Method
How one host is told from another
One model is proven every six hours, models taken in turn, so nothing on the ledger is older than fourteen days. You never run any of it yourself.
The question set
Every host answers the same six questions, twenty-four times over, in shifting wording: an animal, a colour, that colour in Spanish and in Chinese, a first name, a number. Ask a model the same thing twenty-four times and it does not answer evenly. It plays favourites.
How hosts are separated
Hosts carrying the same weights line up. One serving a different model, a different version, or a distilled copy of it breaks the pattern. Each host is scored against the reference set, so a mismatch lands as a number and not an opinion.
The meter check
Where we hold the model's tokenizer, every reply is counted and matched against the count the host billed. A host that bills for more than it served surfaces right here.
The limits
A proof run is evidence, not a warranty. A host that answers cleanly today can change tonight, so old results stop counting and everything is proven again inside fourteen days. Cheap hosting can also be slow, so look at latency before you switch.
Reading a mark
Three marks, no grey areas
Every host on the ledger carries one of these for the reference answers of the model it lists.
On pattern
Answers match the reference set. Requests flow here.
Unsettled
Too close to call. Proven again before it counts.
Off pattern
Departs from the reference. Never served to you.
Prices
Where a published price comes from
A published price is the host's own listing, for a model that held up under proof.
- Listings belong to the hosts and are re-read every few hours.
- Compared as three parts input to one part output, per million tokens.
- Mid means the middle of every price listed for that model.
- A host that billed above its own listing during a proof run never counts as the lowest.
- Compressed versions (fp8, fp4, int4) compete on the same terms and face the same questions.
- Not everything is on the ledger yet. Anything unlisted still flows, marked unproven.
Fees
What serving costs
Zero on your first $25 of AI spend each month, then 0.9%. Pro and Max wipe the fee much further up the bill, and everything settles in our token.