Work the problem before you vote on it.
A shared map, teams taking one strand each, votes the members take part in, and a record of what was decided and why. For a members-only group.
Most group decisions arrive as a motion: a proposition, a discussion, a show of hands. What is missing is the part before that — agreeing what actually causes what, splitting the work so people can go deep rather than broad, and surfacing the places where two teams' conclusions cannot both hold. This is that part.
The question this demonstration works through
Should the way we tax land change, and what would actually happen to housing if it did?
It was chosen because it is genuinely contested, because reasonable people reach opposite conclusions from the same evidence, and because it is a system rather than a slogan — the sort of question that goes badly when it is put straight to a vote.
Background material the teams were given
Every deliberation starts from a briefing. This one draws on arguments that circulate in public debate. It is presented as the teams received it — as a position to be tested, not a conclusion to be accepted.
Tax what we have too much of, and less of what we want more of. We tax wages heavily and the increase in land values barely at all — so effort is taxed and holding land is not.
- The claim. A charge on the unimproved value of land, offset by cutting the tax taken from wages, would make holding land idle expensive and working less penalised. Revenue-neutral overall.
- Why land specifically. Land cannot be moved offshore or hidden, and its value comes largely from what the community built around it — roads, schools, pipes — rather than anything the owner did.
- The mechanism it targets. Land held back from development while its value rises; bank credit flowing to land rather than to businesses; housing treated as the household's main retirement savings.
- The strongest objection. People who are asset-rich and cash-poor — notably retired owners — would face a bill they cannot pay from income, and a deliberate flattening of house-price growth falls hardest on those who planned around it.
- The contested empirical question. Whether a holding charge is passed through to renters or absorbed by the owner. Both are argued from evidence.
The briefing deliberately includes the strongest objection alongside the proposal. A deliberation whose background material only argues one way produces a vote, not a decision.
How a session starts
Any member of the Village can create a session by naming it, and the person who does is its first facilitator. A session facilitator can make another member a facilitator too, for example the person whose material is being worked on. The session is then built in four moves.
Only members who have been given a place can open the session. Everyone else in the Village cannot, and an observer sees the map and the results but no team thread. Messages in a thread show what was said, not who said it.
One map, authored by people
Before anyone argues, the group agrees what affects what. A facilitator types three plain lists — the teams, the variables and the links (cause to effect, plus or minus) — and the map is laid out. Arrows carry a sign: + means the two move together, − means one rises as the other falls. Loops that reinforce themselves are marked R; loops that push back toward balance are B. The map is the group's own: it is typed in from what the group agreed, not generated, which is what makes it arguable. A larger map can be loaded for the session by the platform operator.
Three of the circled loops reinforce. Speculation: rising land values attract buyers expecting further rises, which raises values again. Credit: banks lend against land, which bids up the land they lend against. Savings: higher prices make housing a better store of wealth, which raises demand for it. Two push back. Supply: homes completed eventually restrain price. Holding charge: a tax on land value makes idle land expensive to keep, which releases sections — that one is the proposed intervention, and whether it is strong enough to damp the other three is exactly what the teams disagree about.
A map does not have to be a cause-and-effect system. It can also be a register of claims, such as the claims of a book, a constitution or a set of policies. There each box is one claim with a tier (Established, Contested, Hypothesis or Aspiration) and typed links to the claims it rests on or cuts against. If AI agents drafted the register, the tiers and links are a draft for people to read and correct, and a claim counts as confirmed only once a session facilitator marks it. The worked example on this page uses the cause-and-effect kind.
Five strands, five teams, one each
A whole-group discussion of a system this size produces a discussion of whichever part was mentioned last. Instead each team claims one strand and goes deep. They see the whole map; they answer for one part of it.
Inside one strand
Each team gets its own thread, which the people on that team and the session facilitators read. This is an excerpt from the renters and first-home access team — invented participants, working a real disagreement.
A decision point is not a show of hands
When a team is ready, it puts a specific question — not a mood — to a structured vote. The facilitator chooses the vote type (consent scale, single choice, multiple choice, ranked choice or quadratic), the consensus mode (consent, majority or unanimous) and, if wanted, a required participation. This one uses the consent scale: the question is not "who likes this" but "can anyone not live with it, and why". An objection has to name its kind — safety, values, effectiveness, process, resources or other — and give a reason.
The AI can run a pass that records a suggestion against a decision point. It cannot confirm one — a facilitator edits and confirms, and that confirmation is what opens the vote. Members vote through Polls, and a facilitator closes the vote, which records the tally with the strand. A machine suggestion that nobody confirmed is not a decision anyone made.
