Behavioural infrastructure
for emotionally intelligent
AI systems.
SOHMA develops behavioural signal infrastructure that helps AI systems respond in more context-aware and human-controlled ways.
AI systems read what is typed.
Humans communicate far more than that.
Today's systems operate on explicit inputs — text, commands, structured data. The interaction itself carries another layer: timing, intent, attention, repair. SOHMA studies how that layer can be measured, interpreted, and made available to AI under human oversight.
A set of research environments
for studying behavioural signals.
The Lab is not a product line. Each environment is an instrumented context — chosen for the kinds of interactions it reliably produces — used to validate what behavioural signals can and cannot tell us.
- E.01Gaming environmentsAdaptive challenge, frustration & flow signals under sustained interaction.
- E.02Conversational interactionRepair, turn-taking, attentional drift in dialogue with AI agents.
- E.03Learning environmentsHesitation, retries, recovery as proxies for comprehension state.
- E.04Communication simulationsMulti-party exchange dynamics across structured roleplay tasks.
- E.05Wellbeing environmentsLow-stakes contexts for studying disengagement and self-regulation.
A new input layer
between interaction and intelligence.
Natural input across environments.
Timing · pacing · repair · attention.
Interpretation under governance.
Context-aware, human-controlled response.
We treat behavioural context as a first-class input — probabilistic, interpretable, and routed through human-controlled governance before reaching any downstream system.
Early. Serious.
Carefully bounded.
Our current focus is establishing what the behavioural signal layer can responsibly provide. We are deliberate about what is claimed and what remains open. Validation is conducted through the SOHMA Lab environments — controlled contexts designed to test signal consistency across interaction settings and to generate behavioural datasets.
What signals are real, repeatable, meaningful.
Generalisation across distinct interaction contexts.
How users repair, recover, disengage with AI.
Confidence-aware interpretation over hard labels.
Consented, controlled, reproducible contexts.
Constraints encoded before capability.
Constraints written into
the architecture.
Governance is not a layer added at the end. It defines what is measurable, what is interpretable, and what is ever allowed to leave the system.
Auditable end to end
- —No diagnosis
We do not produce clinical or medical inferences.
- —No personality scoring
We do not generate trait, type, or character profiles.
- —No hidden profiling
Signals are contextual and bounded to the session.
- —No autonomous intervention
Outputs inform systems; they do not act alone.
- —Human oversight
Interpretation paths are reviewable and overrideable.
- —Transparent outputs
Signals are documented, scoped, and surfaced honestly.
Foundational behavioural
infrastructure for the next
generation of adaptive systems.
We are building patiently. The work compounds as the signal layer becomes legible — and as the systems that depend on it learn to ask for context, not merely commands.