Signal · AI Labs
Train on how humans behave in the real world.
Train on how humans act, decide and adapt. Skillprint turns real gameplay into rights-cleared, model-grade human intelligence data: what happened, joined to how people decided, adapted, felt and performed.
Research, white papers and the live benchmark are at skillprint.co/benchmark.
The missing layer in AI training data
Game state data explains what happened. Skillprint adds context for why.
Most datasets capture outputs, pixels, physics or action traces. Skillprint adds fluid human reasoning and behavior, synchronized to the same timeline.
The Skillprint layer
Only from real playWhat you receive
One session. Up to six synchronized views of the human and the world.
Each stream is keyed to the same session and timeline, and can be delivered together as model-ready bundles.
01
Visual stream
Timestamped screenshots and/or video of real human play, ordered by session and time.
02
Game-state stream
Game, level, configuration, objects, goals, points, outcomes, physics and real-time difficulty changes.
03
Player state and decisions
Cognitive skill levels, mindset, experience, controller inputs, goals, strategies and outcomes.
04
Cognition, flow and mood
Assessed cognitive skills plus flow and mood signals, with a human-reviewed golden subset.
05
Human ground truth
Player-reported mood and blind A/B experiment arms, including a non-AI-assisted control.
06
Longitudinal profiles
Session- and player-level scores across 14 cognitive skills and 9 moods, tracked over time.
The dataset
Each configured session is a labeled reasoning example.
Gameplay captures the process behind an outcome: the world state, the decision, the player's cognitive and emotional state, what happened next, and how that same player changes over time.
Discuss your data needs01
Real world relevance
Gameplay maps to how people actually reason and decide under changing conditions, not to how they describe it afterward.
relevance02
Structured labels at scale
Session data is synchronized against one timeline, making visual state, telemetry, cognition, flow, mood, controls and outcomes directly comparable.
labels03
Collected from real play
Human gameplay comes from real players across partners’ live games rather than only hired or curated capture sessions.
live play04
Rights and consent built in
Data streams are sourced from rights-cleared games and are player-consented and anonymized before delivery.
provenanceWhat it answers
Where do humans and AI perform best together?
The data carries blind A/B arms and a non-AI-assisted control, so it answers questions a model-only leaderboard structurally cannot.
01
Does this model actually improve the person using it?Measured against the same person's unaided baseline, in the same task, on the same day.
02
Which model amplifies which kind of person?Lift is not uniform. A model that helps a cautious planner may do nothing for a fast improviser.
03
Where is the human still better alone?The control arm runs with no AI in the loop, so the cases where assistance hurts are visible rather than assumed away.
04
What does the lift cost?Goal attainment scored beside token cost and latency, so efficacy is judged against what it took to get there.
What labs can do
Train and evaluate how models act, adapt and recover in dynamic environments.
Five things the same corpus supports, from fine-tuning a world model through to proving a system actually helps the person holding it.
01 · adaptation
World-model adaptation
Fine-tune on state to action to next-state pairs across multiple environments and human/agent strata.
02 · reasoning
Reasoning and planning
Learn from decisions, strategies, outcomes and cognitive-skill labels rather than action traces alone.
03 · personalization
Personalization
Study how the same environment and intervention affect people with different cognitive, mood and experience profiles.
04 · evaluation
Human+AI evaluation
Use blind A/B arms, unaided controls and human-reported outcomes to measure whether systems genuinely improve performance and experience.
05 · recovery
Failure, recovery and adaptation
Capture unsuccessful attempts, strategy changes, retries and recovery paths so models can learn how people recognize errors and revise their approach.
Work with us
Tell us what you want to measure.
Three ways to work with us: license the gameplay data we already hold, commission collection built to your spec, or benchmark your models against real human play.
01
LicenseExisting human intelligence gameplay datasets.
02
CommissionRights-cleared collection around a target environment or action vocabulary.
03
BenchmarkBuild a human+AI benchmark around the outcome your team cares about.
The benchmark
See how models score against real human play.
A leaderboard position says a model is good at the benchmark. It says nothing about whether a person got further with it in their hands. Signal scores people unaided, AI alone, and people amplified by AI inside real games. The live leaderboard sits on the benchmark page.