Evol-1 T05: extraction cascade v0 (rules + local LLM)

[thanhhale288]

Summary

• Phase / scope: Evol-1 T05 — thác trích chỉ tiêu v0 (tầng 1 luật + tầng 2 LLM local), không cào listing sàn.
• What changed:
  • Package crawlers/extraction_cascade/ + runner: fetch website DN → tier1 rules → optional pinned Ollama tier2 → agree/conflict/abstain tables.
  • Cohort 28 listed + ~100 frame_pilot URLs (DNS-filtered domain hypothesis); artifacts under data/processed/extraction_cascade/ (summary, indicators, conflicts, PROVENANCE).
  • Pilot run with --llm: ~89/128 fetch_ok; tier compare includes real conflicts (e.g. COD/social) for advisor slides — not national estimates.

Scratch / docs

• Spec: docs/evol-1.md T05; design: docs/proposal-v4.md §4.5 cascade.
• ADR touched? No (no Digital VA / OECD formula change).

Test plan

• PYTHONPATH=. pytest tests/extraction_cascade/ -q (14 passed)
• PYTHONPATH=. pytest -q (full suite / CI)
• cd frontend && npm run build (if UI changed) — N/A
• Manual: cascade artifacts present; no marketplace listing scrape; caveats in summary/PROVENANCE
• No invented GSO/OECD/CafeF/marketplace numbers; fetch fail → skip, not invent

Checklist

• Branch named cursor/evol1-task05-extraction-cascade-v0
• Did not commit .env, secrets, large raw dumps, or unrelated model binaries
• Formula/domain changes updated CONTEXT.md (+ ADR) — N/A

Made with Cursor