# Seed registry — Wikidata harvest

Generated by scripts/seed_wikidata.py on 2026-09-13T07:17:29+00:00. 32368 companies, 13209 public (40.8 %), 12781 with a stock exchange, 10969 with a ticker, 4364 with an employee count, 5294 without an industry mapping (16.4 %), 206 countries.

Files: one NDJSON per UN region (wikidata-<region>.ndjson), one JSON object per line — see docs/SEEDS.md for the schema, the diversification rules and how to add companies. scripts/seed_edgar.py rewrites the files in place with SEC data; run it after every assemble.

# Tiers

# Regions

# Countries

# Stock exchanges (primary listing kept per company)

# Top-level industries (a company counts once per top-level sector)

# All industries

# Harvest report — 2026-09-13

• Pipeline: scripts/seed_wikidata.py all --target N (candidates → dissolved → select → details → assemble), then scripts/seed_edgar.py, then catlas seed. Every SPARQL result is cached under data/seed/wikidata/sparql/; per-stage counters in data/seed/harvest-stats.json.
• Candidates: 51,286 unique items with a website from 44 P31 classes × sitelink bands (≥ 3), 250 stock exchanges scanned in 13 P414 queries (13,541 currently listed), employees ≥ 500 / revenue bands per class, plus country and industry boosts. First seen per band: class 39,684, listed 7,664, revenue 1,748, employees 1,723, industry-boost 354, country-boost 103, ticker 10.
• Dissolved check (P576 / P582, 171 queries of 300 QIDs): 3,580 of 51,286 candidates excluded.
• Prefilter: 89 generic / invalid websites dropped, 1,639 duplicate registrable domains (highest sitelinks kept), 45,978 kept (12,935 listed).
• Selection (target 33,000; caps US ≤ 35%, other countries ≤ 12%, unknown country ≤ 3%, narrow classes 2–5 %): country minimums 2,160, then 11,964 listed companies, then 18,876 by sitelinks ≥ 3; 12,791 listed companies selected in total.
• Non-company exclusion: 652 items removed — 515 for a non-company P31 class (museums, libraries, universities / schools, government agencies, NGOs / charities / foundations / nonprofits, religious organisations, political parties, trade unions, hospitals, sports clubs, YouTube channels; top: nonprofit organization 170, government agency 40, university 30, museum 27, learned society 20, public university 20, non-governmental organization 19, charitable organization 18, association football club 15, educational institution 14, hospital 13, library 11) and 137 for a description matching museum / university / school / ministry / agency / charity / foundation / association / club / church / channel with no ticker, exchange, employee count or revenue. Items with a current listing or a ticker are always kept (listed football clubs). List: data/seed/dropped.json.
• Assemble: 33,000 selected companies with details + 0 from the industry top-up (every top-level industry ≥ 60); 0 subsidiaries sharing their parent's domain and 0 duplicate domains dropped; 1,458 related_domain_conflict notes → 32,368 companies.
• Tiers by importance quantile (1% / 6.67% / 26.67%): tier 1 = ranks 1–324, tier 2 → 2,483, tier 3 → 11,116, tier 4 = rest.
• Country minimums: all met (JP 3382, DE 2310, GB 1548, CN 1374, KR 846, IT 844, ID 731, CA 723 …).
• EDGAR (2026-09-13): 920 companies newly matched to a CIK (ticker, then exact normalised name), 2,543 already had one → 3,463 of 32,368 with a SEC CIK; 1,319 rows enriched, 202 industries from SIC (5,092 still without an industry mapping), 816 submissions JSON read.