The practice
How systems model the world — and what they cannot see
They are observing the world at the wrong level.
Collective life is continually rendered by something: an algorithm scoring a household, a policy category deciding who counts as a carer, a market pricing a region, a cultural frame settling what belongs where. The rendering carries consequences whether or not it is accurate, and its models are increasingly sophisticated.
Yet the gaps persist. Communities resist development that the data said they wanted. Households are locked out of systems designed to serve them. Platforms produce divergent and sometimes catastrophic outcomes for users the aggregate models said were fine. Cities invest in belonging and see attachment decline.
The pattern is consistent. Outputs and proxies get measured in place of what matters. Aggregation erases the variance that carries the meaning. Relationships, obligations and forms of value that have no field in the model are read as absence — and the household, the community, the user, the place come out wrong in specific, structural ways.
Propensities works below the level of indicators and above the level of anecdote. Collective life — in households, communities, markets, cultures — is organised in ways that the systems observing it cannot fully register. Each project builds an instrument that registers some part of what falls outside: the relational arrangements, forms of value, and affective conditions that the governing account has no category for. The gap between the two is the finding.
The lens is assemblage analysis: how a system of observation is composed, what it makes visible and what it structurally cannot see, and how it behaves when the world it acts on does not conform to its assumptions. Where the encounter has not happened yet, Relational Systems Simulation renders it in advance, so the failure modes are readable before deployment rather than after. The projects differ in scale and domain; the lens and the move are the same.
Methods
Relational Systems Simulation (RSS)
A researcher-directed design method combining synthetic ethnography, scenario-based design, and LLMs used as prismatic devices — constrained generative instruments that refract empirically grounded demographic, wellbeing, and cultural data into composite intergenerational scenarios. RSS generates anticipatory knowledge about how AI-mediated systems interact with culturally complex households before those interactions produce documented consequences. The second traversal tightened the method in three ways: cases are keyed to household configuration rather than ethnicity, since within-group variation is what culture-keyed archetypes flatten; expectations for each case are written and dated before the scenario is rendered, so departures are findings rather than reinterpretations; and every case now carries a trace sheet recording its design position, its departures, and the corpus anchor behind each claim. Deployed in Beyond the Score and Care in Translation.
Composite cases are built from published research and public data. They model household configurations rather than identifiable people or communities, and they are read as hypotheses to be tested. Validation by people close to the configurations is part of the method: the Beyond the Score cases were validated by Māori, Pacific and Asian young adults. Where commissioned work will inform decisions affecting a community, that community takes part in shaping and reviewing the cases, and in Aotearoa New Zealand the work is guided by a commitment to biculturalism.
Theory
What Language Does: A Mechanism Theory of Differential Effects in Digitally Mediated Communication
A theoretical paper developing a mechanism-level account of why the same digital platform produces divergent outcomes across users. The mechanism is communication itself — addressed expression operating through four conditions: elaboration, ratification, narrative continuity, and positional integrity. Where these conditions are present, self-organisation proceeds; where they are absent or distorted, the same environment produces harm. The framework has direct implications for AI governance: sycophancy is not a design quirk but the systematic destruction of positional integrity. Under review