Mentat vs Polymarket: cryptographic settlement vs corporate dispute desks

The honest framing first: Polymarket is the best prediction market product available today by any commercial metric. Pristine UX, deep liquidity on US politics, real-time price discovery, mobile app, native order book. If you want to trade tier-one event markets at scale today, the answer is Polymarket. We are not pretending otherwise.

We are also not pretending Polymarket is the right architecture for the next ten years of prediction markets. This post is the careful technical comparison.

What Polymarket actually is

Polymarket runs on Polygon PoS, settles outcomes via UMA’s optimistic oracle, and is operated by Polymarket Inc. (and, post-2024, a regulated subsidiary structure). Markets are staff-listed: users submit market ideas, Polymarket’s team approves and deploys them. Trading is permissionless globally (modulo jurisdictional restrictions) but creation is not.

Resolution is layered:

1. UMA optimistic oracle. A proposer submits a result. A challenge window opens. If unchallenged, the result is canonical. If challenged, the dispute escalates to UMA’s DVM token-weighted vote.
2. Polymarket Inc. discretion. Markets that Polymarket considers ambiguous can be invalidated by the company. This is the corporate dispute desk: a final backstop for cases where UMA voters reach an answer the company disagrees with or the market spec is judged unresolvable.

This architecture has produced excellent product outcomes. It has also produced repeated public disputes during high-profile resolution events — particularly across the 2024 US election cycle — where the trigger condition was contestable enough that UMA voters split and Polymarket had to make a judgment call.

What zkTLS settlement changes

Mentat’s resolution path is built around the bet that most market triggers can be mechanized down to a deterministic check against a canonical data source — and that the contestable cases should be flagged and rewritten by the AI pipeline before the market ships.

The settlement layer is the MarketSettlement program on Solana. It accepts a zkTLS proof artifact and verifies on-chain that:

• The proof corresponds to a TLS 1.3 session against a domain in the market spec’s primary_source_list.
• The proof’s transcript commitment matches a trigger_condition (JSONPath query or phrase match) declared in the spec.
• The TLS session timestamp falls within the spec’s acceptance window.

If those checks pass, the market resolves and payout executes. No vote, no challenge window for the common case, no corporate desk. The chain verified the source directly.

The architectural inversion: Polymarket trusts UMA voters; Mentat trusts that TLS 1.3 and the certificate authority chain do what they are designed to do. Those are different trust assumptions and they have different failure modes.

Where Polymarket clearly wins today

Liquidity. Polymarket has millions of dollars of depth on flagship markets. Mentat is pre-mainnet. There is no contest on this dimension today. We expect liquidity to follow the right architecture, but the architecture has to ship first.

Brand and discovery. Polymarket is the default association with “prediction market” in mainstream conversation. That is a real asset.

Mobile UX. Polymarket’s mobile app is excellent. Mentat M3 ships a web trading interface; mobile is M5.

Specific market categories. Polymarket has invested heavily in US political market quality. The depth of liquidity and the editorial team’s domain knowledge produce better markets in that category today than a permissionless protocol would.

These are not architectural advantages but commercial ones. Architectural advantages compound; commercial leads decay if the architecture is wrong.

Where Mentat is structurally different

Permissionless market creation. Anyone can deploy a market through the AI authoring pipeline, subject to curator review. Polymarket caps creation behind a staff queue.

Cryptographic settlement. zkTLS proofs verified on-chain replace UMA voting for the common case. Latency drops from hours-to-days on disputes to seconds on proof verification.

Open source stack. The whole protocol — Vue frontend, FastAPI backend, DSPy agents, Solana programs, Rust indexer, proof service — is Apache-2.0. Anyone can fork, white-label, or self-host.

Solana fee economics. Polygon PoS fees are low but not negligible; Solana fees are sub-cent and confirmation is sub-second. This matters for high-frequency UI, for the long tail of small markets, and for trading strategies that don’t survive higher fee floors.

Schema standardization. Every market conforms to a canonical schema with explicit trigger conditions, source allowlists, fallback logic, and invalidation clauses. Indexers and integrations all speak one format. Polymarket markets are well-curated but spec-on-spec inconsistencies exist because authoring is human-by-human.

Where the comparison is genuinely close

Dispute handling on edge cases. Polymarket’s UMA + corporate desk hybrid does handle genuine ambiguity. Mentat sends ambiguous outcomes to a dispute lane — challenger bond, governance review — but governance design is M5+. We are not claiming we have solved every edge case; we are claiming we have removed most edges through better authoring.

Regulatory perimeter. Polymarket has navigated CFTC engagement and (as of writing) is operating under a settlement framework that gives US users limited access. Mentat is permissionless at the protocol layer; jurisdictional compliance is a deployment concern. Whether that ends up being a better or worse posture depends on enforcement evolution.

A concrete settlement walkthrough

Take a market: “Will Bitcoin’s spot price as reported by api.coingecko.com/api/v3/simple/price?ids=bitcoin&vs_currencies=usd exceed $150,000 USD at any point between 2027-01-01T00:00:00Z and 2027-06-30T23:59:59Z?”

On Polymarket today, this market would:

1. Be drafted, reviewed by Polymarket staff, and deployed.
2. Trade until expiry.
3. At resolution, a UMA proposer would submit YES or NO.
4. A challenge window opens; anyone with a bond can dispute.
5. If undisputed, result is canonical. If disputed, UMA DVM vote decides.
6. Polymarket may override if it judges the result invalid (e.g., if CoinGecko had a documented outage at the relevant timestamp).

On Mentat:

1. The creator types: “BTC above $150k at any point in H1 2027 per CoinGecko.”
2. The Scout agent surfaces the concept; the Draft agent emits structured JSON with the API endpoint, JSONPath query ($.bitcoin.usd), trigger threshold, acceptance window, and fallback logic.
3. The Validator agent confirms the trigger is deterministic and the source is in the allowlist.
4. A curator reviews the spec, accepts it, and deploys to MarketFactory.
5. The market trades.
6. When the threshold is crossed, an off-chain proof service generates a zkTLS proof attesting that api.coingecko.com returned a price above $150,000 at a timestamp inside the window.
7. The proof is submitted to MarketSettlement. The program verifies the proof, verifies the trigger condition, and resolves the market YES. Payouts execute.

No vote. No dispute window for the common case (one opens for the edge case where the proof is allegedly forged or the trigger is misinterpreted). Resolution latency: seconds after the threshold crosses.

What the comparison cashes out to

If you are a trader today, Polymarket. If you are building a prediction market product, Mentat — because the architecture you start from defines what you can ship in two years, and zkTLS settlement is a better architecture than UMA + corporate desk for the markets the long tail of the space wants.

We are not the only people noticing this. Several teams across crypto are converging on zkTLS-style settlement for various data delivery problems. Mentat is the bet that prediction markets specifically benefit most because resolution is their hardest problem and “trust the source, not the network” is their best available answer.

The history of crypto infrastructure is the history of replacing trust layers with cryptographic ones at the moment the primitives become cheap enough. Settlement is the last big remaining one in prediction markets. zkTLS is what cheap-enough looks like.