Before0.88 s
After0.16 s
Time to the first token of an answer, one test turn. Bars are to scale.
File 001 · 2026-10-09 · running
Why our AI was slow to answer, found in one night.
Our engine's median time to first token was 1.27 s (397 requests, all traffic). Most of it traced to cache-block rounding, an estimated 0.8-1.5 s per turn. A config change then cut a test turn from 0.88 s to 0.16 s.
- How it was measured
- Live engine metrics on our own machine, 2026-10-09; the rounding share is an estimate from the uncached-token distribution, not a direct measurement. People turns were not measured on this engine.
- What would prove it wrong
- If an exact end-of-request cache checkpoint does not remove most of that re-read time on byte-stable turns, the claim is wrong.
- Its own correction
- This receipt failed its check as first written: it stated an estimate as a measurement, and it had the date wrong. An agent that didn't write it re-read the source numbers and wrote the sentence above.
File 002 · sealed
File 003 · sealed
A sealed file is registered before its run and not opened yet.