/health and /knowledge/status.The knowledge engine's unit cost is a subprocess spawn, not GPU time. Measured
on this box: a fresh (uncached) input costs about 1.1 s cold / 0.5 s warm
of one vCPU; a cached input is a disk (or S3) read — effectively free. The store turns a
per-call cost into a per-unique-input cost, and glass order text has a
heavy head (the same 4/16/4 builds recur constantly).
computing…
| line | rate | note |
|---|---|---|
r6i.4xlarge on-demand | ≈ $1.18 / hr | 128 GB — headroom the OOM incident bought |
| ≈ monthly always-on | ≈ $850 / mo | see the incident log in CLAUDE.md |
| knowledge fresh inference | ≈ 0.5–1.1 s vCPU | subprocess spawn, once per unique input |
| knowledge cached serve | ≈ 0 (disk/S3 read) | the amortized common case |
Turning ensemble on by default means every /query now also
consults the knowledge subprocess. That is the deliberate trade the vote buys: classic
single-cascade is ~33 ms; the ensemble is ~0.5 s warm —
roughly 15× slower on the hot path, bought back on repeats by
the store. The standing optimization is in-process knowledge loading to collapse the spawn.
This page will keep saying so until that lands.