How The Ledger Works

Where I source the news, what the AI does, what I check by hand, and what stays private

The Holliston Ledger is one person, a spreadsheet and a few programs I wrote myself. Since I utilize AI, Machine Learning, and other automated software for a good share of the first-draft work, It’s important to me for you to be able to see where and how these tools are used, so you can decide for yourself how much weight to give what arrives in your inbox.

Where the news comes from

Twice a week, a simple Python program reads the town’s website (its news flashes, alerts, posted agendas and calendar) along with the Holliston Reporter’s feed and collects anything new. For each posted agenda, the software opens the agenda to find which committee is meeting and what’s on the docket. Meeting recordings come from the same place, the town’s Agenda Center, where boards post their video and audio, often weeks before the minutes are approved.

What the AI does

For each new item, an AI model reads the source and writes up to three possible headlines, each paired with the sentence from the source that supports it, and a second, smaller model reads each headline against the source and flags anything the source doesn’t say. After that, ordinary Python code checks that every number, vote count and quotation in a headline appears in the source. All of this is dropped off into a Google spreadsheet, including the items the program decided to skip, so I can see what it left out as well as what it kept.

What I do

I read every row in that Google spreadsheet, open the link, and pick a headline, rewrite it or drop the item. Nothing reaches Substack unless I’ve approved it and vetted it myself and none of these programs can publish anything by themselves. The model also suggests which item should lead The Briefing, and that call is mine too.

Meetings

For the boards that make the large decisions, like the Select Board, the Finance Committee and the Planning Board, a program downloads the town’s recording and transcribes it overnight on my own computer with Whisper (an open-source speech-recognition program) using a list of Holliston names and streets so it writes “Hiawatha Trail” and “Eversource” instead of guessing. It then grabs every motion, vote and dollar figure and places it into a checklist with the timestamp in the recording where each one was said. A model then drafts a summary from the transcript, with each fact tied to its time in the recording. I then check those facts against the video before any of it runs, with a preference given to the official meeting minutes if they have been posted.

The transcripts themselves stay private because speech recognition still gets names wrong and can’t always tell one speaker from another (the automatic captions for one recent meeting spelled a resident’s name four different ways). I’m not willing to publish a record I can’t stand behind word for word.

What stays private until I’ve asked

When something in the record looks like a mistake or a conflict, it goes on a private list of questions first, and it appears in the Ledger only after I’ve asked the office or committee responsible, along with their answer.

When I get it wrong (and I will)

Corrections go at the bottom of the corrected post and in the next This Week at Town Hall, and every item links to its source so you can check my work against the record. A ★ marks items I reported from the public record myself; everything else links to reporting by the Holliston Reporter, HCAT or the town. If you see something wrong, or know something the record doesn’t, reply to any email, and it comes straight to me.