Tidel.ai

Data for models that
imagine what happens next.

Training data for world models — gameplay video paired with the ground truth behind it: 4D geometry, tick-level world state, and human-written visual reasoning.

The four product lines below are live, browsable samples packaged in the format we ship in production. Together they cover the loop a world model has to learn: watch the world move (4D game capture), read the state that produced every frame (engine-native logs), reason about what it sees (chain-of-thought visual and STEM data), and build the objects its worlds are made of (parametric 3D assets). Nothing is estimated by an off-the-shelf perception model — depth, pose, state and events are read straight from the engine, and reasoning chains are written by people.

A toy game world dissolving into its simulation view — wireframe, depth and telemetry
World-Model Data · Catalog 4 feeds · video → state → reasoning → assets · access code tidelaccess
Feed 1 · See

4D game capture with real telemetry

Gameplay video from a proprietary capture pipeline over AAA titles, paired with the ground truth a world model actually needs — measured from the engine, not estimated by a perception model.

Feed 2 · Read

Engine-native world state

Every video frame ships with the exact world state the simulation computed at that instant — entities, components, events. Nothing is detected; everything is read. One 60 Hz clock across all modalities.

Feed 3 · Reason

Visual reasoning with human chains of thought

The seeing skills frontier models still fail — on the BabyVision benchmark (arXiv:2601.06521) adults score 94.1% while the best model scores 49.7% — and the reasoning-first data that trains them.

Feed 4 · Build

Parametric 3D objects for the worlds themselves

Simulation-ready objects generated by build scripts: change a dimension, rerun, and the mesh, UVs, PBR textures and renders all regenerate consistently. Real-world scale, named movable parts, and the editable Blender source with every piece.