How it works
Suggest — share what you are hearing, in your own words. Type a message, record
a voice memo, upload a screenshot, or reply to the weekly WhatsApp prompt. Similar suggestions
are grouped by AI and the strongest themes rise to a shortlist.
Stay informed — opt in to a weekly digest and a weekly prompt delivered by email
and WhatsApp. The digest summarises what the community has been saying and what the live
consultations are finding. The prompt highlights the emerging themes and invites you to add more.
Research — search the knowledge library and the official record of Parliament
(Hansard) by topic, person and date; save useful finds for the whole team.
Consult — a human moderator promotes a shortlisted theme into a live consultation.
Participants vote agree / disagree / pass on short statements. The votes reveal common ground
statistically — nobody picks a side, the data finds it.
Act — every parliamentary question, council motion, or briefing that draws on
the exercise is logged with a traceable link back to the specific contributions that informed it.
Contribution → action, visible to everyone, permanently.
The journey of a topic
1 · Anyone suggests a topic
The front page invites anyone to suggest, in their own words, what the next public conversation
should be about — by typing, recording a voice memo, or uploading a screenshot. No form-filling
beyond a sentence or two: "Nobody can find an NHS dentist around here" is a perfectly
good suggestion. You sign in with your email address; only your first name is shown against your contributions.
2 · Similar suggestions form an emerging shortlist
Suggestions that point at the same underlying issue are grouped together by AI, which also drafts
a neutral title and framing question for each theme. The shortlist shows how many people raised
each theme, with every verbatim suggestion expandable underneath — the AI summarises, but never
hides, what people actually said. Themes strengthen visibly as more people raise the same issue.
3 · A human promotes a theme into a live consultation
AI drafts; people decide. A moderator or MP promotes a shortlisted theme, edits the title and
framing question, sets the dates, and writes the statements participants will vote on — with an
optional AI assist that drafts candidates from the public's suggestions, deliberately including
statements that different sides of the debate would each endorse. The suggestions that fed the
theme are permanently linked to the consultation they became.
4 · The live consultation runs on four connected layers
Open Listening — share lived experience in your own words. These accounts feed the statement
pool and give representatives concrete human stories rather than abstractions.
Common Ground — vote agree / disagree / pass on short statements. The votes place every
participant on an opinion map and statistically discover the opinion groups — nobody picks a side.
The payoff is two lists: statements both groups endorse, and statements that genuinely
divide them. This separates what communities actually agree on from what they fight about.
(The method — vote-matrix analysis and clustering — was pioneered by vTaiwan and Pol.is.)
Anyone can take part at the depth they have time and confidence for, and each step teaches the craft of
being useful — concrete beats general, single claims beat compound ones. Share an experience in a minute,
vote on statements, propose a statement of your own for the community to vote on (a moderator checks each
one first), or build an evidence brief — a structured claim-evidence-action submission that
moderators pass directly to representatives.
What It Led To — the action tracker. Every parliamentary question, council motion, campaign
briefing or published document that draws on the exercise is logged with its status and a traceable
list of the specific statements and experiences that informed it. This is the trust mechanism:
contribution → action, visible to everyone, permanently.
5 · Closed consultations remain public record
When a topic closes, voting stops but nothing disappears: the final opinion map, the common-ground
findings and the full action trail stay published — so the platform accumulates proof that
participating leads somewhere.
The toolkit around the conversations
Surveys — structured questionnaires alongside the open consultations, with one-click
sign-in links so invitees don't need to remember a password, live results, and PDF export.
Knowledge library — a shared, searchable collection of useful documents, good speeches,
successful campaigns, media appearances and case studies. It includes a full search of
Hansard, the official record of Parliament: filter by topic, by member, by date range
and by house across spoken contributions, written answers, debates and votes — including how
each member voted in a division, a timeline of when Parliament discussed a topic, and the
ability to save a search to re-run later or file a result into the library permanently.
Expertise — MPs tag their own areas of expertise, so anyone planning a project can find
the right colleague in seconds, and see what each MP has recently said in Parliament on the
subjects they know best. Visible to MPs and admins only.
Parliament context — every live consultation automatically shows recent Hansard debates
related to its topic, connecting what the public is saying here to what Parliament has said there.
The evidence layer
The pilot is built to answer a research question — does digital deliberation measurably repair
trust? Every action taken on the basis of the exercise is logged with a traceable link back to
the specific contributions that informed it. An evidence dashboard publishes the count of actions
with public provenance — the figures a funder, party or replicating institution needs. The methods
and code are open by design.
