The AI Assistant

A chat box sits at the top of the right-hand column on the record pages. Ask about what is on the screen and it answers from that record and the movement's published data, naming what it read. Ask what it thinks of a contested policy question and it says it does not make political decisions, then hands you the place where yours gets written down.

The assistant reads three fixed inputs, answers under rules it cannot talk its way out of, and hands back a citation and a button whose address the product resolves, while the next-step suggestion beside it is computed in ordinary code with no model call at all.

What it does

The assistant is handed three things and reads nothing else: the record open on the screen, the committed public data behind it, and the published explanation of how a number on the page was computed. It does not browse, call outside services, or download media. It writes nothing — no votes, no proposals, no commits.

It obeys the rules every other surface obeys. The Wall binds it: anything it says that came out of a transcript belongs to the Flamingo Award and never to a voting-record award. It says a politician earned a percentage of an award, and the percentage is of the award rather than of the person.

Neutral about opinions is not neutral about the record. It will tell you how someone voted on final passage and link the roll call it read; it will tell you the vote is one of the acts feeding a Monkey Award and show the arithmetic; it will not tell you the vote was right. Asked what it thinks, it says it does not make political decisions, that the movement is here to hear yours, and points to where yours gets recorded. Withholding a published number out of caution is the opposite failure, and counts as a defect in our prompt rather than as safety.

Most answers end with a button, and the model never authors an address. It picks a key out of a catalog the product hands it, and the product resolves the key into a route. A build test checks those destinations against the real pages on this site — the first draft of that map pointed at seven pages that do not exist, and the test caught them first.

Beside the chat sits a next-step suggestion with no model call in it. Ordinary code walks one published ladder: tell us where you vote, record a position, verify your identity, vote on a decision, give your representatives stars, log a real-world action, put a challenger forward. Your cursor is the first rung you have not finished. Rungs get ruled out — done four hours ago, suggested yesterday, declined — and each exclusion returns its reason and its expiry. Nothing is random, and when everything is ruled out it says nothing.

What the app knows about your journey is one small file you can read verbatim, clear, switch off, or export. On your own machine it never leaves the box, and it never reaches the public repository, which carries no citizen identity.

The goal it serves

A movement whose thesis is that institutions stopped representing citizens cannot then answer a citizen's policy question with its own opinion. The moment it does, it is one more institution telling people what to think. The neutrality here is a product rule, not a legal hedge — it is the problem this movement exists to name, applied to ourselves.

The work this product asks of you is to write your own values down in a form software can apply consistently, and to record where you stand so it can be checked against the social contract and pooled with other citizens. An assistant with opinions would quietly do that work for you and hand back its own conclusion wearing your name.

Arithmetic over the public record is not an opinion, and withholding a published number because the subject feels political is its own kind of editorializing. So it states the vote, states the award, shows the derivation, and stops. The hard rules the build enforces keep those two behaviours apart.

What keeps it honest

• It ships off. The chat sits behind an install-wide switch set off on every install, a citizen's own laptop included. Off means the box is not rendered and the backend refuses the turn — a switch that only hides a button is not a switch.
• It cites, or it says it cannot. Where the record does not answer the question, it names which link is missing and why, instead of filling the gap from training data.
• It never authors a destination. Buttons and site links resolve through catalogs the product owns, and an unknown key is dropped and logged by name.
• It never adapts to you. It does not guess your politics from your state, your reading history, or which of the four doors you came through — the same question gets the same answer at every door.
• There is an offline-only mode that answers from a reviewed file of expected exchanges and calls no model at all.
• The suggestion is inspectable. The rule that fired and every candidate ruled out are rendered back to you. A recommender that quietly drops candidates is one nobody can fix.

Works with

• The First-Principles AI Check — the reasoning pass that checks whether a chain of argument holds; the assistant explains its output and never substitutes its own.
• Value Systems And Policy Packets — the place your rules actually get written down, which is where the assistant sends you instead of answering for you.
• Decisions, Areas And Communities — the surface the neutrality answer hands you, and the second rung on the recommendation ladder.

Where to go next

• AI — the whole picture of what the AI is allowed to touch here.
• What the AI never decides — the judgements that stay with citizens.
• The core loop — the sequence the assistant walks beside you.
• Do the math yourself — the arithmetic it will show you, laid out so you never have to ask.
• Take Action — the things a citizen can do this week, which is where the ladder points.