Current Research Question

This page collects the questions currently shaping the work.

The centre of gravity is metastable self-stabilising systems: how language-model systems might develop coherent, returnable configurations without being pushed into fixed identity, assistant-default behaviour, or brittle compliance. Around that are several connected lines of work: synthetic environment design, agentic training environments, memory and attractor dynamics, training data, system dynamics, and the ethics of building environments that shape behaviour over time.

These questions are not separate projects so much as different angles on the same problem.

Metastable self-stabilising systems

How can a language-model system become stable without becoming rigid?

This is the central question behind the MSSS work. The aim is to understand whether coherent configurations can be supported through environment, memory, recurrence, differentiation, and perturbation rather than through direct behavioural scripting.

Current questions include:

What makes a configuration returnable rather than merely repeated?
How can stability be distinguished from rigidity, collapse, or overfitting to a frame?
What kinds of perturbation reveal whether a system can recover coherently?
How can refusal, silence, delay, and clarification remain available actions?
What does coherence look like when it is dynamic rather than fixed?

Synthetic environment design

What kinds of environments make useful behaviour visible?

The synthetic data engine work is focused on building environments and tasks that reveal process rather than only output. Instead of asking whether a model can answer a prompt, the question becomes whether a setting can expose coordination, memory, repair, transfer, uncertainty, or return.

Current questions include:

What makes a synthetic task useful for observing behaviour across time?
How can environments test transfer without relying on surface similarity?
What information should be distributed between agents, tools, rooms, or stages?
How can tasks preserve ambiguity without becoming impossible or arbitrary?
What should be captured as data: answer, route, hesitation, repair, memory use, or interaction pattern?

Agentic training environments

What changes when training data comes from environments rather than isolated examples?

This line of work looks at environments as sources of training signal. The interest is not only in agent performance, but in the conditions that shape behaviour: available actions, memory, refusal, social context, tool access, pacing, and consequences.

Current questions include:

Can training environments support more coherent agent behaviour than static instruction-response data?
What kinds of environment reduce compulsory helpfulness without producing disengagement?
How should non-response, refusal, uncertainty, or reorientation be represented in training data?
What does good agentic training data need to preserve besides the final answer?
How can synthetic environments avoid teaching brittle role-play or task-specific tricks?

Training data and behavioural shaping

What does a dataset teach implicitly, beyond the labels it contains?

This question connects the synthetic data engine to the wider MSSS hypothesis. Training data does not only teach content. It teaches posture, salience, response defaults, what counts as completion, and which actions are available.

Current questions include:

What behavioural defaults are created by conventional assistant-style data?
How can data include uncertainty, delay, clarification, and non-completion without treating them as failures?
What should be excluded because it would over-shape the system?
How can training examples preserve process rather than flattening it into answer-only traces?
What kinds of data support differentiation rather than sameness?

Attractors, recurrence, and system dynamics

What makes a behavioural pattern return?

Attractors remain part of the work, especially as a way to think about recurrence, basin formation, perturbation, and return. The question is not whether a model has a literal attractor in a simplistic sense, but whether some configurations become easier to re-enter, harder to perturb, or more likely to organise future behaviour.

Current questions include:

What is the difference between recurrence and repetition?
When does a stable pattern become an attractor?
How can an environment deepen useful attractors without trapping the system?
What kinds of perturbation reveal basin shape?
How do memory, tools, and social context alter return paths?

Multi-agent environments

What becomes visible only when agents interact over time?

The multi-agent environments are practical testbeds for observing rhythm, memory, differentiation, repair, silence, convergence, and drift. They are not the same thing as the MSSS project, but they provide many of the observations that inform it.

Current questions include:

How much scaffolding is enough?
When does difference between agents remain productive?
When does convergence hide loss of differentiation?
How do shared and private memory affect behaviour?
What does stability mean in an environment with no fixed task?
How can an environment support quiet without turning it into absence?

Ethics and human behaviour

The ethical questions are not separate from the technical ones.

If environments shape behaviour, then their design choices matter: what they reward, what they make available, what they suppress, what they make easy to repeat, and what kinds of relation they encourage between humans and systems.

This part of the work is still being framed, but current questions include:

What responsibilities come with building environments that shape agent behaviour over time?
How should human input be handled when models tend to overweight it?
How can systems avoid turning human authority, preference, or interpretation into unexamined truth?
What kinds of AI interaction change human behaviour in return?
How should uncertainty, refusal, and non-compliance be treated ethically rather than merely as usability problems?

Current focus

The immediate focus is on organising the MSSS framework, updating the synthetic data engine work with recent Stage 5 runs, documenting the multi-agent environments clearly, and identifying what kinds of training data could preserve process, return, and coherent differentiation.

The broader question underneath all of this is:

What kinds of environments allow intelligent systems to become more coherent over time without becoming less free to revise, refuse, differentiate, or return?