lazywatch: 68 page checks, 5 model calls, 0 bylaws read by me

Описание и применение

Автоматический мониторинг правительственных страниц о краткосрочной аренде в Канаде с использованием локальной модели Gemma 3 4B, которая активируется только при изменении текста, выдавая проверенные предложения и цитаты.

Задача

Keeping up with infrequent changes to government short‑term rental rules without manual browsing or excessive API calls.

Результат

Over a weekend, 68 page checks were performed with only 5 model invocations. Four of the five generated summaries were verified; one was discarded by the string‑match rule. No pages were manually read.

Рабочий процесс

• 1. Daily fetch each monitored page via HTTP or Browserbase if needed.
• 2. Compare fetched text to yesterday’s copy; if unchanged, skip.
• 3. If changed, extract only the changed paragraphs.
• 4. Run Gemma 3 4B locally on the changed text to generate a one‑sentence summary and a verbatim quote.
• 5. Verify the quote appears exactly in the new text; if not, discard the summary and output the raw diff marked NOT verified.
• 6. Aggregate results into a weekly digest.
• 7. Store page snapshots and diffs in Tiger Cloud Postgres with pgvector for future reference.

Доказательства

Once a day it fetches each page and compares the text to yesterday's copy. If nothing changed, that's it. Done. No model, no tokens, no reading.

Источник

Gemma gets the changed paragraphs, nothing else, and writes one sentence plus a quote.

Источник

Gemma writes the sentence. A string match decides if it stays.

Источник

68 page checks, 5 wakes, 4 of 5 quotes verified, 1 thrown out by the string match, 0 pages read by me

Источник