v0.2.0 — The AI-Era MDM Release

v0.2.0 — The AI-Era MDM Release

This release transforms AURUM from a working pipeline into the definitive open-source MDM reference repository.

What's New

📊 Sample Data — 3,000 rows across all 7 domains
Previously only 4 domains had sample data. Now all 7 are covered with deliberately dirty, globally diverse synthetic data. Fictional company names throughout — no real organisations referenced.

Domain	Rows
Customer	600
Product	500
Vendor	500
Asset	400
Employee	400
Location	300
Counterparty	300

📚 41 Use-Case Playbooks (use_cases/)
Three tiers of real-world MDM scenarios:

• Tier 1: 35 single-domain use cases — 5 per domain across all 7 domains
• Tier 2: 5 cross-domain pair scenarios (2–3 domains interacting)
• Tier 3: 1 grand scenario — New Store Opening, all 7 domains in one business event

📖 Narrative Documentation

• THE_INTELLIGENT_REFINERY.md — MDM in the age of AI agents: LLM rule generation, Isolation Forest anomaly detection, MCP-native pipeline, agentic stewardship
• THE_MDM_WORLD.md — 4-act MDM journey: single domain → silos → partial cross-domain web → full golden web

⚖️ Global Compliance Guide

• DATA_SOVEREIGNTY_AND_COMPLIANCE.md — AURUM mapped to UAE PDPL, ADDA, DDA, KSA PDPL + NDMO, Qatar PDPL, GDPR, CCPA, India DPDP, China PIPL + DSL, Singapore PDPA, Australia Privacy Act, POPIA and more. Official public document links included for every framework.
• DISCLAIMER.md — all entities fictional, coincidence disclaimer

🔧 CI Fixed
All 5 jobs green across Python 3.10 / 3.11 / 3.12.

---

Full details in CHANGELOG.md

v0.1.1 — Architectural rewrite of survivorship pipeline

Changed

• Survivorship pipeline rewritten with three-step architecture:
  standardize → validate → survive. Real MDM tools cleanse before
  validating before surviving; v0.1.0 jumped straight to validate.
• Names now use independent per-field survivorship (the matcher
  established identity; survivorship picks the cleanest version of each
  attribute independently).
• Geography (city, country) now uses linked-tuple survivorship —
  the whole tuple comes from one source. Prevents the "Dubai, UK"
  frankenrecord failure mode where independent field selection produces
  geographically impossible combinations.
• Trust score reformulated: 0.6 × completeness + 0.4 × source
  diversity. Single-source records cannot exceed ~0.73 — diversity matters.
• Matcher re-weighted: name 0.65, email 0.25, phone 0.10. Added a
  name-boost floor so strong name matches alone can carry a pair.
  Threshold lowered to 0.65 with the boost.
• Sample data generator now emits geographically-consistent dirty
  pairs (Dubai always with a UAE variant; London always with a UK variant).
  Format variation is realistic dirt; geographic mismatch was unrealistic.

Added

• Standardization layer (standardize_name) reverses known dirt
  patterns before validation: leetspeak (@→a in non-email context),
  trailing punctuation. Pluggable via standardizer callable for
  domain-specific rules.
• Transitive cluster builder (build_cluster_ids) — connected
  components on the match graph. A→B and B→C cluster {A, B, C} even if
  A→C wasn't directly scored above threshold.
• Demo assertions as CI guards:
  • len(matches) > 0, len(cluster) >= 2
  • No @ characters in golden names (validator regression check)
  • Nameless-record assertion (fails loud if validator/standardizer break)
  • Frankenrecord geography assertion (golden city + country must
    co-exist as a pair in some cluster source)

Fixed

• v0.1.0 produced a "golden record" called S@R@H Smith with a misleading
  trust score of 1.0. Names are now standardized before validation.
• v0.1.0 found zero matches above threshold despite duplicates in the data.
  The matcher reweighting and name-boost floor fix this.
• v0.1.0 trust score measured only attribute fill rate, ignoring source
  diversity. Any random row with all fields filled scored 1.0. Fixed.

Documentation

• README rewritten with explicit Component Status table — every component
  marked ✅ Working, 🔧 Stub, or 📋 Planned. No more overpromising.
• ROADMAP.md added — captures v0.2.0, v0.3.0, v0.4.0, v1.0.0 milestones
  and triggers.
• CHANGELOG.md added — this file.

v0.1.0 — Initial reference release

[v0.1.0] — 2026-05-02

Added

• Initial reference implementation across all 5 stages
• 7 domain models (Pydantic)
• Sample data generator (deliberately dirty across 4 domains)
• ASSAY schema inspector
• UNEARTH Customer profiler
• REFINE matcher and survivorship (v1 — superseded in v0.1.1)
• UNFURL publisher stub
• MARK lineage tracker
• MCP server with 3 tools
• Power Platform Customer Dataverse schema
• Certification program outline (3-tier curriculum)
• CI workflow (Python 3.10/3.11/3.12)
• Trust files: SECURITY, CONTRIBUTING, CODE_OF_CONDUCT, CITATION, LICENSE