The first implementation of Structured Democratic Dialogue that runs without a room
True Causes asks a group the question almost nobody asks — one pair at a time, does A actually influence B — until a map of causes exists instead of a ranked list of complaints. A wall per group, ideas routed between groups on corroboration, and every group's map merged into one.
Twelve problems, arranged by cause. Ringed items on the left are the real drivers: quiet, structural, rarely on anyone's mind. The right-hand side is what people actually feel week to week. The dial sets how the room decides what matters.
Two findings, both fifty years old, both still designed around rather than designed for.
Thomas Dye called it the erroneous priorities effect: in complex situations the problems people rank highest are reliably the ones furthest downstream. Salience runs opposite to depth. A warehouse names picking errors, an ER names waiting times, a ministry names public trust — and every one of those is a consequence.
John Warfield called the second one spreadthink: put fifteen informed people in a room and their individual priorities barely overlap. Not because anyone is wrong, but because a complex situation exceeds what one person can hold. In our runs, 61–89% of all proposed items receive at least one vote. Dispersion is the normal state, not a failure of the group.
Put those together and the standard tools break in a specific way. A poll measures salience and calls it priority. A workshop reaches consensus on whatever the loudest coalition framed first. Both produce agreement. Neither produces a cause.
Why not just ask better questions? Because the constraint is not question quality. In simulation, once an item survives the vote, structuring recovers 64–97% of the true drivers. A driver that receives zero votes is unrecoverable at any later stage. What gets structured is the whole ballgame.
Why not let AI decide? A model can read an argument into a structure — ours does. It is not allowed to judge one. Every verdict on this platform is computed by a fixed rule or cast by people, so there is no step where a model's preference becomes the group's conclusion.
The flow is Structured Democratic Dialogue, implemented rather than approximated. One triggering question, small stratified groups, and a strict separation between what is a fact and what is a value.
One idea per card, one claim per idea, answering a single triggering question.
Authors sharpen meaning. Duplicates merge only when both authors consent — the defence against a popular framing multiplying itself.
Participants link ideas they judge alike; a link holds when two people make it independently. Names come last, and never gate a vote.
A fixed sticker budget for importance, plus seats reserved for overlooked ideas that someone can back with evidence.
Does A significantly influence B? Answered pair by pair at 80% of votes cast, with transitive consequences inferred for free.
The causes, deepest first. A feedback loop counts as one deep cause, and items nothing connects to are shown rather than quietly dropped.
Propose what to do, then judge each action against each root cause separately — not a wish list.
Every group runs all seven stages on its own wall. What makes that add up to one answer rather than forty is the layer underneath: shared claim identity, corroborated routing between groups, and a merge that turns many local maps into one.
Structured Democratic Dialogue already scales further than most assume — co-laboratories routinely run one to two hundred participants with a trained facilitator, and the practice has fifty years behind it. What it has never had is a way to run without the room. This is the first implementation that does.
Three things bind a co-laboratory to its venue. Everyone reads the same wall, which caps how many ideas can exist. One facilitator holds the floor, which caps how many people can speak. Everyone is present at once, which caps the whole thing to a day of shared calendar. The obvious substitute — run many sessions and staple the outputs together — fails in a specific way: each session reaches its own conclusion, nobody can tell whether two of them said the same thing in different words, and whoever captures one session captures that slice of the answer.
Participants sit in groups sized so the wall stays readable and the stakeholder mix stays varied — on the order of a hundred people and the ideas they produce. Within a group, each influence question goes to a panel rather than the whole crowd, and nobody is asked to read everything: each person is dealt a randomised slice. One person's load is fixed whether the question has two hundred participants or two million.
A room is not the unit — the question is. Open a dialogue and the platform recognises other rooms already asking it, however differently worded, and joins them into one deliberation. Their groups become groups of the same question, and claims move between all of them.
The same claim written in two groups is counted as one claim raised by two groups. That single choice is what makes scale computable: without it, a claim in forty rooms is forty items and nothing can be counted at all. Claims several groups have raised are then carried to the groups that lack them, alongside a smaller allowance for ideas only one group has seen, so a lone true insight can still spread. Support limits volume, never visibility, and writing more cards buys no extra reach.