Under the bonnet — how it actually works
Finding the opinion groups. Every vote is stored as +1 agree, −1 disagree or 0 pass, forming a
matrix of participants against statements. Each statement's column is centred on its mean, then the
first two principal components are extracted by power iteration (the second deflated against the
first) to project every participant onto a 2-D plane — that projection is the opinion map.
Those coordinates are then clustered with k-means to separate the opinion groups. Nobody is assigned
to a group by demographic, party or self-description: the groups fall out of voting behaviour alone.
You need at least 3 votes to be placed on the map, and at least 4 mapped participants before a map is
drawn at all, so a handful of votes can't manufacture a pattern.
Why the same votes always give the same map. Both steps are deterministic — the power
iteration starts from a fixed vector rather than a random one, and k-means is initialised at the two
extremes of the first principal component rather than at random points. Re-run the analysis on the
same votes and you get an identical map, which is what makes the findings reproducible and auditable
rather than a black box that shifts each time you look at it.
Finding common ground. For each statement the agreement rate is calculated within each
opinion group, counting only votes actually cast — so a statement reads "80% agree" of the people who
expressed a view, never diluted by people who haven't seen it. A statement is only eligible to be
called consensus or divisive once each group has cast at least 3 votes on it. Consensus is
scored as the lowest agreement across the groups (so something only counts as common ground if
both sides back it), and divisiveness as the gap between them.
Statements arriving mid-conversation. The voting deck is rebuilt from whatever you personally
haven't voted on yet, so a newly approved statement is automatically put to people who had already
finished. A new statement starts with a nearly empty column, contributes almost no variance and so
barely moves the map until enough people have weighed in — it earns influence rather than being
handed it. A known limitation, stated plainly: the matrix cannot yet tell "hasn't voted on
this" apart from "passed on this", so someone who has voted on only a few statements is pulled toward
the centre of the map and can look more moderate than they are.
Grouping the suggestions. The emerging shortlist is built by giving an AI model the whole
suggestion set at once and asking it to group suggestions pointing at the same underlying issue, then
draft a neutral title and framing question for each. Short civic suggestions cluster poorly on text
similarity alone, so this is deliberately a language-model judgement rather than a maths one; if it
returns nothing usable, the system falls back to embedding the suggestions as vectors and grouping
them greedily by cosine similarity above 0.6. Results are cached until the suggestion set changes.
Which AI does what. Claude Haiku 4.5 handles the language work: grouping suggestions, drafting
titles, framing questions and candidate statements, writing the weekly digest, summarising Hansard and
library search results, matching projects to expertise, and reading text out of uploaded PDFs and
screenshots. Voice memos are transcribed by Whisper, and the fallback clustering uses a
sentence-embedding model — neither of those has a Claude equivalent. Every AI output is a draft:
a human edits and decides.
The stack. It runs entirely on Cloudflare's edge — a single Worker for the API, D1 (SQLite) for
data, R2 for uploaded files, and static assets served alongside. The frontend is plain HTML, CSS and
JavaScript with no framework and no build step; the project has zero runtime dependencies, so there is
no supply chain to audit beyond the platform itself. Sign-in uses email and password hashed with
PBKDF2-HMAC-SHA256, and sessions are signed tokens that can be revoked instantly for any account.
Requests to Parliament's Hansard API are proxied and cached server-side for an hour, so the app is a
considerate consumer of a free public service and never exposes your browser to a third party.
Where the AI is — and isn't
AI does a handful of narrow jobs: transcribing voice memos, extracting the key point from
screenshots and documents, grouping similar suggestions, drafting neutral titles and candidate
voting statements, writing the weekly digest, summarising library and Hansard search results,
and matching a project description to the colleagues who know the subject. It never decides
what proceeds, never moderates, and never replaces a human judgement — and the opinion mapping
is classical statistics, not generative AI at all. Every AI output is either editable by a human
before it takes effect or accompanied by the verbatim source material.
Privacy and openness
Taking part requires an email address to sign in — this is used only to let you return to your
contributions and receive the weekly digest, plus an optional WhatsApp number if you want the
prompts there. Only your first name is shown against your contributions;
your email address is never published. No other personal data is collected
beyond what you choose to write or say. Everything runs on infrastructure costing pennies at pilot
scale, which is itself part of the replicability story: the model is designed for other funders,
parties and civic institutions to adopt and scale.