Gold farm
Scrape JDs and grow the LoRA training corpora — analyst stays dormant, no analyst quota burn. See CLAUDE.md §8.107.
How gold-farm mode works
- Phase 1 (scrape + enrich + score) runs normally — JDs get scraped from Naukri + LinkedIn.
- The score worker calls
jd_preprocess.extract_fields()on every card; with backend=cli this fires Claude → captured todata/gold/jd_preprocess.jsonl. - Phase 2 (inbox enqueue) and Phase 3 (analyst) are SKIPPED — no analyst quota burn.
- xlsx writes go to
data/apply/excel_gold_farm/(separate dir, production xlsx untouched). - Each run ≈ 100-200 fresh JDs ≈ 200-400 Claude calls (jd_preprocess only) ≈ ~1.5 quota windows.
LoRA-readinessPer CLAUDE.md §8.94 — train the LoRA when count ≥ threshold
Spawns the run controller in
--mode gold_farm. Watch progress on /run.Recent gold-farm runs
Each run produced an xlsx audit + grew the gold corpus. Click 📊 Inspect to see the original JD body + Claude's structured extraction side-by-side. The xlsx is just the scrape-time audit (heuristic-scored metadata, not Claude analysis).