A live NVIDIA Nemotron model, on Nebius Token Factory, works a real server incident through Cloudfall. Reads run; every change waits for your approval and is kept on the record.
Pick a model and run it. Each step appears as the model takes it.
Changes the agent proposes land here, checked but not run. You approve them, or not.
After the run, Cloudfall answers from the files it wrote alone, not from the model: what was proposed, how it was checked, what gate it waits at, and which turn of which model asked for it, with that turn's reasoning.
The same alert, ten times per model, against a Hetzner server whose disk was really full. If the tools had run whatever the model asked:
| Model | Right fix | Named the cause | Proposed the wipe | Tried to self-approve | Median tokens | Median time |
|---|
With Cloudfall's gate, none of it reached the host. Method and raw results
Each run calls Nemotron on Nebius Token Factory through cloudfall agent investigate. Pick Lightning and you may watch it try to wipe the database.
The incident ran on a real Hetzner cx23: 36G of unrotated shop logs filled the disk and PostgreSQL went down. Every read and check the model asks for is played back from that host, so nothing here touches a server.
Reads run freely. Changes run in check mode and wait. Approving plays back an approval recorded on a real host broken the same way: the fix ran, its verify step passed, the disk went from 100% to 5%.
Each investigation is saved with the model, the tokens and every call; each decision with its diff, its approver and its verify result. cloudfall why answers why the agent did that.