RegWatch

Tells a small EU online shop which regulation changes apply to it and what to do by when, with the source sentence quoted under every to-do. Nothing is invented: a to-do is stored only if its quote is verbatim in the official page (regwatch/schemas.py, checked in code, never in a prompt).

uv sync --dev
uv run regwatch seed     # loads the recorded run in examples/seed-run (idempotent)
uv run regwatch serve    # http://localhost:8000  →  /month

Demo: https://port8000-jmda1596h1jt4g9.tunnel.applications.eu-west2.nebius.cloud (Nebius Serverless Endpoint regwatch-web-6, eu-west2, deployed 2026-10-27 with four example stores, the agent on and five EU laws; serves the recorded seed run; stays up through judging, Dec 1–15).

Runs on Nebius · NVIDIA Nemotron · Tavily

Piece	Role
Nebius Token Factory — nvidia/Nemotron-3_5-Lightning	document → obligations, each with a verbatim quote (regwatch/understand/normalize.py)
Nebius Token Factory — nvidia/Nemotron-3-Ultra-550b-a55b	(change, store profile) → applies / doesn't apply / unclear + grounded to-dos (regwatch/understand/classify.py)
Nebius Serverless Endpoint + Serverless Job	web app and the crawl, one container image (ops/)
Tavily search (advanced, include_domains, time_range) + extract (advanced, markdown)	the only ingestion path; one live crawl costs 40 Tavily credits (14 searches, 2 extracts; 2026-10-06) and 52 in the Nebius Job (2026-10-10, more new pages); 56 Tavily credits measured offline against recorded fixtures (24 searches over five regulations)

Only nine domains are ever searched or extracted (config.SOURCE_DOMAINS): eur-lex.europa.eu, commission.europa.eu, green-forum.ec.europa.eu, gesetze-im-internet.de, bmj.de, bmuv.de, ihk.de, help.shopify.com, shopify.com. The feed footer shows the credits and model cost of every crawl.

How it works

1. Search — regwatch/ingest/search.py: Tavily advanced search per (domain × regulation query), domain-filtered, deduped by URL.
2. Extract — regwatch/ingest/extract.py: Tavily advanced extract in batches of ≤ 20 URLs → Document (markdown + content hash).
3. Diff — regwatch/ingest/diff.py: new / changed / unchanged by content hash; unchanged documents cost no model call.
4. Normalize — regwatch/understand/normalize.py: Lightning turns a new document into instrument, dates and obligations with verbatim quotes.
5. Classify — regwatch/understand/classify.py: Ultra decides applies / doesn't / unclear for the store profile and drafts to-dos.
6. Store — regwatch/jobs/crawl.py: one CrawlRunRow per crawl with Tavily credits and model USD; regwatch crawl --stage all.

Why you can trust a to-do

• Quotes are checked in code: RegChange.validate_quotes() (quote is a substring of the full page) and Applicability.ungrounded_todos() (a to-do's quote is one of the obligations) must be empty before anything is stored — and classify_all re-checks at the store boundary.
• A model answer with an ungrounded quote gets exactly one retry; what is still ungrounded is dropped and logged, never shown. Citation URLs are set by code; a due date the model proposes is kept only if it is one of the change's effective dates, otherwise it is nulled.
• "Unclear" is a first-class answer: the page says which missing profile fact would decide it.
• Every change page has a "Report a wrong call" link (REGWATCH_REPORT_EMAIL).

What the app shows

• /profile — six fields (country, ship-to, categories, platform, company size, where the goods come from; the last two may be "not sure"); no account, no email.
• /feed — changes grouped by regulation with applies / doesn't / unclear pills; footer with credits and USD.
• /month — to-dos bucketed Already in force → Due this month → Coming up → No fixed date; check-off persists (htmx). / lands here.
• /change/<id> — why it applies in your profile's terms, to-dos with their quote and source link, obligations, source metadata.

Run it locally

No keys needed (recorded Tavily and model replies from tests/fixtures/):

uv sync --dev
uv run regwatch seed
uv run regwatch crawl --stage all --backend fixtures --llm-backend fixtures --no-out   # prints "… 56 Tavily credits …"
uv run regwatch serve

With keys: copy .env.example to .env, set TAVILY_API_KEY= and NEBIUS_API_KEY=, then uv run regwatch crawl (live search + extract + both models; writes nothing to examples/seed-run). The demo seed is recorded from such a crawl with uv run regwatch record-seed --url … [--exclude …] (grounded changes only; undated to-dos skipped; the crawl's Tavily credits carried into the footer). Schema changes are not migrated — delete a stale local regwatch.db first (the example stores' slug / name / blurb columns are the one exception: make_engine adds them to an older DB).

Example stores

The shared demo profile stays locked (REGWATCH_PROFILE_LOCKED=true). The demo URL is open to anyone with no login, and checking a new store costs about 35 Nemotron 3 Ultra calls (~$0.24), so entering your own store works only when you run RegWatch yourself. A row of example stores on the feed, month and profile pages lets a visitor switch between shops whose results were worked out ahead of time (a cookie, nothing written). Each is examples/seed-run/stores/<slug>.json; the default store (de-furniture, equal to profile.json) uses the change files' own results, every other store reads stores/<slug>.applicability.jsonl. uv run regwatch classify-presets fills that file: Nemotron 3 Ultra on each (store, seed change) pair that has no line yet, the same enforce() grounding and ungrounded_todos() check as the crawl, one appended line per result. A rerun only fills gaps, so after new change files are recorded it classifies just those; regwatch seed then loads them.

