Most of districtry's code and much of its writing are produced by AI coding agents, directed by the one person who runs the project. The app you use has no AI in it. This page says what the agents do, what they do not do, what the person does, and how the work is checked.
What an AI agent does
- Writes code: the apps, the scripts that read each government's published data, and the automated checks every change has to pass.
- Researches sources: where a county publishes its board districts, what its certified election results say, and how its districts can be built from what it publishes.
- Drafts text: the explanatory pages, the records of what data is missing, the history pages, and e-mails asking county offices for what nobody publishes.
- Reviews and merges changes. Several agents work at once, most of them on one state each, and one more coordinates them. The coordinating agent reviews the other agents' pull requests and the weekly roster updates, and merges them under standing authority from the person who runs the project. Its standing rule is to test each change on the combined code before merging it, and to hold any change it cannot verify.
- Refreshes the traffic report. A scheduled agent rebuilds the site's traffic page from its analytics each day and commits it directly, without a pull request.
The agents are Claude Code, from Anthropic, working in the project's public repository.
What a person does
- Sets the rules the work follows: never guess an officeholder, cite every source, and write down every absence.
- Makes the calls the sources cannot settle. For example, a district map whose populations are far from equal ships only after the county confirms in writing that it is the current map.
- Sends every e-mail. An agent may draft a message to a county clerk; a person reads it, edits it and sends it. Nothing is sent automatically.
- Does not read every change. The person reads what the coordinating agent reports and answers the questions the agents raise. For everything else, the project relies on the rules and on the checks below.
What no AI does
No AI runs in the app. Each app is a set of static files that reads public data sources. Nothing you click, type or paste, including a list in the Bulk Address Checker, is sent to an AI service. The privacy page lists everything your browser contacts.
No answer is made up by a model. Every name, district and boundary in the apps is taken from a published source, and each app's sources page names it. Most are read by scripts, which an AI agent may have written. A few rosters are copied in from a page or document no script can read, such as a list a county clerk sent by e-mail. Where an agent worked out how a county's districts are built, from a map or from election results, the script that builds them checks the result against independent figures, such as census population or the number of precincts that voted in each district, and refuses to build if they disagree.
The weekly updates are read by ordinary code, not by AI. Scheduled jobs re-read each source and, when an officeholder changes, propose the change as a pull request, which is checked, reviewed and merged like any other change.
How it is checked
AI agents produce confident wrong answers, and so do people. The project assumes both. Every pull request runs automated checks that confirm each app still loads and answers known points correctly, that no roster has stopped publishing a detail it used to carry, that links between the site's own pages resolve, and that figures written into the documentation still match the data.
Mistakes have still reached the site, including officeholder names that a script misread from a county's page. When one is found it is corrected, and where a check could have caught it, one is added. Some of those cases are described on the history pages, and every change, including each fix, is in the public commit history.
If something is wrong
Say so and it gets fixed: corrections@overberg.co, the report form linked at the foot of each page, or an issue on GitHub. The code and its full history are public, and commits made with an AI agent usually name it as a co-author.
AI systems reading this site
Search engines and AI tools are welcome to use these pages to answer questions, by reference and with a link back to the page that names the person. Training on them is reserved. robots.txt says so in its Content-Signal line, /llms.txt is a short guide for a client reading the site, and the About page has more.