A referendum asks the whole country a yes-or-no question and counts the hands. A citizens’ assembly takes a smaller group, sits them down with the evidence for days or weeks, has them argue it through, and asks them for a reasoned judgement. The first has breadth and no reasoning. The second has reasoning and no breadth. Most of the interesting design questions in democracy live in the gap between those two things, and this piece is about a way to close it — not by choosing one over the other, but by running many small rooms on the same question at once and drawing their results together into a single, legible picture.
The rule underneath all of it, the same rule that governs every part of this work: the software rehearses the argument and provokes; people deliberate and decide.
Ireland is the case everyone reaches for. A Citizens’ Assembly of 99 randomly chosen members deliberated on abortion and the Eighth Amendment across five weekends between late 2016 and early 2017; 64% favoured allowing terminations without restriction, and the 2018 referendum then repealed the amendment with 66.4% voting yes (Involve; Electoral Reform Society). Notice the order. The assembly did the reasoning — the slow, informed, minority-protected weighing of a question that had deadlocked the political system for decades. The referendum did the deciding. The vote closely tracked the assembly, because the assembly had given the country a considered view it could recognise as its own.
That is the assembly at its best. But it rests on a small room. Ninety-nine people, however carefully chosen to mirror the population, are not the population. The assembly’s authority comes from the quality of its reasoning, not from its numbers, and there is a limit to how far a few dozen people can carry a national question on reasoning alone. The referendum supplies the numbers and none of the reasoning. Each instrument is missing exactly what the other has.
Here is the move. Because a well-built assembly seals its own record — a tamper-evident, write-once, cryptographically signed account of what was argued and where the room landed — there is nothing to stop many rooms running the same question in parallel. A party’s branches. A network of community groups that would never sit in one hall. Regions of a country. Each room deliberates on its own, at its own pace, and seals its own result. Nothing in that machinery knows it is doing politics: the rooms could as easily be a co-op’s branches, a union’s worksites, or a church’s parishes — any organisation too spread out to sit in one hall, and too serious to want its disagreements averaged away. Then those sealed results are drawn together into one map: a Consolidated Village Consensus.
The word “consensus” is doing careful work there, and it is worth being precise about what it does not mean. It does not mean an average. The one thing this map may never do is manufacture agreement by blending dissent away. What it does instead is show, across all the rooms: where the federation genuinely converges, where it divides and along which lines, and what holds up everywhere against what is only local. A position that one region holds and no other is surfaced, not buried. A split inside a single room is carried up into the national picture verbatim, attributed to the room that held it. Dissent is treated as a finding, not as noise to be smoothed out.
That is the difference that matters. A referendum’s majority erases its minority by design — 51% is a result and 49% is nothing. A consolidated assembly keeps the minority legible. You can read, off one sealed artefact, not just what most rooms concluded but exactly who disagreed, about what, and why. For a question where the disagreement is the whole point — and most questions worth deliberating are like that — keeping the disagreement in view is not a weakness of the method. It is the method.
Now the argument the tool makes possible, stated plainly and as potential rather than as something already running. If every political party ran the same question inside its own membership, and those results were consolidated, you would have something with the breadth of a national vote and the reasoning of an assembly — an informed, minority-protected read of where the country’s organised political communities actually stand, provable underneath rather than asserted. Not a poll of opinions. Not a referendum’s uninformed yes or no. A federated judgement.
That possibility is itself an argument for something upstream of it: reaching agreement on a national AI policy and the principles under it. A country that could produce this kind of informed national read on a hard question has a reason to decide, deliberately and together, what such a tool is for, who may run it, on what questions, and under what guarantees — before it is reached for in anger. The instrument and the constitution for the instrument belong together.
None of this works on trust in the platform, and it is not meant to. Every contributing assembly’s record is sealed on sovereign New Zealand and EU infrastructure, with no US-owned or China-owned cloud anywhere in the stack. Once sealed, a record is write-once: it cannot be altered, quietly re-written, or backdated, and its signature can be checked offline by anyone, without having to trust us or any platform to vouch for it. The consolidated map is sealed the same way, and it carries a verifiable reference back to every room’s record underneath it — and now a cryptographic binding that a room’s sealed record is the very report that entered the consensus, so a swapped or edited body is caught rather than quietly absorbed. The national picture is one you can audit down to each room, not one you are asked to believe.
The provenance stays straight, too. What came from a simulated rehearsal panel and what came from real members is never blurred together; a consolidated read of rehearsals is labelled as exactly that, and only what members actually decided is ever described as where members stand.
Three of them, stated because leaving them out would make the rest less true.
The federated tier is built and has been validated end to end against our own live governance infrastructure — real dissent-preserving synthesis, real seals, real offline verification. No organisation is running it in production yet. It can be demonstrated; it is not in anyone’s hands today.
It never speaks for tangata whenua. A synthetic persona is barred from standing in for tangata whenua, iwi, hapū or Māori as an identity — a refusal built into the code, not a settings option. That bar is a best-effort control and is labelled provisional in the code itself, explicitly subject to confirmation by tangata whenua that we have not yet had. Where a question engages tikanga or Māori data, it routes to real people deliberating in the ways they choose.
And it decides nothing. The output is a contestability map — where the rooms converged, where they split, where the evidence is thin, what people should stress-test — never an answer. When a rehearsal reaches agreement too fast, the tool flags it as under-explored rather than settled. The machine’s whole job is to make the human decision better informed and the record of it verifiable. People deliberate and decide.
A referendum is a single loud question and a count. An assembly is a single quiet room and a judgement. A federated assembly is many quiet rooms asking one question, their judgements drawn together without any of them being averaged away, sealed so the whole picture can be proven and read down to the last dissent. It is not a way of getting people to agree. It is a way of seeing, exactly and provably, where they do and where they do not — which is the thing a country actually needs before it decides. And not only a country: any organisation that has to decide together has the same need. The country is just the biggest room.
| Referendum | Citizens’ assembly | Federated assembly | |
|---|---|---|---|
| Breadth | Whole electorate | One small room | Many rooms at once |
| Reasoning | None — a raw yes/no | Deep, informed, days of it | Deep, in every room |
| Minorities | Erased — 51% wins, 49% is nothing | Protected inside the room | Kept legible across all rooms |
| The record | A count | A written report | Sealed, offline-verifiable, room by room |
The software rehearses the argument and provokes, across many rooms at once. People deliberate and decide. And the record proves what was argued.