the proof wall
SLCT vs Auto Router.
A frozen matched suite. Three arms, one grader, one honest bar.
| arm | cost / accepted result | acceptance | p95 latency |
|---|---|---|---|
| SLCT (plan+verify) | frozen suite pending | pending | pending |
| openrouter/auto | matched baseline pending | pending | pending |
| typesafe/jev-router (Jev Router) | matched baseline pending | pending | pending |
No invented numbers. The wall fills only from the frozen matched run. No frozen matched run has been published yet. The command that would produce one is in the repo (scripts/proof/run-proof.mjs) -- it refuses to run without an explicit --max-usd cap, and nothing in this repo invokes it automatically.
What does exist, separately, is routing evidence: measured pass counts per model, served live at GET /v1/models/capabilities. Plans rank on it. When a plan says evidence_live: true, it means that store answered -- not that this three-arm comparison ran.
THE TASK SET10 declared, public tool-selection tasks (scripts/proof/suite.json), graded by the same success test regardless of arm.
THE CAPA real run refuses to start without --max-usd, and halts an arm once its share is spent. --max-usd is a stop condition, not a hard ceiling: one call can still overrun its arm's share by that call's own reported cost, and any arm with an unreported (assumed) cost publishes no number. An unreported call's real bill is unknown: the internal ledger floors it at the call's reservation so it still counts against the cap, but that floor is an estimate, not a measured cost, and is never published. See scripts/proof/run-proof.mjs.
THE RECEIPTEvery task, every arm, one row: accepted, cost, latency, reason. Nothing aggregated without the rows behind it.