tos.watch

Get an email when a company changes what it does with your data. The newsletter is free at tos.watch. The code is Apache-2.0 (see LICENSE).

tos.watch watches about 40 services in five tracks: AI assistants, dating apps, things you type into (writing tools, keyboards, notes, translation, transcription, email), developer tools (telemetry, training on your code, credit-based pricing), and the broad consumer apps most people use every day. The exact list is pipeline/watchlist.json.

Not affiliated with any company it tracks.

What's here

• pipeline/: Python 3, stdlib only. For each tracked document it:
  • pulls the history from Open Terms Archive, or fetches the page itself when OTA has no declaration (pipeline/lib/direct.py);
  • diffs it paragraph by paragraph;
  • asks the Jev classifier a few narrow questions about each real change;
  • drops scraper flap and duplicate captures;
  • writes one alert per document version to alerts/.
• pipeline/questions.example.json: the questions Jev answers. It's a generic set that runs but isn't tuned; see "Questions" below.
• pipeline/draft_emails.py: turns each new alert into a Buttondown draft email. Nothing is ever sent automatically.
• site/build.py: builds the static site from alerts/:
  • a landing page with search;
  • a page per service and per change, with word-level diffs by @pierre/diffs, Apache-2.0;
  • an RSS feed.
• worker/: a Cloudflare Worker for signup, site requests, feature votes, and a mock on-demand check. Emails live only in Buttondown; D1 holds anonymous request counts.
• SELF_HOSTING.md: run your own copy for the documents you care about.
• ROADMAP.md: what's deliberately not built yet, and why.

What isn't here

tos.watch is open core. The public code runs end to end. What stays private is the judgment built on top of it:

• the tuned question set, and the hand labels it was checked against;
• the dataset-selection work;
• the alert history, with its classifier scores;
• the hosted service.

The alerts themselves are public on tos.watch.

Code here, data elsewhere

This repo is code only. Everything a run reads or writes, apart from pipeline/watchlist.json, lives in a data directory set by TOS_WATCH_DATA: alerts, the Jev cache, run state, snapshots, your own questions.json and the generated site. Its layout mirrors this repo (alerts/, cache/jev/, pipeline/state.json, site/…). Unset, the data directory is this repo, and all of it is gitignored. tos.watch keeps its data in a private repo; see pipeline/lib/paths.py.

scripts/check-public.sh runs in CI and fails if run data, generated pages, local paths or personal addresses are ever committed here.

Data sources and attribution

Policy history comes from Open Terms Archive collections, including:

• the Platform Governance Archive (ZeMKI, University of Bremen);
• the GenAI Governance Archive.

They're licensed ODC-By 1.0, which requires attribution for any public use. The site footer, the feed and every alert carry it. Alerts quote short excerpts cut to the part that changed, never whole documents.

The fonts are Public Sans and Source Serif 4, both SIL OFL 1.1. They're self-hosted, so the site makes no third-party requests. Their licences are in site/fonts/.

Running the pipeline

export OPENROUTER_API_KEY=...
python3 pipeline/run.py             # pull, diff, score, write alerts, rebuild the site
python3 pipeline/run.py --dry-run   # collect, extract and filter; skip scoring and alert writing
python3 pipeline/run.py --no-pull   # reuse OTA clones; HTTP sources are still collected

--dry-run is not an offline or read-only mode: source collection runs before the dry-run exit, so it can clone/pull OTA history, fetch HTTP sources and write source snapshots. --no-pull only skips OTA clone/pull; it does not skip the HTTP source collectors. --out also creates its alerts directory during a dry run.

Reruns are cheap. Jev answers are cached on disk by a hash of the change (cache/jev/), and each alert file is written once per document version. --since YYYY-MM-DD limits Jev spend on a large backfill.

Questions

pipeline/lib/jev.py loads its question set from pipeline/questions.json, falling back to pipeline/questions.example.json.

A question set has four parts:

• a version, which goes into the cache key, so bump it when you change wording;
• the questions themselves;
• the topic_questions that can trigger an alert;
• an alert_threshold.

The example set is a starting point. Before trusting a threshold, check your questions against changes you've labelled by hand.

Email drafts

export BUTTONDOWN_API_KEY=...
python3 pipeline/draft_emails.py             # one draft per new alert
python3 pipeline/draft_emails.py --dry-run   # print the drafts; no API call, no state written
python3 pipeline/test_draft_emails.py        # new alert -> one draft; rerun -> none

Without BUTTONDOWN_API_KEY it's a dry run that exits 0. The data directory's pipeline/drafts_state.json records which alerts already have drafts. The first authenticated run marks existing history as handled without drafting it; later runs draft new alerts once.

Site

(cd build && npm ci)   # locked fonts and diff-renderer dependencies
python3 site/build.py       # alert pages, service pages, RSS feed

Python builds the site without Node dependencies, keeping a plain diff fallback. Installing build/ dependencies enables @pierre/diffs rendering.

以上为 README 的开头部分。其余内容请阅读原文。 查看完整 README ↗