2026-08-27 - Asked to drop an engine, and the engine was fine
local-cpu scored zero and looked like dead weight. It is
the second-best engine on this machine - 4 of 12 novel
words against 2 for the four-billion-parameter model on the accelerator
above it. The zero was the scorer demanding a JSON array while the
proposal loop reads three fallbacks deep. A benchmark stricter than its
consumer measures the benchmark (D335).
measure now takes the caller's scoring function. The
strict one is a default, not the rule.
Quoting the reply is what found it. The failure line
now carries what was said, and it said
<think> </think> Here is a list of **12 new variant markers**...
- engine working, model on topic, and a reasoning block the managed path
suppresses with --reasoning off that the in-process path
cannot. None of that was visible behind "not a list of words".
Three measurements, three wrong answers, all mine: by speed, by volume, by strictness. The third nearly deleted something that works.