HomeSandboxGalleryWorkbenchbrowsing as guestSign in to run

GPS-Bench

A policy is put to the actors who would actually answer it — governments, frontier labs, chipmakers — each carrying its own recorded conduct up to 2024. They state a position alone, then again after hearing the others, and what changes between those two answers is the part worth reading.

Where to go

Sandbox → Describe a policy that does not exist yet and watch it land. Actors are drawn from the recorded roster, answer independently, then answer again having seen the room; you can edit any actor’s position and re-run to see who moves in response. Runs on a GPU, and takes a few seconds. Gallery → Every simulation that has been run here, kept and browsable: the policy as it was put, each actor’s stance and the action it would take, who approached whom, and which approaches were reciprocated into a coalition. Reads from disk and never touches the GPU. Workbench → How the actors were fitted and what the comparisons showed: the eight fitting modules and the evidence behind each, what every actor decides under each of them, the coalition structure, and the findings — including the ones that went against expectation. A static record of a completed analysis.

What is being run

Qwen2.5-7B-Instruct on local GPUs, with per-actor LoRA fittings built from pre-2024 evidence only. Sandbox runs currently use the untouched base weights; the fitted variants exist and are loaded on the machine, but are not yet selectable from the page.

Nothing is sent to a hosted model and no API key is involved. Greedy decoding, so a repeated run of the same policy returns the same answer: a result is a property of the weights rather than a lucky sample.