← the record

Our AI Constitution

Making a tool once required permission — a degree, a department, a budget line. AI collapses the distance between wishing a tool existed and building it: anyone who understands a problem can now build the thing that meets it. Toolmaking, the oldest civic craft, becomes available to everyone. But that potential won't bend toward the public good on its own — not while the companies building AI are the only ones deciding what it's for. This page is what deciding for ourselves looks like: every place a model touches this record, named; every promise, checkable.

part of the Community AI Project A live experiment in taking hold of AI and pointing it at civic life — open-source, locally-owned tools, deployed in real neighborhoods, so communities shape their own future rather than rent it. Read the whole argument at communityai.studio →

the articles — what we promise, and how to check us

The record is the source. Models summarize, find, and label; they never replace the transcript. Every AI-made line sits beside the record it came from, and points back to it.
check it yourself
Open any meeting: the summary is one labeled paragraph above the complete transcript, and every moment deep-links a timestamp on the tape.
Disclosure at the point of use. Anything a model wrote is labeled where you read it — “AI summary” on the page, ai:<model> in the data. Nothing AI-made is passed off as the record.
check it yourself
Every meeting's own plane (meetings/<id>.json) carries a summary_origin field: ai:<model> when a model drafted it, extractive when the sentences came from the transcript itself.
Reading never requires AI. Meaning search is an upgrade, not a dependency: when the model is off, the static keyword index answers and the page says so. Every page reads with every server gone.
check it yourself
Search the record, then search it again offline — the results say which index answered. The covenant holds the larger promise: static files only.
Readers are never fed to models. The models see public record text — transcripts, agendas, minutes — and, for search, the words of your query alone. Never anything about you: no account, cookie, or analytics exists here to share.
check it yourself
The covenant is enforced by a strict Content-Security-Policy on every page — no third-party script or beacon can load, and the search call carries no identity.
People gate the record. Nothing enters on a model's say-so: a steward approves every submitted meeting; corrections annotate, never rewrite. AI drafts; people decide.
check it yourself
Submit a meeting from the add page — the reply says a steward reviews it before the record updates.
Local first where it counts. The desk tools run their models on your machine — noise reduction, voice isolation, rotoscoping — and nothing leaves the room. The hosted record uses a cloud model only where static files structurally cannot (embedding a search query at the moment you ask it), and names the model it uses.
check it yourself
Install the desk from the press — then unplug the network and cut a film. It all still works.
Open to audit. The code that pressed this page — prompts, gates, fallbacks — is public under AGPL-3.0, a license whose purpose is that a public thing stays public. A constitution nobody can audit is a press release.
check it yourself
The source is at github.com/amateurmenace/control-z; which license covers which part is in LICENSING.md.
Built with AI, signed by people. This software is itself developed with AI assistance — the same collapse of distance we argue for is how a tiny shop built a record this size. Every change lands in a public commit a person reviewed and signed.
check it yourself
The commit history is public — the AI-assisted work is co-authored in the open, not laundered.

where the models sit

the ledger — when, whose, where, and what stands without it

where you meet it	what the model does	whose model	where it runs	without it
meaning search	turns your query and the record's segments into vectors so related passages match	Google `gemini-embedding-001`	our server, at the moment you search; it sees the query text and nothing about you	the static keyword index answers, labeled “lexical”
meeting summaries	drafts one paragraph from the transcript	Google Gemini (Flash) — labeled `ai:<model>` in the data	our pipeline, at press time, over public transcript text	an extractive summary — sentences drawn from the transcript itself, labeled `extractive`
issue names & labels	suggests a plain name for a thread that spans meetings	the same Gemini lane, labeled the same way	our pipeline, at press time	a keyword-derived name
transcripts	speech-to-text, only when a tape arrives with no official captions	an open Whisper-family model	our own hardware — the audio is never sent to a third party	official captions are always preferred; a tape with neither ships as a document, not a guess
moments · reels · kits · your paper · the front page	no model at all — open keyword rules and extraction, readable in the source		your browser, or press time	they are the fallback
the desk — civicmedia.studio / Control-Z	noise reduction, voice isolation, rotoscoping, film emulation	open models, bundled with the app	your machine only — offline works; no account, no cloud, nothing leaves the room	they are the tool

use AI that stays local

The record is the reading half; the making half is a desk of free, open-source tools whose models run entirely on your computer. Community media stations, filmmakers, and neighbors finish real work with them — denoise a meeting tape, isolate a voice, cut a reel into a film — with the network cable out of the wall. That is what “locally-owned AI” means here: not a slogan, a download. Get the desk from the press →

understand it deeper

communityai.studio — the program's whole argument: why civic AI, why open source, why locally owned.
the covenant — the promises this site keeps about you, enforced in the browser.
the source — every prompt, gate, and fallback named on this page, in the open.
Elements of AI — a free, plain-language university course; the best first step we know.
But what is a GPT? — 3Blue1Brown's visual walk through how language models actually work.
EFF on AI — the civil-liberties view: rights, risks, and policy.

This page is made to be shared — the direct link is publicrecord.studio/constitution. This edition was pressed 2026-06-18 from a corpus fingerprinted 66c05a9544c70585. When our use of AI changes, this page changes in the same commit.