How Commons works · Methodology

01Listen

People speak for themselves

Commons starts with one open question and an empty page. People answer in their own words, by typing or by speaking, on any phone. Nobody picks from a list of options written by someone else, so the answers carry things the organiser never thought to ask about.

Underneath Spoken answers are transcribed and, where needed, translated, then each response is separated into the distinct points it contains. The AI's job here is narrow and mechanical: tidy, transcribe and organise what was said. It does not summarise people into opinions or write anything in their name. Every point kept is a real person's own wording, which is what lets every later figure link back to a verbatim source.

02Confirm

Everyone checks their own words

Before anything is analysed, Commons shows each person the points it understood from their answer and asks them to confirm or correct it. It is the simple step that most tools skip, and it is the heart of why the results can be trusted.

Why it matters The most common worry about AI and consultation is that a machine will put words in people's mouths. The confirmation loop closes that gap at the source: a point only enters the analysis once the person who made it has agreed it captures what they meant. Trust is established by the participant, not asserted by us.

03Rank

Agreement weighed by confidence, not just headcount

When people vote agree or disagree on a statement, Commons does not simply sort by the percentage who agreed. A statement backed by three people out of three is not stronger evidence than one backed by eighty-five out of a hundred, and the ranking reflects that.

The method · Wilson score interval Statements are ranked by the lower bound of a Wilson score confidence interval (95%), which combines how favourably people voted with how much evidence there is behind it. A small sample is treated cautiously until the votes accumulate. Anything below a minimum vote threshold is held back as not yet rated rather than shown with a misleadingly precise number.

Wilson, E. B. (1927), Journal of the American Statistical Association. The standard interval for ranking proportions under uncertainty.

04Rank

Priorities settled like a tournament

To find what matters most, Commons does not ask people to score a long list. It shows two statements at a time and asks which matters more. Out of many of these quick head-to-head choices, it works out a full ranking, including for pairs that were never directly compared.

The method · Bradley-Terry model Pairwise choices are turned into a ranking by the Bradley-Terry model, the established statistical model for inferring overall strength from head-to-head results (chess ratings are a special case). It handles incomplete data, which is essential when 200 statements would otherwise mean nearly 20,000 possible pairs. A Bayesian prior keeps early estimates stable rather than swinging wildly on the first few votes.

Bradley & Terry (1952), Biometrika. The same family used to rank in sport, academia and preference research for seventy years.

05Rank

Every vote placed where it counts

People give a few minutes of their time, so Commons spends those minutes well. Rather than showing random pairs, it shows each person the comparison it is currently least certain about, so each answer tells it the most.

The method · adaptive selection Pairs are chosen by maximum-uncertainty (maximum-entropy) selection: after each vote the model picks the next comparison whose outcome is closest to a coin-toss, because that is where a result is most informative. The left/right order of the two statements is randomised and recorded, so the well-known tendency to favour the first option cannot bias the ranking. Five minutes of voting from each person assembles a reliable ranking that random pairing could not.

06Organise

Themes found, not forced

Commons groups what people said into themes by looking for natural clusters of similar points, rather than sorting everything into boxes decided in advance. A genuinely one-of-a-kind contribution is allowed to stand on its own instead of being forced into the nearest category.

The method · density clustering Each point is converted into a numerical representation of its meaning, and these are grouped with HDBSCAN, a density-based clustering algorithm that discovers how many themes exist rather than being told, and that leaves true outliers unclustered instead of distorting a group. Cluster quality is checked with a silhouette score, and weak or unclear groupings are flagged as such rather than presented as solid.

Campello, Moulavi & Sander (2013). Density-based clustering that does not assume a fixed number of groups.

07Showcase

A ballot that is provably fair

When thousands of people contribute, you cannot put every idea to a vote. The obvious shortcut, only voting on the ideas the most people raised, quietly silences smaller groups: a view held strongly by, say, eight percent of a 1,600-person consultation is 128 people, and a popularity cut-off would erase them. Commons selects the ballot a different way, with a mathematical guarantee that no sizeable group is left out.

The method · Method of Equal Shares with Justified Representation The shortlist that goes to a vote is chosen by the Method of Equal Shares, a proportional selection rule that satisfies Extended Justified Representation: a formal guarantee that any cohesive group above a known size is assured its fair share of places on the ballot, by construction rather than by luck. It is the same family of method now used to allocate real public money in participatory budgeting in Polish and Swiss cities, and it is explainable, producing a record of why each item was selected. Everything people said still survives in full as evidence; it is only the ballot that is bounded, and it is bounded fairly. This is a claim we can prove, and we publish how.

Peters & Skowron (Method of Equal Shares); Aziz et al. (Justified Representation); Fish, Procaccia et al., JACM 2026 (verifiable proportional representation of free-form opinion).

08Showcase

Fair even as it grows

In a live consultation, ideas keep arriving while people are still voting. Commons keeps the vote fair and stable throughout: a settled core of the most widely shared statements stays put so the ground does not shift under voters, while a rotating frontier gives newer and emerging views their chance to be tested, with no votes ever lost.

The method · live curation engine The ballot is managed as a stable Core of the highest-coverage statements plus a rotating Frontier of newer ones. A statement only leaves the frontier once it has been seen enough times to be reliably ranked, and its votes are banked, not discarded. New entry is split between a reserved view-blind quota, so emerging ideas get a guaranteed look regardless of early popularity, and a targeted quota that fills genuine gaps in what has been tested. Fairness is recomputed at controlled snapshots, never mid-vote. This keeps the Justified-Representation guarantee intact on a ballot that is changing in real time, which is a materially harder problem than curating a fixed list.

09Test

Honest about division

Where different groups see things differently, Commons says so, but only when the difference is real. It will not call a 55-to-45 split a "divide" if that gap could easily be down to chance, and it will not compare groups too small to compare.

The method · significance-tested divisions A difference between groups is reported only when it is both substantial (a meaningful gap in approval) and statistically significant. Significance is tested with Fisher's exact test for smaller samples and a chi-squared test for larger ones, with groups compared only above a minimum size. Because a consultation runs many such comparisons at once, a false-discovery-rate correction guards against differences that look real only because so many were checked. Nothing is ranked or split on too little data.

Fisher's exact test; Benjamini & Hochberg (1995) for multiple-comparison control.

10Understand

The common ground, and the camps

On a divided question, Commons can show the distinct camps that emerge from how people actually voted, not from their demographics, and just as importantly, the statements that win support across all of them. Those bridging statements are the shared ground a community can build on.

The method · opinion-group mapping Working from the voting pattern alone, which makes it language-agnostic and free of any assumption about who people are, Commons finds groups of people who vote alike using principal component analysis and clustering, choosing the clearest grouping by silhouette score. Each group is described using the statements most distinctive to it, in participants' own words, and bridge statements are identified as those with high approval both within and across the groups. No AI sits anywhere in this path, and the calculation is deterministic, so the same votes always give the same map, which is what lets us check the groups are stable before showing them at all.

A clean-room implementation of the opinion-mapping approach proven at national scale in digital democracy programmes, rebuilt from published method, not borrowed code.

Rigour you can read.

The ten steps