Counting groups only means something if you can tell when two of them said the same thing in different words. The platform spots the likely pairs and puts them to the people who wrote them. It never merges anything itself: two members have to agree that two wordings are one claim, and either can take it back.
The asymmetry is deliberate. A false merge is an attack — fold your claim into a popular one and you inherit its support — while a missed merge only costs a little efficiency, since the claim still travels as two items instead of one. Claims that contradict or reverse each other are never offered as the same claim, however similar they look.
Each group structures the wall it actually deliberated on and produces its own map. Those maps are then folded together: the same claim is one node, every connection carries how many groups certified it independently, feedback loops collapse into single deep causes, and roots are read off the merged graph. Where two groups certified opposite things, the contradiction is published as a conflict and is referred, like any deadlock, to two other groups that must agree — never averaged away.
Why a hundred, not fifteen. A group is sized so its wall is still worth reading and its stakeholder mix is still varied — two constraints pulling opposite ways. Above that, each member is dealt a random slice instead. At 5% exposure on a large wall, the crowd still surfaced 11 of the 12 real drivers.
Where a deadlock goes. Not back to the group that split, and not to everyone. The question is referred to two other whole groups, chosen in a way nobody can steer toward a friendly audience. Each answers on its own, and they have to agree — if they don't, the pair is published as contested. One group is never enough: in simulation a single referral hands back a wrong answer 29.6% of the time under attack, where two or more does not.
What it costs to rig. Suppose three people coordinate to keep one cause off the map. They cannot vote it down — a blocked question is never recorded as “no” — so they vote to deadlock, and that alone cut the map to 4 of every 10 causes found. Referring deadlocks to two other groups took it back to 8 of 10. Taking the first pass away from any group that deadlocks as a habit, 9.5 of 10. Reshuffling which group gets which questions, 9.8 of 10. After that, three insiders cannot bend the result; they would need roughly one person in five across the whole verified population. Not “this is safe”, but this is what breaking it would take.
No model decides anything. A language model is used in one place — reading a plain-language argument into premises so the logic check can run — and it is forbidden to judge whether the argument is valid; a fixed engine does that. Deciding that two claims are the same is left to people.
Support has to be earned. A claim carries weight only when independent groups raise it — and “independent” is checked, not assumed. Groups whose walls turn out near-identical are flagged, and their agreement is reported rather than counted twice. In simulation the quarantine holds well past the point where an attacker can plausibly manufacture groups, and even when it is breached, capture plateaus instead of running away.
Each was added because a simulated adversary broke the version without it. The number beside each rule is what it recovered.
3×10⁻⁴wrong edges per decided question at 30% individual error
0.43 → 0.98driver recovery against three colluding insiders
zero votesstill enough to reach the map, with a source attached
5×worse recovery when cluster boundaries were allowed to count
80% / top-5the two thresholds, never interchanged
Every number on this page comes from simulations that ship with the platform and can be rerun in a minute. A claim we cannot reproduce is not on this page.
A simulated organisation with twenty problems, five of which really are the deep causes. Fifteen people work through it, two hundred times per row. Because it is a simulation we know the right answer, so we can count how many of those five each method ends up with — the only honest way to score a process, since no real room has an answer key.
| The room ranks by what hurts today | Stop at the vote | Carry on and build the map | Map, with seats held for evidence |
|---|---|---|---|
| not at all | 3.1 of 5 | 2.7 of 5 | 2.7 of 5 |
| a little | 2.4 of 5 | 2.2 of 5 | 2.6 of 5 |
| mostly | 0.8 of 5 | 1.9 of 5 | 2.6 of 5 |
| almost entirely | none | 0.2 of 5 | 2.6 of 5 |
Follow the bottom row, where the room ranks purely by what it feels. Crowning the
top-voted items names none of the five real causes. Building the map on those same
top-voted items rescues 0.2, because almost nothing worth finding survived the vote to
begin with — a cause dropped at the ballot cannot be recovered by any later stage.
The last column is the fix, and it is the most important number here. Hold five of the
fifteen shortlist places back, let anyone claim one for an overlooked idea by attaching a
source, and recovery stops tracking the room's mood: 2.6 of 5 at every level of bias.
