Skip to content

Voice

Voice is a privacy-first civic participation service. A resident talks privately about a local problem, checks what Voice understood, decides exactly what to share and with whom, and later hears what was done about it. A representative sees shared, source-backed issues and records what they will do, with a date, an authority and an owner. Anyone can read and search the public issues without an account. A participant can have their personal data erased, and the public issue stays. It is an ankka application, and it keeps its evidence in ankka-reasoning.

This is a demonstrator for a pilot: synthetic data, invited participants, one area.

The seven things it does

  1. Hear. A participant starts a session with an invite code and no name, email or verified identity. They talk to Voice about a concern. Voice proposes a structured reading of what they said: the problem, where, when, who it affects and how. The participant edits it, confirms it or discards it. Nothing they said leaves the conversation at this point.
  2. Share. For a confirmed concern, the participant writes or approves the exact text to share and picks one of four modes: private, de-identified, for representatives only, or publicly under a name they choose. A reviewer checks the text, with an automatic privacy check that flags distinctive circumstances, health, minors and small groups. An approved text joins a shared public issue.
  3. Understand. An issue links evidence from cited sources, each with its state: candidate, accepted, rejected or superseded. Questions show their competing explanations, and what is not known is said.
  4. Deliberate. Participants say how much an issue matters to them and which trade-offs they would accept. These are opinions, shown apart from evidence, as counts with their method and a caveat. A count below the suppression threshold is not shown.
  5. Act. A representative opens an investigation, records a dated action with its authority, owner and rationale, and later a decision and an outcome. Each is public, with who recorded it.
  6. Return. A participant who follows an issue gets one update in their inbox for each material development: what was recorded, why they are told, and links to the record. The update never says a problem is solved unless the record does.
  7. Erase. A participant can ask for their personal data to be erased. Their key is destroyed, so every protected field becomes unreadable everywhere; their private records are deleted; a de-identified text they published stays on the issue with no link to them. A participant with no session for the dormancy period is erased the same way.

What it is not

It is not a poll, a petition or a vote. It does not say how many people agree, weight responses or infer demographics. It sends nothing in a representative's name and decides nothing. The AI in it replies, proposes and suggests; people confirm, approve, record and decide. The full list is on Limitations.

The example

Every page uses one example: the issue "Rural school bus withdrawal", from the seed pack the repository ships. The Route 14 school bus between Kingsway and Northfield was withdrawn at the end of March. The council's route review says it carried eleven pupils a day; the parents' survey says twenty-three. Both are accepted evidence on the hypothesis that the route was withdrawn because too few pupils used it, one supporting it and one opposing it, so a reader sees the contradiction. A second question, whether the ring road safety works will be finished before the new school term, has no evidence, so Voice abstains. Every source in the pack is fictional and says so.

Where to start

For models and agents

This documentation is published in forms a model can read directly. llms.txt at the site root lists every page with a one-sentence description, llms-full.txt holds every page in one file, and each page is also served as Markdown at its own path with a .md suffix. The same pages are rendered as an agent skill, voice, in the repository.