laya-v6-bahn

Laya (421M parameters, ModernBERT-large), fine-tuned to check an agent's decision card for known traps before a person sees it: a stale GPS fix, a replaced instruction, a wrong escalation. Text only.

Built for Controlane, a control room for car delivery, at the Prompt Operations Hackathon (24 to 25 Oct 2026). Live at https://controlane.bobs.build.

Results

On 4,395 synthetic test cases from 203 template families held out of training completely:

Checker	Catch rate	False alarms	Time per check
Laya v6 (this model)	86%	8%	10 ms
Laya, as released	60%	54%	10 ms
Jev (typesafe/jev-1.14)	92%	13%	229 ms

31 of 34 on the hand-labelled set. For higher accuracy, see bahnator-1.1-pro-preview.

Training

The recipe of Laya's official notebook (RLCD loss, temperature calibration) on one RTX 6000 Ada on Verda. Version 6 of 6; synthetic train split only, no test case used.

Use

import laya  # laya 0.4.15
agent = laya.Agent("path/to/this/repo", device="cuda")
agent.predict(state, questions)

All training data is synthetic and fictional.