Self reporting stops at what people say
It cannot show what someone does when attention, pressure and adaptation are live. It captures the story a person tells, not the behaviour underneath it.
How it works
How a few minutes of play becomes a picture of how you think.
It starts with why play works at all, and ends with a model you can put inside a product. This page walks the whole route.
Why gameplay
Play is fundamental to human nature. That is exactly why it is such an honest way to measure someone.
Every other way of measuring a person interrupts them. A questionnaire asks them to describe themselves. A test tells them they are being assessed, and they behave accordingly. A game asks for none of that. People play because they want to, and while they are playing they are simply themselves.
A test captures one moment, then ages
It interrupts the experience to take a reading, and goes stale as mood, readiness and performance move through the week.
Analytics show activity, not ability
Clicks, scores and completion rates say what happened. They say nothing about how the person got there.
What play reveals
A game is a microcosm of real behaviour.
Games ask people to do the things work and learning ask of them, compressed into a few minutes and repeated hundreds of times. Games, simulations, quizzes and interactive experiences can all contribute the same signals.
“Games are a microcosm of reality.”
Demis Hassabis · Google DeepMind
How they decide
Which option, how quickly, and what they do when the first choice was wrong.
How they adapt
What happens to their play when the rules shift or the pace climbs past comfortable.
How they focus
Whether attention holds through a long session, and where it starts to slip.
How they feel
Confidence, frustration and flow, read from behaviour rather than asked for afterwards.
The ontology
One vocabulary is what lets a puzzle game and a racing game report on the same scale.
Reading behaviour only works if every game speaks a common language. The Human × Game ontology is that language: 87 features every game is rated on. The map below is not a taxonomy drawn in a circle — it is those features correlated across roughly seven hundred rated games, so how far apart any two sit is measured. The six colours are the groups that fall out of the data, and they cut straight across our own Skill, Mood, Goal and Genre categories. Explore it below.
Human × Game ontologyRadius is distance · hover a term to trace it to the root
01
Mechanics become actions
150+ game mechanics resolve into the verbs a session actually asks for: obtain, avoid, choose, solve, collaborate.
02
Actions become signals
Those verbs, with their timing and accuracy, are what every human reading is inferred from. Never the genre label.
03
Signals become strengths
50+ cognitive skills, 60+ personality traits and 9+ moods combine into one comparable Skillprint.
Gameplay signals
What a session actually gives up.
A session is not scored on whether someone won. It is read on how they played: the shape of their decisions, what they did after a mistake, and whether they held their form as the game got harder.
Which option, how fast, and how consistently across a rising difficulty curve.
What a player changes after a failure, and how many attempts it takes to find a new route.
Whether accuracy and timing hold across a long session, and where concentration drifts.
Confidence, frustration and flow, inferred continuously from behaviour rather than asked for.
How far ahead a player is working, and whether they hold a plan or improvise each move.
In multiplayer and cooperative formats, how a player shares load, signals and defers.
Model architecture
Watch. Score. Tune the game, live.
A fine tuned vision model watches gameplay, scores the player against the ontology, and returns parameter updates that keep the game in flow. Nothing is added to the game itself, and the player is never interrupted.
See the developer view01
Game and SDK
Unity and JavaScript SDKs sit inside the game, alongside the loop you already built.
Flow02
Session capture
Play is captured as timestamped chunks, so a reading can always be traced back to a moment.
chunks03
Fine tuned vision model
Schema constrained inference at temperature zero, so the output is always machine readable.
VLM04
Structured output
Skill and mood readings returned for every chunk, against the ontology's own vocabulary.
structured JSON05
Profile and parameter updates
The Skillprint profile updates, and difficulty tunes live to keep the player in flow.
parameter_updatesTraining and validation
Fine tuning gives us the eyes. A live benchmark keeps us precise.
Supervised fine tuning on gameplay frames paired with expert structured output, then a blind model against model benchmark run inside real games — with a control arm where no AI is in the loop at all.
What the model learns
Fine tuning01
Supervised fine tuningGameplay frames paired with expert structured output, so the model learns the reading rather than the score.
02
Schema constrainedTemperature zero with JSON enforced, so every response is machine readable and comparable across games.
03
Tuning qualityTraining and validation loss tracked in the tuning logs for every run.
How we validate it
Benchmark01
Live benchmarkBlind A and B arms inside real games, plus a control arm with no AI in the loop at all.
02
11 frontier modelsGemini, Claude, GPT and open weight models, scored by one neutral scorer.
03
Statistics that hold upBayesian averaged ratings, 95% confidence intervals and Welch's t test.
Explore the products
That is the whole system. Now pick the route that fits you.
The same model, four ways in. Each one has its own page, its own outputs and its own next step.