Let the agent check (unclear changes)

On a change marked unclear, a button starts an agent: Nemotron 3 Ultra picks one tool per step — search one allow-listed domain (Tavily basic depth), read a page it was shown, ask you a multiple-choice question, or decide — and each step is shown with its Tavily credits and model cost. Its to-dos pass the same enforce() grounding as the crawl, an evidence quote must be verbatim in a page it read, and a reason that cites no quote or answer of yours is dropped (an applies / does_not left without one becomes unclear). The result is shown only to the visitor who ran it; the shared seed rows are never written. Caps in code: 6 model steps and 10 Tavily credits per run, $1 of model tokens and 60 Tavily credits per UTC day (counted in the app's database, which resets when the Endpoint restarts) for all visitors together, 5 runs per visitor per hour. REGWATCH_AGENT_ENABLED=false (the default) hides it, and agent spend is not counted in the feed footer's credits per crawl.

Deploy on Nebius

The public URL serves the recorded run in examples/seed-run (recorded with regwatch record-seed from the sixth live crawl: 4 documents, 40 Tavily credits; re-classified on 2026-10-14 for a small company, 26 changes; plus 7 changes from 5 documents appended on 2026-10-15 with record-seed --append from a live crawl of the green-claims, packaging and product-safety acts: 43 Tavily credits; 33 changes in all); the Serverless Job runs the same image with the live crawl entrypoint against a freshly seeded database — Job #1 on 2026-10-10 re-ingested the 4 seed documents as unchanged and left them alone (52 new documents, 47 new changes, 52 Tavily credits, $0.2464 in model tokens; a second pass over one database adding nothing is verified offline by tests/test_crawl_all.py::test_crawl_all_twice_is_idempotent) — live results reach the URL by re-recording the seed and rebuilding the image, not by a shared database, because Serverless Endpoints have no persistent volume we could verify.

One image, two modes (ops/entrypoint.sh serve|crawl; crawl exits 2 without both keys and never uses fixtures). Console values: ops/serverless-endpoint.yaml, ops/serverless-job.yaml; step-by-step: docs/runbook-nebius.md. A daily schedule exists as .github/workflows/daily-crawl.yml (Jobs have no cron field) but is switched off.

Costs and measurement

Live figures come from the sixth live crawl (2026-10-06: seed + crawl --stage all on a fresh database), priced at the Token Factory catalog prices checked 2026-10-10 (Lightning $0.06 / $0.24, Ultra $1.00 / $3.00 per Mtok). Offline figures are fixture replays and are labelled as such.

Item	Figure	Basis
Tavily credits per crawl	40 live (14 searches, 2 extracts); 56 offline (24 searches, 1 extract of 16 URLs; the three regulations added 2026-10-15 replay empty searches)	usage on every Tavily response; documented pricing for the offline row
Lightning (normalise), 29 documents → 60 changes	682,890 prompt / 58,210 completion tokens, 115 calls	$0.0549 live
Ultra (classify), 60 changes	102,279 prompt / 84,830 completion tokens, 61 calls	$0.3568 live
Full crawl	8 min 03 s	$0.4117 live in model tokens + 40 Tavily credits
Grounding	live: 105 to-dos stored, 0 ungrounded; offline: quotes verbatim first try 20/22	validate_quotes, ungrounded_todos

Prices live in Settings (.env.example). Tooling friction with numbers: FEEDBACK.md.

Scope cut

Pre-approved: if live classification of the EU Deforestation Regulation proves unreliable, EUDR is dropped and the withdrawal function stays; everything else is unchanged.

Status / what's not done

• Live since 2026-10-10: the web app on a Nebius Serverless Endpoint (URL above) and the crawl as a Nebius Serverless Job, with keys from a Nebius MysteryBox secret; a Job without the secret refuses to run (exit 2, never fixtures).
• No daily schedule (decided 2026-10-19): .github/workflows/daily-crawl.yml stays off, because a crawl's results reach the URL only by re-recording the seed and redeploying. Crawls run on demand.
• The demo serves the recorded seed run. "Report a wrong call" opens a mail draft to REGWATCH_REPORT_EMAIL (set on the demo).
• "Let the agent check" ships off by default (REGWATCH_AGENT_ENABLED); the public Endpoint has it on, with keys and the read-only demo profile (runbook step 12), capped at $1 and 24 Tavily credits per day. From the next image there is also an all-time cap over every agent step in the database (REGWATCH_AGENT_TOTAL_USD, default $10; REGWATCH_AGENT_TOTAL_CREDITS, default 300). Once either is used up, the agent switches itself off.
• Not done: dates in English only (no locale), no re-classification right after a profile change (the next crawl does it). Open issues: docs/issues.md.

Seed regulations

Directive (EU) 2023/2673 (withdrawal button, § 356a BGB, from 19 June 2026) and Regulation (EU) 2023/1115 (EUDR, as amended by Regulation (EU) 2025/2650: medium and large operators from 30 December 2026, micro and small operators from 30 June 2027). Added 2026-10-15: Directive (EU) 2024/825 (green claims and guarantee information, applied from 27 September 2026), Regulation (EU) 2025/40 (PPWR, applies from 12 August 2026; the e-commerce empty-space rule from 1 January 2030) and Regulation (EU) 2023/988 (GPSR, applies from 13 December 2024).

License

MIT