SELAT // Capability re-rankReplay of the recorded eval // 0 intents // live catalog incl. x402 Bazaar // Oct 2026

Which endpoint can actually do it?

SELAT's keyword ranker finds endpoints on the right topic. Jev, TypeSafe's System One model, then asks of each shortlisted endpoint whether it can perform the action the intent asks for. A gate swaps SELAT's pick only when Jev is confident and SELAT's pick looks incapable.

Intents judged0

Candidates scored0

SELAT alone0

With Jev gate0

Regressions0

Jev spend$0.0000

Swap when Jev's top ≥ 0.60 and SELAT's pick ≤ 0.40

Intents

SELAT right, kept Jev fixed it regression still wrong no capable endpoint shortlisted

Teardown

IntentSELAT shortlistJev scoresGatePick

Bars: Jev's P(capable). │ swap threshold for Jev's top │ ceiling for SELAT's pick. LABEL = keyword-rule ground truth.

How well Jev separates capable from not

Gate settings

top-1 /0 // regressions

Click a cell to apply it. Outlined cells hit the best score with zero regressions.

Where the gain comes from

■ SELAT alone ■ with gate

Method

Each intent ran through selat search --top 8, then Jev scored every candidate (one batched call, or one call per candidate). Jev sees the service name, method, templated path, a short description and input names. Never URLs, prices or wallet data.

Ground truth

"Capable" comes from keyword rules, not human review, so a few labels are wrong. Read the counts as a comparison between SELAT alone and SELAT with the Jev gate on the same labels, not as absolute accuracy.

What Jev can't fix

On 0 intents no shortlisted endpoint could do the job (places near a landmark, UV index, pollen, a person's work email…). Those are retrieval misses. The gate made 0 swaps on them; always-on Jev made 0.