- Identifier
- TC-P-002
- Version
- 1.0
- Date
- Author
- Rudolf Malle, Kirabi Works, Inc.
Summary
AI sorts, summarises, screens, matches and drafts. A person approves anything that is shown to residents as a reference, and AI never decides anything about a person.
Where AI runs
| Where | What it sees | What it produces | Who checks |
|---|---|---|---|
| News sorting | Title and excerpt | Relevance score, topics, functions | Threshold in config; admin can hide or show |
| News summaries | The publisher's page text, capped | A 2 to 3 sentence summary and one why-it-matters line | Shown with an AI summary tag; admin can edit, remove or redo |
| Post screening | The text of a question, answer or comment | Ok or hold, with a fixed reason | Held posts wait for City Hall, never auto-removed |
| Question matching | The new question and up to 5 candidate card titles and answers | One card id or none, with a confidence | Used only above the configured threshold; the resident can still post |
| Card drafts | The question, the shown answers and their links | A draft card or none | City Hall approves, edits or dismisses |
| Poll drafts | Aggregate counts and public titles only | Five poll questions | Admin approves |
What AI never sees
- Emails.
- Claim tokens.
- The full text written by residents for any purpose other than screening that text itself.
- Cities with fewer than 5 residents, in poll drafting.
What AI never does
- Publish anything to residents.
- Decide who is a resident.
- Score or rank people.
- Send email.
- Issue or deny any Passport status.
Guards
Every output is validated against fixed rules (lengths, options, no dashes, allowed values) and dropped if it fails. There are daily caps on every call type and no retry loops. Gateway failures fall back to the publisher's text or to a human queue.
Reporting an error
Every card and summary has Report. City Hall reviews within one business day; corrections carry a new checked-on date.
Change log for prompts
- October 14, 2026: prompts for sorting, summaries, screening, matching, card drafts and poll drafts recorded as version 1.0.
Limitations
Summaries are written from the publisher's page at the time of fetch and can lag the publisher's own corrections. The models used are third-party services reached through a gateway; we do not train models on resident content.
Changes
- October 14, 2026: Version 1.0 published.
Cite this
Malle, R. (2026). How we use AI in Trial City (TC-P-002, version 1.0). Kirabi Works, Inc.. https://trialcity.app/papers/how-we-use-ai