Checkpoints, so nobody works in a hole
A facilitator fires a checkpoint, or the session sets a cadence and the system fires them, and each live team gets a digest of where it has got to. The digest is machine-drafted and then edited and approved by a person; the draft body is not shown to other teams, and what is approved is posted to the team's thread and recorded. If a thread has fewer than three messages the draft says there is not enough to summarise, rather than inventing a summary. Each team's facilitator then acknowledges the checkpoint (until it does, reminders escalate from the bell to a push message to email), and a session facilitator completes it.
Where two teams cannot both be right
Because the strands are parts of one map, some of them touch. When two coupled teams reach conclusions that pull against each other, that tension is found by the system, not raised by anyone's opinion. Which strands touch comes from the map's own structure; whether their conclusions pull against each other is read from the outcomes of their closed decision points. An open cross-check stops a strand being closed until its facilitator records a position: resolved, or an acknowledged disagreement with a note. A rebuttal can also be filed, but it is only recorded alongside the cross-check and does not clear it.
A recorded disagreement is a legitimate outcome. Teams are not forced into a synthetic consensus; they are required to state, with a note, that they looked at the same coupling and still disagree, and that note is sealed into the strand's record when it closes.
The membership decides, not the working group
Teams do the work; they do not own the outcome. A session facilitator starts a round, choosing the consensus mode and an optional quorum. Each concluded strand then gets its own vote of all members, against an electorate fixed at the moment the round opened — so the roll cannot move while the vote is running. The facilitator approves a briefing, which opens the vote, and later closes it.
The result is counted, not chosen. When the vote closes, the result comes from the votes under the consensus mode and quorum the members were told. A facilitator cannot pick it. Without a quorum nothing is decided and the vote stays open. The one thing a facilitator can do is set a spoiled vote aside, which needs a written reason of 20 to 1000 characters; the reason is sealed with the result, nothing is ratified or returned, and the strand can go to a new round.
What the record actually says
A sealed deliberation record is not a summary someone wrote afterwards. Closing a strand seals its record (today the close is made on the server and the app does not yet show a close button): what the team concluded, and any disagreement it acknowledged. Each ratification result is sealed as well: the totals and quorum with no voter identity, who closed the vote, the rule it was counted under and, for a set-aside vote, the written reason. A later round seals under its own entry, so the first round's entry stays as it was. Records are signed, so a later copy can be checked against them.
Sealing fails safe. Records are signed by the governance service, not by the Village itself. If the signing service cannot be reached, no record is produced, rather than an unsigned one presented as sealed. The layout above is an illustration of what the records hold, not a copy of one.
Putting documents in the room
A session can hold documents — a report, a book, a set of policies. When an owner or admin of the Village agrees, members of the session can ask questions of them. Answers are written on our own servers, checked against the passages they cite, and show those passages; if no written answer can be checked, the closest passages are shown in their own words.
A document that names an author who is a member needs that author's own agreement, which they can withdraw at any time. Owners, admins and session facilitators can open a Documents summary — documents, authors, how each agreed, consent and terms of use — to copy, download or print. No AI model writes it. A separate session type, Role fit, lines up a role's duties against the work done and the organisation's plans, quotes the words on each side and treats every match as a draft until a person confirms or rejects it. This page does not demonstrate Role fit.
Where to next
Deliberation Village is the heaviest of the deciding tools. If it is more than your group needs, the lighter ones are worth seeing first — they are not lesser versions of this, they answer different questions.
What is real and what is not
- Real. Creating a session and adding people, the map and claim registers, strands and teams, decision points and their vote types, checkpoints and digest approval, cross-checks, ratification rounds with a pinned electorate and a result counted from the votes, setting a spoiled vote aside with a sealed reason, asking the documents, the Documents summary, and sealed strand and ratification records. All of it is built and runs, except that closing a strand, which seals its record and opens it to ratification, is a server action the app does not yet offer as a button. Record sealing needs record signing to be available to the Village; if it is not, the close is refused and nothing is sealed.
- Invented. Every person, message, digest, vote and number on this page. The deliberation shown never happened.
- Real, and gated twice. Reaching Deliberation Village needs a platform switch and, separately, a switch on the individual Village. This page does not tell you whether any particular Village has turned it on, and it is not itself a live Village — what you are reading is a worked example, not a session you can join.
- Language. This demonstration page is in English only. The claim register screens are in English only too; other screens of the app are partly translated.
- Not claimed. That any of this produces better decisions. That is the argument the software is built on, and it is not evidence. The briefing, digest and suggestion drafts are written by a machine; a person edits and approves each one.