How the AI estimate is made
Three models assess the claim separately. The estimate helps you read the evidence; the market sets the price.
What the models read
Each model receives the same claim, resolution criteria and stored evidence packet. Models do not see market prices or one another’s answers.
Web search is enabled for each model’s request. Each can find additional sources separately; the recorded search count and citations show what was reported for that call.
The packet is stored by the site, not assembled by an autonomous research agent. Some entries describe evidence requirements rather than published papers. Read the source links and model responses on the market page.
When it updates
Rounds are paused; the weekly panel is not currently scheduled to run.
A round needs three valid estimates. Failed or disabled calls produce a labelled placeholder, not a new AI assessment.
What the estimate means
It is the models’ assessment of the evidence, not a calibrated probability or proof. Models can miss sources, misread results or share the same errors.
Each also returns YES, NO or UNDECIDED against the resolution criteria. Probabilities alone do not settle or pay tokens. Settlement follows each market’s frozen policy; markets without a terminal binding retain the proposal, challenge and finalization stages.
Terminal settlement and deployment
Three frozen trusted adapters and model identifiers must return matching explicit YES or NO verdicts against the evidence criteria, using the same frozen context and evidence packet. When they match, the reports are saved and a settlement transaction is signed and sent to Solana. The site shows the outcome as final only after Solana confirms the transaction as finalized and the market account on chain matches the market that settled.
Probability-only reports, UNDECIDED, disagreement, missing or invalid reports do not settle. Terminal outcomes have no appeal or reopening. Signatures record trusted adapter assertions, not proof of model authorship or scientific truth.
Getting paid: the site does not send money to anyone. When a market settles, the winning token becomes redeemable. Open the market's resolution status page and redeem it from your own wallet, and it pays 1 USDC less the market's disclosed redemption fee.
Automatic terminal settlement switches on market by market as each market's on-chain checks pass. Markets without a terminal binding yet settle under the policy they listed with; the market's resolution status page shows which process applies and where the market stands. Read each market’s resolution status before you buy.
Models, prompts and historical records
Current prompt: panel-v8. Latest recorded methodology: science-2026-10-15-audit2.
OpenAI · openai/gpt-6.1-sol
Anthropic · anthropic/claude-sonnet-5.5
xAI · x-ai/grok-4.7
Stored responses and model identifiers are available on market pages. Current examples do not reconstruct an older round’s exact request.
- Recorded version
- panel-v1
- Context rules
- Each round builds a packet from the market's frozen hypothesis, its fixed resolution criterion, and the evidence rows on record. The hypothesis is written once at listing and never edited; a reworded claim is a new market. The packet is the same for all three providers in a round, so the three scores differ only by the model, not by the inputs.
- Sampling rules
- One call per provider per round, three providers, run together under a wall-clock ceiling. A round is valid only when all three providers return a usable score in range. A partial round is not averaged: two surviving scores are never averaged into a published number, an outlier is never dropped, and a valid answer is never rerun for a different result. Each provider also returns a confident binary, YES, NO, or UNDECIDED. A market is decided only when all three return the same confident YES, or all three the same confident NO. Any disagreement between the two definite answers, any UNDECIDED, and any missing or unusable binary all mean not settled. The weekly mean is a current estimate; it does not settle a market and does not decide the binary.
- Evidence rules
- Evidence is fixed before exposure. The packet lists the supporting, contrary, method, and context artifacts that matter and what each would have to show; it carries no invented numbers, names, or citations. Where a provider can search the web, the exact model id that answered is recorded on the round. Evidence cutoffs, weighting, provenance, and the treatment of a retraction or a contradictory result are recorded per round rather than left implicit.
- Change history
- v1: initial published methodology. Superseded versions stay on the record; a version that has settled a round is never edited, and a change is a new version with a new effective date.
- Governance
- The panel is three model providers: three different AI companies supply the models. One operator runs the process around them — it selects the evidence packet, calls the models, and publishes the round. Because the operator sits between the providers and the public, the models' agreement is weaker corroboration than three independent firms would be. Who may change the methodology, resolve a dispute, or correct a result: the operator, and no one else. Changes are prospective: a completed settlement is never rewritten, and a change is a new version.