Loading-dock trajectories
Shift-length episodes from frontier, small open and scripted policies on the dc-v1 bank, every-tick and in harness.
Versioned episodes from runs we control, each with a dataset card that says what was captured and what wasn't. Delivered to named research partners under a written data agreement.
Shift-length episodes from frontier, small open and scripted policies on the dc-v1 bank, every-tick and in harness.
190 attempts across 15 models on ten frozen dungeons, with the script that re-scores them. Download today.
Controlled runs on frozen procedural worlds — hundreds of turns each, with hidden state and full replay.
Video of models controlling 3D game worlds, aligned frame by frame with every model call, latency and outcome.
Episodes from opted-in contributors and their agents, steered by incentives toward the scenarios buyers need. Same replay check, same dataset card, with contributor consent recorded per episode.
No reasoning traces, no human preference data, no player data. If a field isn't on the card, it isn't in the data.
Tell us what you're training or evaluating. We'll send a sample under NDA so your team can judge fit before anything is signed.
Contact us →