Gigabytes ofbehavioral data collected
Get in touch Predicting humansunder pressure
We study the fast, pre-verbal layer of human behavior — turning measurable reactions into intelligence that helps people and systems perform when it matters most.
Loading interactive brain
From surprise to recognition. One record of human behaviour.
Our proposed framework follows what happens when a completely novel stimulus meets a player — and how their response changes when the brain recognises the same pattern again.
Model-freeModel-based · Proposed framework
A new stimulus. An individual reaction. A learned pattern.
- 01 / Model-free
Completely novel stimulus
A surprise with no matching learned pattern for this situation.
- 02 / Individual reaction
Different players. Different reactions.
An initially unpredictable response, varying from person to person.
- 03 / Model-based
The pattern is recognised
On a later encounter, a matching pattern in the brain supports a response informed by prior experience.
Transition condition: the same pattern is encountered again and matched to prior experience.
Working hypothesis: initial reactions have a person-specific random component; experience can support subsequent pattern recognition. This progression remains to be tested.
The distinction in this framework: whether the response draws on a learned pattern or arises in a novel situation. Response speed does not define the categories.
Millions of realhuman decisions.
Every AI today models the part of the brain that reasons. We model the part that reacts — the fast, pre-verbal layer nobody is building. Read straight from gameplay telemetry at 64 Hz, resolved to 15 ms.
5,500+years of cumulative
years of cumulative
human activity
Capturing decisions and reactions under pressure.
People representedin our datasets
Telemetry granularitydown to 15 ms
Initial results that
move the baseline.
We trained Ember and Orbit — the first models built to predict fight-or-flight outcomes — and benchmarked them against leading open-source and frontier LLMs over 1,000 predictions. They are more accurate, orders of magnitude faster, and orders of magnitude cheaper — because they read behaviour directly from telemetry instead of reasoning over text.
ModelAccuracyAvg. time / predictionCost / 1,000 predictions
EMBER by Mesocort AI73.5%0.01 ms$0.000002
EMBER by Mesocort AIORBIT by Mesocort AI71.2%0.34 ms$0.000062
Qwen 3.5 9B63.5%2.58 s$2.33
Qwen 3.5 9BGemini 3 Flash62.7%4.27 s$10.16
Gemini 3 FlashGPT 5.461.6%8.97 s$48.72
GPT 5.4DeepSeek V4 Pro58.9%0.91 s$10.66
DeepSeek V4 ProMajority-class
baseline51.5%——
baseline51.5%——
Best-performing configuration shown per model family. LLM results measured over 1,000 predictions; Orbit & Ember over the full validation set. Full methodology in our scientific paper (under review, available under NDA).
Grounded in research,
built in the real world.
We work with top researchers and research institutions.
Centaur AI
Human cognition foundation model
Marcel Binz, PhD — Author of Centaur, the foundation model of human cognition, who reviewed our scientific paper.
Masaryk University
Psychology of digital behaviour
Prof. David Šmahel — Joint research on esports athletes, pairing in-game and physiological data during competitive play.
VŠB — Technical University of Ostrava
AI research & PhD talent
Prof. Ivan Zelinka — Among the top 2% most-cited AI scientists worldwide: swarm intelligence and evolutionary computing.
Centaur AI
Human cognition foundation model
Marcel Binz, PhD — Author of Centaur, the foundation model of human cognition, who reviewed our scientific paper.
Masaryk University
Psychology of digital behaviour
Prof. David Šmahel — Joint research on esports athletes, pairing in-game and physiological data during competitive play.
VŠB — Technical University of Ostrava
AI research & PhD talent
Prof. Ivan Zelinka — Among the top 2% most-cited AI scientists worldwide: swarm intelligence and evolutionary computing.