about

Small municipalities don’t have the resources to surface information and analysis that lets both officials and constituents really understand the issues in front of them. Budgets are published, agendas are posted, AFRs are filed — but the work of pulling those threads together into a single picture you can interrogate is a full-time job nobody has the budget for.

Red Hook Watch is an attempt to address that. The aim is to make it possible — for a resident, a trustee, or anyone reading the news — to follow a specific number from the headline back to its line item in the published budget, and to ask honest questions like is the sewer fund self-supporting? or are the Town and Tivoli paying their fair share for police? without doing the spreadsheet archaeology themselves.

• Objectivity is the goal.Where the data views describe the public record — budgets, meetings, allocations — we try to describe what’s there, not what we’d like to be there. Corrections are welcome, in public, on the GitHub repo.
• Where proposals are presented, we try to present and analyze all proposals.If three options are on the table for a sewer-fund repayment plan, we lay out all three with the same level of detail, not just the one we’d pick.
• Some bias in which issues we analyze is unavoidable.A two-person side project can’t cover everything; what gets attention is partly a matter of where we notice questions worth asking. We actively welcome input on what we’re missing — if a topic deserves analysis, tell us.
• Every figure traces to a source document. Pages link the PDFs, OSC filings, and meeting minutes the analysis was built from. The pipeline that turns those documents into the JSON the site reads is open source and reproducible end-to-end.

We use large language models to do something that would otherwise be flatly impossible at this scale: ingest 20 years of meeting minutes, OCR scanned PDFs, organize the data into typed records, and produce structured analysis across thousands of source documents.

The analysis itself is human-driven— the questions to ask, the methodology to follow, the framing of the findings. AI does the work that makes this a side project two people can run on weekends rather than a team of full-time analysts: extraction, normalization, summarization, and routine arithmetic at corpus scale.

We also lean on LLMs to push toward objectivity. Asking a model to summarize a debate or describe a position tends to produce more neutral prose than writing the summary by hand, and we can audit the result against the source. Where AI outputs feed published numbers, the model name and extraction prompt are recorded with each record so reviewers can spot-check.

AI is fallible. Quotes can be mis-attributed; sums can be off; subtext can be missed. Every page is structured so a reader can click through to the source and verify. When something looks wrong, please tell us — the same model that wrote the summary can fix the prompt that produced it.

• Not an official government site.The Village and Town publish their own pages. This site republishes and analyzes what they’ve already made public; it doesn’t set policy or speak for them.
• Not an official publication of any office. Although one of the contributors (Frances Uku) sits on the Village Board of Trustees, this site is not a Village publication, isn’t produced as part of her duties as a Trustee, and doesn’t speak for the Board. It is a personal project. There are no advertisers, sponsors, or paid contributors.
• Not infallible. Both extraction and analysis can be wrong. Send a note via the contact page and we’ll fix it.

Red Hook Watch is a personal project. It is not a Village publication, isn’t produced as part of any sitting official’s duties, and isn’t a campaign vehicle. Frances’s role as a Trustee is disclosed because it’s relevant context for how some of the framing and topic selection comes about.

See an error, want analysis on a topic we haven’t covered, or have a question about how a number was derived? Use the form on the contact page and we’ll get back to you.