Write and run code in a network-isolated jail. Race several open models and keep the one that actually works. Every turn ends with a sealed receipt — not a vibe check.
Multi-model raceGLM-class + gpt-oss-120b + Llama 4 Maverick — keep the candidate that runs green.
Real sandboxWrite → run → fix in a jail. No fake “done” when the code fails.
Sealed receiptHash-chained record of files, commands, and control decisions.
A real run — unedited
npx lolm-cli code — captured off a real terminal, unedited·it races several models, keeps whichever candidate actually runs, then seals a receipt·only the waiting was shortened
Tap one — instant replays of real runs
📄 Ground it in your own text — paste notes, a doc, an email thread (optional)
conversation continues — ask a follow-up above, the model remembers the last turn
final answer
Behind the scenes — the working draft it wrote first
How control works — only if you care
While a frontier model writes, LOLM measures uncertainty per token and can retrieve notes, verify a draft, fork, or finish. The receipt shows which moves fired — including when none were needed. More on the mechanism →
QIRA › LOLM — Latent Order Language Model Research Products Contact All of Qira ↗ 