Research

Research for adaptive intelligence.

Biology constrains the starting state. Culture carries accumulated knowledge. Individual experience changes what happens next. Embreier studies how to represent those changes rigorously enough for a model to learn from them.

Biology · Culture · Individual experience · Learning

The thesis

Model the agent as something that changes.

An agent changes. It arrives with priors, inherits what others worked out, and is changed by what happens to it. The representation holds all of that, so it can study it.

01BiologyWhat an agent starts with
02Culture and environmentWhat it inherits from others
03Individual stateWhat the evidence suggests about it now
04ActionWhat it does in that state
05OutcomeWhat the world answers
06LearningWhat changes as a result
07New stateWhat the next action starts from

07 returns as the new 03The next action starts from a changed agent

The programme sits at the intersection: how the state of the agent itself changes through action, consequence and experience.

The programme

The Human Model
research programme.

Human-development science supplies candidate variables and priors. Each one earns its place in the model by improving a real decision.

0.1First principles

First-principles decomposition.

Break human development into state families, process families and transition primitives, so each candidate variable has a defined place, a unit and a timescale before it is modelled.

0.2Convergence

Cross-disciplinary convergence.

Neuroscience, biology, developmental and applied psychology, pedagogy and behavioural science each describe how people change. Where they converge, we hold a candidate prior. Where they diverge, we hold a question.

0.3Culture

Culture and universals.

Most of what a person knows was worked out with others. We separate what holds across people from what a team, a site or a culture carries, and study how practice travels between them.

0.4State

Probabilistic human state.

A person is partly observable. Their task-relevant state is held as belief with uncertainty and an expiry, kept apart from the evidence it came from, and open to their correction.

0.5Empathy

Functional empathy.

Reading what a person needs next from what they have seen, tried and demonstrated on the task, so the machine can choose the move that helps: a cue, a question, an explanation or a step back.

0.6Reciprocity

Reciprocal learning.

The machine learns which help works for this person. The person builds capability through the work. We study both directions, including how help today shapes independence tomorrow.

0.7Dual outcomes

Dual human and world outcomes.

Every transition is scored twice: on what happened to the process, and on what happened to the person or team acting in it. Performance today and capability later are measured separately.

0.1
First-principles decomposition.Break human development into state families, process families and transition primitives, so each candidate variable has a defined place, a unit and a timescale before it is modelled.
0.2
Cross-disciplinary convergence.Neuroscience, biology, developmental and applied psychology, pedagogy and behavioural science each describe how people change. Where they converge, we hold a candidate prior. Where they diverge, we hold a question.
0.3
Culture and universals.Most of what a person knows was worked out with others. We separate what holds across people from what a team, a site or a culture carries, and study how practice travels between them.
0.4
Probabilistic human state.A person is partly observable. Their task-relevant state is held as belief with uncertainty and an expiry, kept apart from the evidence it came from, and open to their correction.
0.5
Functional empathy.Reading what a person needs next from what they have seen, tried and demonstrated on the task, so the machine can choose the move that helps: a cue, a question, an explanation or a step back.
0.6
Reciprocal learning.The machine learns which help works for this person. The person builds capability through the work. We study both directions, including how help today shapes independence tomorrow.
0.7
Dual human and world outcomes.Every transition is scored twice: on what happened to the process, and on what happened to the person or team acting in it. Performance today and capability later are measured separately.

Every theme ends in a variable with a passport, a simpler proxy it must beat and a result that would remove it.

The testbed

Somewhere a prediction can be scored.

A developmental claim is only testable where state is observable, action is consequential and the outcome arrives while anyone still cares. Industrial operations give all three.

Two operators working at a control panel in an industrial control room.
Why industry first Observable state Accountable action Measurable consequence Recurring decisions
Industrial fault traceHuman decision retained
S₀Line statePressure anomaly
O₁Observation11 sec after
A₁ActionSequence C tested
R₁₂OutcomeRestart · 14 min

Evaluation

Measure what actually improves.

Same problem. Same data. Same input budget. Measure the difference.

A · System only
Strong frontier model

The same process evidence, the same tools and the same memory budget.

B · Static profile
A plus a static profile

A static role, task or team profile added to A.

C · Dynamic actor state
B plus governed actor state

A dynamic, governed person or team state in context, and the adaptive intervention policy.

The primary metric and its threshold are preregistered before data is seen. C earns its place by beating B net of burden and governance cost. Prediction, calibration, retention, transfer and independence are measured separately, on future time and on new sites.

What improves

Measure what lasts.

Four dimensions, measured separately, on future time and in a new situation. Every one of them can move while immediate performance holds steady.

0 / 7

01 RETENTION still there later 02 INDEPENDENCE greater independence 03 TRANSFER other lines, other sites 04 ADAPTIVE EXPERTISE new situations, unseen A NEW SITUATION IMMEDIATE PERFORMANCE · UNCHANGED

Research frontier

The work behind adaptive intelligence.

Human development is the research foundation. State, intervention and improvement are what we build outward from.

Foundation

NESTED LAYERS · ONE CONSTRUCT Foundation

Human-development construct

The research framework for representing human learning and adaptation, and for keeping observation, inference and judgement apart. Everything below is built outward from it.

Read →

Systems

STATE CONTEXT ACTION OUTCOME WHAT THE SYSTEM HOLDS, AND WHAT RETURNS System 01 Representing state

What belongs in the record, what the system is allowed to believe, and how uncertainty stays visible.

ACT ASK ABSTAIN THE MOMENT, AND THE THREE MOVES System 02 Choosing intervention

Which intervention helps, when to ask instead, and when the right move is to watch.

RETAINED INDEPENDENT TRANSFERS ADAPTIVE MEASURED SEPARATELY, OVER TIME System 03 Measuring improvement

Retention, independence, transfer and adaptive expertise, each measured on its own.

The frontier

Build the frontier of adaptive intelligence.

The next generation of AI will need to understand more than information. It will need to understand change, adaptation, context and consequence.

We are bringing together researchers across AI, neuroscience, biology, human development, behavioural science, human factors, philosophy, ethics and HCI to build it.

Join the research frontier →

AI · Neuroscience · Biology · Human development · Behavioural science · Human factors · Philosophy · Ethics · HCI