SouqDataSouqData

Data Quality Changelog

A public, append-only record of every correction, purge, and methodology change on SouqData. If you downloaded data before one of these events, your snapshot may be inconsistent with the current platform — re-sync or consult the event notes below.

Machine-readable feed: /api/v1/data-quality/changelog

criticalMar 15, 2026·purge·D49b

Purged fabricated director records generated by Claude Haiku

An audit of director ingestion revealed that records with source='claude-haiku' were generated without a real source document for exchanges where our scrapers were blocked (Cloudflare / WAF). Haiku produced plausible-sounding names and assigned them across unrelated companies. All 824 appointments, 493 directors and 1,038 skill rows tagged source='claude-haiku' were deleted on 2026-03-15. Data from yahoo_finance, qe.com.qa, claude_extraction (real PDFs), liqaa_agm, annual_report and role_inference was preserved. Clients who downloaded director snapshots prior to this date should re-sync.

Tables
directors, director_appointments, director_skills
Exchanges
MSX, BHB, BK, ADX, ASE
Rows affected
1,038
Version
2026.03.14 → 2026.03.15
Root cause

Ingestion script called Claude Haiku to fill gaps on Cloudflare-blocked exchanges without requiring a verified source document. MSX was 100% fabricated; BHB had ~130 fabricated entries.

Remediation

All claude-haiku-sourced rows deleted. New policy: Haiku may only extract from a real PDF or scraped HTML that is attached as source_url. MSX has since been rebuilt from msx.om (410 appointments, D76c). BHB is pending re-source.