Built for prediction markets, tokenized assets,
trading, and autonomous agents.
Markets, smart contracts, and autonomous agents act on what happens in the real world — tokenized stocks and treasuries, insured events, agent transactions — but reality arrives as fragmented news, conflicting sources, and rules written in natural language. Today’s oracles were built to move data, not to judge it.
Prices, randomness, structured API answers. Solved, and still the floor everything else stands on.
LiveHuman voting is slow and contestable. Black-box AI is fast but unaccountable.
Every source, reasoning step, and conclusion is audited, challenged, and anchored on-chain.
Proven first in prediction markets, extending to trading intelligence, tokenized assets (stocks, treasuries, collectibles), parametric insurance, and agent-to-agent commerce.
What happened?
Evidence-backed settlement for prediction markets and on-chain contracts.
See productWhat's likely to happen?
An agent skill that returns a probability for an event, with the claim and sources behind it.
See productWhat's the number?
Drop-in feeds for majors, judgment-based pricing for everything else.
See productWhat should happen next?
Oracles that monitor, reason, and act within policy.
See productFour products on one reasoning engine. Pick one to see what it does, the numbers, and how it works.
Settle real-world outcomes with evidence, not votes.
Reads a market's own rules, gathers evidence from multiple sources, and returns a verdict with its full reasoning trace and an on-chain proof. Every case is public in Explorer. Official resolution layer for 42, and live with Myriad.
Running in live prediction market workflows, where resolution accuracy, latency, and economic impact are measured in production.
*all data are up to date as of October 7, 2026.
Proof of Reasoning makes every judgment verifiable. The Verification Chain makes it enforceable, without a single vote. And none of it requires the evidence to be public.
Three proofs, each produced by a separately staked agent, then a layer that enforces them. Authenticity says the evidence is real, logic says the reasoning holds, determinism says the answer is settleable — and a bonded challenger can prove any of the three wrong. Verification replaces voting: faults are proven, not debated, and only the role at fault pays.
Read the whitepaper ↗The Verification Chain ↗Full diagram ↗Every piece of evidence carries a cryptographic receipt of the source it came from, and of the moment it was read.
Source integritySearch, verification and deduction are committed as a trace. A dispute re-runs only the contested step.
Logical coherenceNeuro-symbolic constraints force one settleable answer. Two conflicting signed outputs are provable by comparison.
Schema complianceAnyone can bond and re-run a claim. One honest challenger is enough, and only the role at fault is slashed.
Challenge bondA fault slashes only the role that caused it. A token never decides what is true.
Every oracle in production assumes the data is public. That assumption rules out most of what institutions actually need verified — bank balances, covenants, credit files, shipment records. Cournot splits the pipeline in two: public questions stay trustless through opML replay, private ones move into attested enclaves from several hardware vendors, and both end in the same auditable reasoning trace.
Privacy architecture ↗Talk to us about a private deployment ↗Three layers, no overlap. Each proves something the other two structurally cannot.
Verification is trustless: anyone can re-run it.
Replay is impossible without exposing the data, so trust is hardened by hardware diversity — and the Anchor stakes more for it.
A probability for anything.
Cournot Intelligence is an agent skill. Install it once in the agent runtime you already use, ask with /cournot, and get back one number: the probability of an event, the exact claim it was computed against, and the sources that informed it. If a request cannot be resolved to a specific claim, the skill stops and asks instead of guessing. Agent platforms such as Alphio use it to assess market events before their agents act.
Agents install it once and call it on demand. The first release covers crypto events.
Claude Code
Codex
*Live since August 24, 2026. Data as of September 20, 2026.
Cournot collects the probabilities implied for the same event across independent sources, weights them into one estimate, and returns it with the claim and the basis.
Example session shown on skill.cournot.ai · rules v1.0 · An assessment of probability, not investment advice.
From BTC/USD to a graded Charizard.
Standard pairs work out of the box. Cournot goes further: tokenized stocks and treasuries, graded collectibles, and long-tail RWAs, where a price requires judgment — source tiering, wash-trade filtering, and a confidence interval on every number. The same engine provides verifiable randomness for draws, drops, and fair distribution.
VRF is mathematically deterministic, so the blockchain can leave the critical path and act only as the verification layer.
*Figures from the Cournot Oracle VRF design. Comparison latencies reflect typical multi-block confirmation waits.
A price nobody publishes, and a random number nobody can rig.
A long-tail asset has no single price. Cournot reads every venue that trades it, ranks them by how they settle, drops the trades that are not real, and publishes a number with its spread.
Illustrative long-tail asset. Majors use drop-in feeds; judgment pricing is for what has no feed.
From judging reality to acting on it.
The next step for Cournot: autonomous agents that continuously monitor conditions, reason over new evidence, and trigger pre-approved actions within defined policies, all with the same verifiable trail. It extends Cournot from prediction markets into insurance triggers, credit intelligence, compliance monitoring, and agent-to-agent commerce.
Verticals where Cournot is extending the same engine from judgment to policy-bounded action.
For each sector we publish our thesis — with production data where we are live, and our reasoning where we are building.
Resolution, not liquidity, is what limits how many markets can exist.
Traders and agents need a sourced probability, not another opinion.
Illiquid assets have no price feed, fragmented sales and grading need a judge, and the packs they come in need draws anyone can verify.
Before an agent pays another agent, someone has to verify the work was done.
A payout trigger is only as honest as the sensor behind it — and often the trigger data is private.
Creditworthiness lives in documents nobody can publish. A verdict can still be public.
Cournot settles verified outcomes on production chains.
Evidence is gathered from news, markets, sports, and public social data.
Common questions about Cournot, its products, and how to integrate.
npx cournot-skills add Cournot-AI/cournot-skills/skills/cournot, then ask with /cournot followed by your question. The skill runs only on an explicit request. Each account starts with free calls; after that, calls are paid per call via x402 / b402 on Base or BNB Chain. Payment is handled by a local client, and wallet credentials never enter the model's context. Agent platforms such as Alphio already use it to assess market events before acting.