Evol-1 T08: Japan URL-finder calibration (gBizINFO silver, n=300)

[thanhhale288]

Summary

• Phase / scope: Evol-1 T08 — calibrate T03 URL-finder on ~300 Japanese manufacturing firms with gBizINFO silver website URLs (leak fence: URLs only in data/raw/jp_labels/).
• What changed: crawlers/jp_calibration sample+eval pipeline; locale ja.json + RQ3 lifts (.co.jp 2LD, TLD bonuses / Accept-Language in locale config); eval artifacts hit 21/300 (7.0%), abstain 80.3% (search blocked HTTP 202, same method as T03).
• Omitted regenerable gbizinfo_search_pool.json (~33k lines) and NTA/HTML caches from the PR.

Scratch / docs

• Spec or status: .scratch/evol1-task08-jp-calibration-pilot/spec.md
• ADR touched? (Digital VA / OECD policy): no

Test plan

• PYTHONPATH=. pytest -q tests/jp_calibration tests/url_finder (43 passed)
• cd frontend && npm run build (if UI changed) — N/A
• Manual: open data/processed/jp_calibration/comparison_vn28.md + fill review_30.csv human columns (~30 mismatches)
• No invented GSO/OECD numbers; JP provenance recorded; silver labels not gold

Checklist

• Branch named cursor/evol1-task08-jp-calibration-pilot
• Did not commit .env, secrets, large raw dumps, or model binaries
• Formula/domain changes updated CONTEXT.md (+ ADR under docs/adr/ if needed) — N/A

Made with Cursor