Mint a market
MenuClose

Products

Kernel

About

Mint a market

Workflow-based resolution beats AI quorums.

  • More reliable
  • Harder to manipulate
  • Reproducible
  • Faster
Same market, two resolvers · illustrative

“Will US CPI be at least 3.0% in the October report?”

AI-judge quorum

3 models vote
  1. 01

    Search the open web

    Headline: “Inflation hits 3%”

  2. 02

    Every judge reads the same page

    • A · YES
    • B · YES
    • C · YES
  3. 03

    YES

    Unanimous. And wrong: the headline was rounded.

Fermah Kernel

rule fixed at mint
  1. 01

    Named source, locked in

    BLS · CPI-U · all items · not seasonally adjusted

  2. 02

    Read the number

    12-month change = 2.96%

  3. 03

    NO

    2.96 ≥ 3.00 is false. Same answer on every replay.

More judges ≠ more truth

Models make the same mistakes.

  • Same training data, same search results
  • So their errors line up instead of cancelling
  • Past three judges, accuracy barely moves
0%1%2%3%4%5%13579AI judges in the quorumreal, correlated judgesfloor no quorum can get belowif judges were independent0%1%2%3%4%5%13579AI judges in the quorumreal, correlated judgesfloor no quorum can get belowif judges were independent

Model: each judge wrong 5% of the time · illustrative

AI-judge quorum compared with Fermah Kernel
AI-judge quorumFermah Kernel
ClarityDecided at the deadlineFixed before the first trade
ReliabilityWhatever search findsNamed sources, frozen at mint
ReproducibilityCan flipSame answer, inputs stored
Handling uncertaintyPicks an answer anywayPauses, extends or voids
Manipulation resistancePlant pages it might readMust compromise the named source
SpeedMinutesSeconds

Where judgment is needed, AI works inside the rule.

kernel workflow · fixed at mint
ai.judge
  • Cited evidence only
  • 2+ model families
  • 2/3 to settle
  • Unsure → review, never a guess

Inside a Kernel workflow, the AI judge is one step, wrapped by four guards.

What the research says about AI judges