The cost is half a cause in the unbiased case, and the return is a floor that holds when a
room is at its worst.
Now read the first row honestly. When a room somehow ranks purely by root cause, all this machinery loses half a cause compared with just voting — hundreds of yes/no judgements introduce mistakes of their own. It is insurance: you pay a premium in the easy case and collect in the hard one. If your group genuinely has no blind spot, save yourself the afternoon.
There is a cliff, and it is human. When people get roughly one in five of the influence questions wrong, the method still finds 3.3 of the 5 causes. At one in three wrong, the result is no better than guessing: a fifteen-person majority can correct scattered mistakes, not a room that is mostly mistaken. This is why one question is put to a whole panel at a time, rather than fifty questions to one person.
Some answers come free. If the group has agreed that A drives B and B drives C, nobody needs to be asked about A and C — it already follows. That shortcut removes 5–24% of the questions, and across 800 paired runs it cost no accuracy at all.
Duplicates are an attack, even by accident. Let one popular complaint sit on the wall as three differently worded cards and it eats three times the attention. In simulation that halved the number of real causes reaching the map, from 2.6 of 5 down to 1.2. Merging duplicates is not tidiness; it protects the ballot.
Contested pairs carry two different disputes. The platform separates them, because only one of them has a mechanical answer.
One word moves. The two premises stay, the two conclusions stay, and the argument goes from forced to fallacious. Groups argue about this for hours without noticing which one they are holding.
What the machine may decide. Whether a conclusion follows — that is computation, and it costs no votes. Whether a premise is true is never the machine's call. It goes to a panel, with the evidence attached.
The method is old and the mechanism is measured. What is new is that both now run at the size of a city rather than a seminar room — which changes what can honestly be attempted.
A municipality. Not a consultation that produces a list of complaints ranked by how loudly they were made, but a map showing which of those complaints cause the others — with the disagreements published rather than smoothed away, and every connection traceable to how it was settled. A council that acts on the roots can say why it chose them, and a resident who disagrees can point at exactly which question they would have answered differently.
A workplace or a union. Fifteen people in a warehouse already know why the picking errors happen; the knowledge is distributed across shifts that never meet. This assembles it without anyone having to be in a room at the same time, and the reserved seats mean the night-shift observation that nobody upvoted still reaches the map if someone can point at the log.
A regulator or a profession. Where expertise is real but contested, the split between what is certified (empirical and causal claims, at 80% of votes cast) and what is only measured (importance, values) means a body can publish a causal account without pretending its priorities are facts.
A diaspora, or anyone the room excludes. The binding constraint on a hall is physical presence. Remove it and the people who could not travel, could not take the day off, or live three time zones away are in the same deliberation as everyone else — on one shared timer, with the same five stickers.
Two old devices, doing real work. Selection by lot: a deadlocked question goes to groups drawn at random from the rest, never to whoever volunteers. And equal standing: nothing here weights a vote by reputation, stake or seniority. Both are ancient. What is not is asynchrony — an assembly needed everyone in one place on one day, which is precisely the ceiling this removes.
Who counts is still a political choice. Every dialogue declares who may take part, and every map can be read against the variety of the room that produced it. The platform does not decide that question; it refuses to let it stay implicit.
What this does not do. It does not make decisions, allocate money, or bind anyone. It produces a causal account with its provenance attached, and an honest record of what a group could not agree on. Whether anyone acts on that remains a political question — which is the right place for it to remain.
The method is not new. The mechanism science around it, the priced adversary model, and the implementation are what we add.
Related systems, and the gap. Polis maps opinion clusters. Remesh and collective-dialogue work scale agreement finding. The Habermas Machine drafts group statements people endorse. All optimise for agreement. None of them certify a causal structure, price an adversary, or keep disagreement as a first-class result.
Open by design. The mechanism, its simulations and its conformance tests are published; a dialogue records which version of the rules produced it, so changing a threshold is a visible act rather than a silent commit.
One group and an afternoon is enough to run the whole thing. You will finish with a map, a list of roots, an honest record of what could not be agreed on, and every connection showing how it was settled — and, once more than one group is working the question, how many of them found it independently.