A living hypothesis about a person, not a replica.
I use the term digital twin to mean a continuously updated, idiographic model of a person (idiographic means built for one specific individual, not a population average), organised across seven analytic dimensions. I treat it as a living hypothesis about that person, not a replica and not a representation that claims to mirror them. It earns its place by being testable, revisable, and contestable. Everything below is conceptual and architectural work from the ATLAS paper. None of it is a validated product, and I do not present it as one.
Where the digital twin comes from, and what it is not
Where the digital twin comes from
A living hypothesis, not a representation
What the twin is permitted to represent
The twin is organised across seven analytic dimensions that define what the model is allowed to hold. They are not a taxonomy of what a person is. They are coupled analytic axes, and their coupling is what disciplines the inferences the twin is permitted to make. A reduction in collaboration-platform activity, to take one common signal, is consistent with deep-work focus, a role change, communication migrating to channels the platform does not capture, leave, caregiving demands, or genuine disengagement. A twin that collapses these alternatives into one confident inference is not sophisticated. It is wrong by construction.
Twin
A living hypothesis about a person, refined by seven analytic dimensions.
Select a dimension to reveal what it holds
A living hypothesis about a person, refined by seven analytic dimensions.
Functional
Biographic information, qualifications, and role history. The structural scaffolding within which inferences from the other dimensions are read.
Psychological
Dispositional traits, values, and motivational drivers, held as parameters that can shift over developmental time as evidence accumulates, not as estimates fixed at first measurement.
Cognitive
Problem-solving patterns, learning trajectory, and decision behaviour, read against the context that gives the behaviour its meaning.
Behavioural
Observable communication, collaboration, and work-rhythm patterns in identifiable contexts, including patterns of conflict and recovery from failure.
Physiological
Sleep, heart-rate variability, and fatigue signals relevant to recovery and stress response. Held only where the person has given explicit opt-in.
Performance
Key-performance-indicator outcomes, delivery patterns, and ratings across a horizon long enough to distinguish stable performance from situational performance.
Ecological
The role context itself: team dynamics, organisational stressors, and the culture in which the person operates. The work environment is inseparable from the inferences drawn about the person in it.
What keeps a twin honest
Simulations that direct attention, not decisions
Before any assessment, I run scenarios on the twin to find where the work should probe. These are hypotheses about what could happen, never predictions or verdicts.
Succession planning
Run candidates through a chief-executive-transition scenario, with the twin generating hypotheses about how each navigates stakeholder tensions before anyone steps into the role.
Continuous signal under strict constraint
The twin runs on continuous multi-source signal rather than a single sitting. The precedent is digital phenotyping (using the everyday digital traces a person already leaves, such as calendar and message patterns, as behavioural signal), where dense everyday data reveals within-person patterns that episodic instruments miss. That same density is why governance is not an afterthought. Under the power asymmetry of an employment relationship, a technically stronger twin can be ethically worse, not better, because it widens the employer's informational advantage over the worker. The streams are also not interchangeable proxies for psychological constructs. Calendar density is not conscientiousness. Email volume is not collaboration.
Purpose limitation
each stream is justified against a named decision, and the twin is not permitted to repurpose data beyond it.
Data minimisation
collect the smallest data sufficient for the inference, not the most data available.
Use separation
data collected for development must not migrate to discipline or termination.
Opt-in for sensitive streams
physiological signal is held only with explicit, revocable consent.
Contestability by design
the person can inspect, contest, and withhold any signal, and opting out does not affect access to the opportunity.
Human authority on consequence
any decision that materially affects the person requires human evaluation, not an automated verdict.
Common questions about digital twins
Related work and engagements
AI System Psychology
Cornerstone Hub
The parent discipline. The science of designing, measuring, and assuring AI systems that decide about people.
AI-Driven Assessments
Cornerstone Hub
The ATLAS architecture the twin lives inside. The digital twin is Layer 2 of a five-layer assessment system.
The Algorithm of You: Hyper-Personalised Wellbeing Models
IPPA World Congress, 2025
How AI can tailor wellbeing models to the individual, and the ethical constraints required to do it responsibly.
Psychologically Safe AI Infrastructure
6-12 week engagement
Build and assure person-specific AI systems with construct clarity, harm analysis, and lifecycle monitoring.
When AI Becomes Your Therapist: The Audit Nobody Is Running
Article
AI therapy bots are moving from beta product to clinical product without the validation any other clinical tool would require. The class-action wave is twelve months out. Here is what an AI-IARA audit catches before it lands.
Construct Drift: The Silent Failure Mode in Deployed AI Assessment
Article
Construct drift is the gradual shift in what an AI assessment is actually measuring after deployment, even when the model weights are frozen. It is the most expensive failure mode in deployed people-impact AI, and almost no one is watching for it.
Discuss your digital twin
If you are considering a person-specific model for assessment, development, or wellbeing, the first questions are not about data. They are about what the system should never be allowed to decide, who can contest it, and what evidence would force it to revise. I am happy to think through those constraints with you before any architecture is built.