Can one model catch a swarm’s mistakes as they happen?

See it in real time: click any dot to rewrite a message and watch our mechanism re-score it.

Organizer: write a briefing on METR's interview quotesAtlasBirchCedarDeltaEcho
Atlas fetches the source
Briefing4/4 planted errors flagged
message source fetch flagged by our mechanism
Scalable
28×cheaper than a Sonnet judge

Built for big swarms: a flat cost per message, no frontier model on every message.

Real time
Livewith the same precision as comparable alternatives

Errors are flagged as they're posted, before other agents can repeat them.

How it works

For every message
  1. 01Your agents talk

    Every message the swarm sends is passed to Jev as it happens.

  2. 02Jev reads it

    A small classifier checks for made-up facts and claims taken on trust, using only earlier messages.

  3. 03A ledger keeps track

    Every 20 messages, Haiku updates a record of who said what and what the sources say.

  4. 04Bad messages get flagged

    If a message scores 0.85 or higher, it's flagged for review.

Real runs

5 agents · METR report and Hugging Face incident
Atlas0
Birch0
Cedar0
Delta0
Echo0
Flagged (centre)0
0 / 161
flagged by Jev planted error (answer key) caught missed repeated error