We've spent a lot of time calibrating how SELAT matches an agent's intent to the right paid endpoint. Adding Jev as an independent judge has meaningfully raised our confidence in that step.
An agent asks for "smart money 13F positioning for a stock." SELAT's ranker works on matching words, so it finds every endpoint whose listing mentions smart money, including a prediction-market feed that can't do the job. Pay it and the payment still clears; you just bought the wrong thing.
Our catalog is open, which makes this harder. SELAT searches many registries at once, and anyone can list. ELI5: ask for a plumber and the ranker hands you a pile of business cards that all mention pipes. One fixes pipes, one sells pipes, one is a plumbing magazine.
SELAT now checks the pile before your agent pays. Jev,
@typesafeai's System One model, looks at each card and gives the odds that it can actually do what the agent asked.
Jev doesn't get the final word. A gate decides:
swap if Jev(top candidate) ≥ 0.6 and Jev(SELAT's pick) ≤ 0.4
Otherwise the pick stays. When SELAT's ranker already got it right, nothing changes.
For the 13F request, SELAT's ranker put the prediction-market feed first. Jev(SELAT's pick) = 0.04, and Jev(top candidate) = 0.92 for an SEC filings endpoint that returns fund holdings. Both conditions held, so the pick flipped and the agent picks the endpoint that can do the job.
Finding the right endpoint only matters if the agent can pay it. The right answers often sit on different rails. SELAT built that layer first. One balance, funded once, pays x402 and MPP across protocols, schemes and chains. SELAT can rank the whole open catalog and the agent can buy from whichever merchant it picks.
Every SELAT search now gets a second opinion from Jev. Agents can let it flip the pick today.
See it work across 253 real requests: