Preface
The Programme asks how the cognition that lives between people — in teams, in institutions, in cultures — can be made a computational object of study, with the discipline such study requires.
Edwin Hutchins (1995) articulated distributed cognition as a foundational lens: thinking is not confined to individual brains but emerges from the coordinated work of people, tools, and environments. Thirty years later the framework remains theoretically influential — touching organisational research, HCI, education, and anthropology — but it has lacked the computational instantiation that adjacent frameworks (Bayesian cognition, connectionism) acquired long ago. Distributed cognition has been studied ethnographically and invoked as a conceptual lens; it has not become a computational research programme.
Multi‑agent large‑language‑model simulation appears to provide the missing instrument. Multiple language‑capable cognitive units, coordinating through dialogue, using shared representations, and producing collective outcomes — this is the architecture distributed cognition has waited for. But the instrument is not yet a discipline. The current literature operates largely at the level of phenomenon generation: simulations that replicate known correlations without pre‑registered novel predictions, without substrate‑invariant verification, and without the validation apparatus that turns generation into theory.
The Programme exists to provide that discipline. It treats multi‑agent LLM simulation as the computational instrument of distributed cognition, and submits its claims to a validation regime borrowed from structural econometrics, climate‑style intercomparison, and the philosophy of science. Its empirical anchors include the dynamics by which private beliefs become — or fail to become — public utterances within a team; the cultural grammars that operate as generative priors on what a collective can think at all; and the long re‑implementation of Hutchins's ethnographic cases as computational systems.
The work proceeds in collaboration with academic laboratories in cognitive science, computational social science, and HCI, and within operating institutions whose collective cognition is itself the object of inquiry.
Horizon
The Programme takes a position.
A team is more than the sum of its members' minds, and a culture more than the sum of its speakers' utterances. What happens between them — what is said and what is suppressed, what is licensed and what is unthinkable, which tensions are surfaced and which are kept private — is itself a cognitive structure. It has dynamics, constraints, and failure modes. It has been the object of ethnographic study for half a century; it has not, until recently, been the object of computational study.
Collective Intelligibility is the property of a collective whose reasoning structure is legible — to its members, to its leaders, and to the institutions that depend on it. The Programme treats this legibility as something to be engineered: simulations that make visible the dynamics by which private beliefs become public utterances; protocols that surface the cultural priors a group is unaware of; instruments that distinguish substrate‑invariant findings from artefacts of a particular model. Where ethnography described, computation can predict — but only if computation submits to the validation discipline that prediction requires.
The Programme refuses two adjacent failure modes: the use of simulation as illustration (no falsifiable claim, no discipline) and the use of simulation as proxy (collective behaviour reduced to a single utility function). Between them lies the harder work it is organised to do.
Method
Four interleaved fronts — each grounded in distributed cognition, cultural cognition, and the philosophy of scientific modelling, and held to a validation discipline borrowed from climate science and structural econometrics.
The Programme draws on distributed cognition (Hutchins 1995), cultural cognition (Kitayama & Markus 1991, Nisbett 2003), paradox theory (Smith & Lewis 2011, 2022), and the philosophy of science (Lakatos's progressive programmes, Mayo's severity testing). Each contributes a particular constraint on what a defensible computational claim about collective cognition is permitted to look like.
Multi‑agent LLM systems are treated not as models of behaviour but as instruments for studying it. Cultural grammars are operationalised as generative priors — sadō, ubuntu, bauhaus, sprezzatura, ayni — that license different modes of tension‑resolution. Role and identity are realised as conditioning, not styling. The instrument is calibrated against ethnographic ground truth before being trusted to produce novel predictions.
The Programme articulates a four‑level claim taxonomy — tuned fit, pre‑fixed parameter fit, novel prediction verified, mature progressive programme — and proposes LLM‑MIP, a multi‑model intercomparison protocol modelled on the climate community's CMIP, as a community standard. Findings that survive substrate variation (GPT, Claude, Gemini) and pre‑registered predictions are admitted; findings conditional on a single model are reported as such.
Simulation without field anchoring is exhibition. The Programme partners with operating institutions to anchor its simulations against real organisational dynamics — psychological safety in cross‑functional teams, cultural reasoning in strategic decisions, voice and silence in deliberative bodies — and submits its predictions to the institutions whose behaviour they purport to explain.
Works in the Programme
Current anchors: the CogSci accepted paper, and a community methodological proposal that adapts the discipline by which climate scientists test their models against each other — applied here to the simulation of how people reason together.
A computational study of the dynamics by which privately held beliefs do or do not become public utterances within a team, under varying conditions of psychological safety.
An open community proposal to give multi‑agent simulation the validation discipline it has been missing — adapted from how climate scientists test their models against each other. Collaborators across simulation, computational social science, and the cognitive sciences welcome.
Persona prompting performs a culture from the outside. What it takes to build multi‑agent systems that reason from inside one — developed in dialogue with cross‑cultural collaborators.
Two paired position papers on what is required for multi‑agent LLM simulation to make defensible claims about social and collective cognition — construct validity at the front, human‑in‑the‑loop validation at the back.
Who this is for
Plainspoken so a prospective partner does not waste a conversation. This Programme is research, not a product — and the lines between who it serves and who it does not are kept deliberate.
An institution with a deep, durable organisational or cultural question and the patience for an 18‑month inquiry — not a deployment timeline. Typical sponsors: foundations, corporate research divisions, ministries, and operating leaders carrying a question their HR function cannot answer.
Academic laboratories in organisational science, computational social science, cultural psychology, and cognitive anthropology. The deliverable is the peer‑reviewed contribution; the relationship is intellectual.
Individuals and institutions who want the discipline of this research to exist at all — and accept that what they sustain is the Institute’s independence, not a deliverable to themselves.
A culture‑analytics dashboard. An HR pilot. A diagnostic product. A platform on which to run an “AI culture survey.” The Programme refuses these on principle — distributed cognition is not legible to a dashboard, and pretending otherwise is the failure mode the Programme exists to refuse. Prospective partners looking for these are directed to operating firms.
The three working modes set out the terms of each in full.
Correspondence
For research collaborations, joint publications, advisory engagements, and field deployments —
A note via the form below — the Director responds personally.