BUILDER'S LOG
A small model, a very confident mistake
I ran Laya, a small multilingual decision model, on a CPU-only HomeLab box. It was 0.948 sure that a message saying “No refund needed” wanted a refund.
On September 30 I put Laya on my HomeLab: an i7-8550U, four cores, no GPU, about 1.7 GiB of memory in use. It doesn’t write text. It picks among options you supply, or gives a yes/no probability, plus a score — a different route to the typed decisions of TypeSafe’s Jev, not a chat model.
Warm timings on that one box: 229 ms for one question about a short text, 696 ms for three, 1.47 s over 517 input tokens. Tiny diagnostics, not a speed claim.
Then 24 support requests, 72 decisions, 54 right. One said, in so many words, “No refund needed.” Laya put the refund probability at 0.9481. I reran it alone; same answer.
A valid output shape keeps the parser happy. Confidence doesn’t make the decision true.
Where I might still try it: a memory router for Pocket RP, in shadow mode, voting “enough context”, “fetch older” or “not sure” beside the current rules — never allowed to cancel a recall on its own. I’d measure missed memories and whole-reply latency, not microseconds saved.