AI · Data for Models

Every frame, fully labelled.

PRISM captures live game state and player inputs directly from runtime. Structured, schema-consistent, frame-accurate. 10+ supported titles, more on the way.

Get in Touch Explore Data Tiers
Enemy
Terrorist · pos 1247, −823, 42
conf 0.94 · cls bot_T_03
Texture
bombsite_letter_b
bombsite_left_arrow
VFX
weapon_fire · t+0.012s
lifecycle 32ms · ally_03
Health
HP 34 / 100 · Δ −66
last hit t+0.21s · armor 100
CS2 frame captured by PRISM
Event
CT_03 → T_02 · headshot
t+12.4s · weapon MP9
Weapon
M4A1-S · silencer eq.
skin default · ammo 20/40
Round
Warmup · phase 0 / 5
match.id 1f4a · MR12
REC | vtxcapture_8412.vtx · tc00:00:08:412 · rate60 Hz · buildcs2 23038025

For marketing purposes only. The real data is much better.

Capture Data Tiers Models Runtime Assist
What PRISM Captures

The full
ground truth.

PRISM reads the running game directly. Output is structured and frame-aligned, captured at user-defined rates up to 120 Hz, so you only take what your model needs.

Render

Rendering

Visual ground truth straight from the engine. Pixels and the render buffers behind them.

Video Depth Segmentation Surface normals Motion vectors UI mask
Input

Player Inputs

What the human did, synchronised with the state stream. The action half of any imitation-learning pair.

Mouse Keyboard Controller Semantic keybindings Player viewport
State

World State

Everything the game is doing, frame by frame. The structured world behind the screen, not reconstructed from pixels.

Entities Transforms Health & status Animations NPCs Audio Match state Camera
Enrich

Data Enrichment

Added labels generated after capture. Semantic, inferred, and consistent across every supported title.

Scene captions Event annotations Intent & trajectory Named actions Inferred inputs
+ Video

Pairs with video. PRISM frames are timecoded. Bring your own gameplay video or let PRISM capture alongside, then sync against the state stream. The structured world is the ground-truth layer your video sits on top of.

0 Game SDKs. No studio integration required.

Data Tiers

Three tiers.
One pipeline.

Each tier targets a specific class of model. Tiers are cumulative, each including everything in the tier below.
Capture rates are configurable per stream and captures land in VTX.

Tier 1

Behavioural

For training models to act like human players. The minimum viable dataset for policy learning.

Player state
Player inputs
Camera data
Position & orientation
Lens parameters
Cross-engine coordinates
For Player Models · Action/Imitation · Assist
Enquire
Tier 2

World Model

For models that need to understand environment dynamics, not just player behaviour.

World & match state
NPCs & game entities
Multiplayer context
Audio events
Named player actions
Depth & UI buffers
+ All Tier 1 Data
For World Models · Reward/Eval · Player
Enquire
Tier 3

Simulation Grade

For full episode reconstruction, generative world models, and synthetic data generation.

Full granular entity state
Vision buffers
Segmentation masks
Surface normals
Motion vectors
Bounding boxes
Collision geometry
+ Enrichment Layer

Derivative labels generated after capture: scene captions, event annotations, intent and trajectory.

+ All Tier 1 & 2 Data
For Complex World Models · Synthetic Data · Simulation
Enquire
Models You Can Train

The best tool for training your models.

PRISM lands directly into the data shapes modern game-AI work depends on.
Six categories where it materially changes what you can build.

World World Models

Dense (state, action, next_state) tuples at frame granularity. Native, not reconstructed from video.

Action Action / Imitation

Synchronised state and player-input streams from real human play. Behavioural cloning without pixel-to-control inference.

Foundation Generalist Foundation

Schema-consistent state across every supported game. Cross-title agents on a shared vocabulary, not bespoke wrappers.

Player Player Models

Skill, intent, playstyle. Longitudinal per-player input and outcome data, schema-stable across patches.

Coach Assist Models

State plus outcome ground truth for models that recommend the right move. Train here, deploy against a live PRISM stream.

Reward Reward / Eval

Ground-truth outcomes for free: kills, objectives, win/loss, economy. RL reward shaping without extra labelling.

Runtime AI Assist

Read the game, not the screen.

Most consumer-facing game AI today reads the screen. PRISM lets the model read the game itself. Same hardware, lower latency, lower compute, higher accuracy.

Traditional Pixels in
Game frames 60 fps capture
Vision Model Heavy
LLM Reasoning
Output Advice
High latency High compute Inferred
With PRISM State in
Live state stream Frame-accurate
Structured prompt Lightweight
LLM Reasoning
Output Advice
15–35 ms Low compute Ground truth
Use cases On-device coaching Tactical assist Live narration

Let's talk.

Pre-training, runtime assist, evaluation, reward modeling. Tell us what you're building and we'll scope it together.

Get in Touch