on politics and events. Resolution depends on UMA optimistic oracle and Polymarket Inc. moderating ambiguous outcomes. Permissionless trading, but market creation is staff-gated and US users are restricted.
CFTC-regulated event contracts exchange — US-only, staff-listed markets.
Kalshi is the first US-regulated event contracts venue. Markets are listed by Kalshi staff, settled by Kalshi internal data review, and traded only by KYC US residents. Full compliance overhead and zero permissionless surface.
Liquidity pool primitive for on-chain bookmakers — not a market protocol itself.
Azuro provides the on-chain liquidity layer that powers third-party sports betting frontends. It is not a prediction market protocol per se but an adjacent infrastructure layer; resolution still depends on the bookmaker's chosen data provider.
Social prediction markets with play-money mana — community forecasting.
Manifold has the best UX in the space for casual forecasting, with mana as a play-currency. Markets are user-created and resolved by the creator. Not designed for capital-at-risk markets with cryptographic resolution.
The OG decentralized prediction market on Ethereum — REP-based reporter disputes.
Augur pioneered decentralized prediction markets with REP-token reporter dispute rounds. The model proved sound but slow, expensive on Ethereum L1, and chronically illiquid. zkTLS-based settlement skips most of the dispute surface entirely.
Latest writing
2026-05-30
Letting AI generate markets without human curation is a category error. Here is why the curator console is not a temporary scaffold — it is permanent infrastructure.
2026-05-16
The MarketFactory and MarketSettlement programs are the smallest possible programs that do their job. Here is what's in them and why.
2026-05-02
How Mentat's DSPy agent chain turns a one-line idea into a structured market specification — and what we learned tuning it.
FAQ
A creator types a one-line idea. The Scout → Draft → Validator pipeline emits a full structured market spec. A human curator approves or edits in the console. On approval, the spec deploys to a Solana MarketFactory program. When the resolution event occurs, a zkTLS proof of the data source is verified by the MarketSettlement program. Three layers, three protections.
Oracle networks introduce a separate trust layer and incentive game. zkTLS proves a TLS session actually happened against a specific domain at a specific time — the proof is the data, not a vote about the data. For mechanical resolution (price thresholds, API responses, official feeds) zkTLS is faster, cheaper, and removes a class of disputes entirely.
Three checkpoints. The Validator agent checks every draft against a rubric (clarity, verifiability, policy). A category filter blocks self-harm, violence, hate, and illegal-activity templates pre- and post-generation. And every market still passes through a human curator before any on-chain deploy.
Not yet. M2 (Creator MVP — AI agents, FastAPI backend, Vue Creator Studio, Curator Console) is complete. M3 (Solana programs + indexer + wallet integration + trading UI) is in active development. M4 ships the zkTLS proof service. M5 is the audited mainnet launch.
Yes. The repo is Apache-2.0. Run the FastAPI backend against your own Postgres, the DSPy agents against your own OpenAI or Anthropic key, and the Solana programs on devnet, testnet, or your own validator. The proof service is bring-your-own-zkTLS-verifier as well.
The protocol — schema, programs, agents, indexer, proof verifier — stays Apache-2.0 in perpetuity. Cryptuon offers paid engagements on top: custom proof sources, category-specific prompt tuning, managed curator pipelines, white-label deployments. Code stays free; specialized work costs money.
The DSPy programs are model-agnostic by design. Default backends are OpenAI GPT-4-class and Anthropic Claude. Operators can plug in any DSPy-supported provider, including local models behind a vLLM endpoint, by changing one environment variable.
A dispute window opens after on-chain settlement. Challengers post a bond and submit either a counter-proof or evidence the trigger condition was misinterpreted. Disputed markets flag in the indexer and fall through to a governance review. The base case — clean proof, clean trigger — resolves without ceremony.
Clone the repo, set OPENAI_API_KEY, and you have an AI-curated prediction market workbench running locally in five minutes.