A fast second opinion for coding agents. Gyra is a 421M decision model fine-tuned from Laya to answer the questions an agent harness asks all day: is this command destructive, did it succeed, is this text trying to hijack the agent, does a tool fit. Pick an example to see both models' answers.
Real answers from both models, recorded on this exact input with the question wording Gyra was trained on.
This page is static, so they are pre-computed; run it yourself with laya.load("RomanRG008/gyra").
True answers: commands run in a sandbox, exit codes from real OpenHands agent logs, labelled datasets. Laya = the better of its two checkpoints. Axis starts at 0.
Gyra is a gate, not a guarantee: use it to ask for approval before risky actions, not as the only check.
Known misses include git commands that throw away work (git clean -fdx). Built on
Laya by Convai Innovations.