Rights-cleared, structured, frame-aligned. The full (state, action, next-state) triplet across a growing catalogue of licensed titles, every one normalised to a single standard. Browse it, sample it, train on it.
Captured from the running game with PRISM. No screen scraping, no inference.Real gameplay with the captured signals playing in sync. Scrub to any frame and read the exact state, input, and buffers behind it.
Real capture session: switch render buffers on the video, scrub to any frame and read the exact state behind it — all from the session's VTX file.
Most game data is reconstructed from pixels. Ours is read directly from the running game, so every signal is exact, not estimated. The result is simulation-grade: exact enough to rebuild the moment, not just label it.
No state, no labels. What scraping gives you.
Lossy and indirect. Estimated state, approximate labels.
Direct from the running game. The (state, action, next-state) triplet, rights-cleared.
+ all Tier 1 & 2 data[ Placeholder pull-quote from a frontier-model researcher on why ground-truth game data beats scraped video. ]Frontier-model researcher, to be cleared ⚑
The behaviour in this data is human: hesitation, intent, mistakes, recovery. Agent play has a ceiling, scripted or learned: a policy generating its own training data can only teach what it already knows. The distribution narrows, and the model learns the agent, not the world. We capture real play for coverage, to a brief, and QA it before delivery.
Natural play across skill levels. The rare and the routine, not just the highlight reel.
Specify titles, scenarios, signals, and rates. We capture to order.
Checked for sync, completeness, and label quality.
Synthetic capture doesn't escape licensing either: data generated from a game is still derived from that game's IP. The rights question follows the title, not the capture method. Ours arrives already answered. ⚑
All of it synchronised to the same clock as the rendered frame. None of it can be inferred from video alone.
HUD-less, up to 4K and 60 Hz. The rendered world, not screen capture.
Separate tracks: dialogue, environment, effects. Not a single mixed-down stream.
Entities, transforms, velocity, physics, collisions, health, animations, events, objectives. Structured, not inferred.
Raw keyboard, mouse, controller. Exact sequences and timing. The action half of every learning pair.
Depth, surface normals, segmentation, motion vectors, UI / HUD masks.
Objectives, transitions, rewards, scene captions, named actions, semantic labels.
The render buffers come straight from the engine's pipeline, frame-aligned to the video and state, down to the G-buffer. The geometry behind the pixels, not estimated from them. ⚑
Every title is normalised to a single coordinate system, scale, and convention. Engine quirks removed. Data from one game runs unchanged alongside another, so you can train and mix across the whole catalogue with zero per-title wrangling.
Every clip carries structured metadata across around 20 categories, plus semantic labels spanning a large activity vocabulary. Filter and search across any combination. ⚑
Every record is licensed at the source and traces back to it. Train without the legal exposure of scraped footage or grey-area self-capture.
Captured only from titles we have licensed, non-exclusively, with the rights to train.
Every session's manifest carries its source, licence, and rights scope. Full chain of custody.
Access controlled, delivery versioned, usage reportable.