v0.5.0 - CRT and CAS validation expansion

v0.5.0

This release expands C-DAG with real CRT and CAS validation paths while preserving the original reference model behavior.

Added

• Public CRT validation path
• Freddie/STACR-style loan-level validation support
• Fannie CAS April 2026 loan-level validation support
• CAS external header-file parsing
• Streaming CAS ingestion for large 3GB+ files
• CAS-specific calibrated policy configuration
• Real CRT validation report updates
• Public mortgage holdout validation baseline
• C-DAG public-facing naming update while preserving causal-credit-risk-engine as the formal implementation name

Validation examples

Freddie/STACR-style CRT validation:

• Rows processed: 10,000
• Accepted rows: 10,000
• Rejected rows: 0
• Decision distribution: APPROVE 9,299 / REVIEW 0 / DECLINE 701
• Replay success rate: 1.0
• Audit-chain verification: true
• Evidence-pack mode: sampled, 1,000 rows

Fannie CAS April 2026 validation:

• Rows processed: 10,000
• Accepted rows: 10,000
• Rejected rows: 0
• Base policy distribution: APPROVE 7,722 / REVIEW 2,278 / DECLINE 0
• Calibrated CAS policy distribution: APPROVE 7,948 / REVIEW 686 / DECLINE 1,366
• Replay success rate: 1.0
• Audit-chain verification: true

Mortgage holdout validation baseline:

• Train rows: 58,579
• Test rows: 41,421
• Test positives: 201
• AUC improved from 0.538963 to 0.573062
• PR-AUC improved from 0.005601 to 0.006059
• Decision distribution after calibration: APPROVE 36,336 / REVIEW 5,085 / DECLINE 0
• Replay success rate: 1.0
• Audit-chain verification: true

Unchanged

• Core inference math unchanged
• Reference demo model unchanged
• Default example risk remains 0.849375
• BUSL-1.1 license posture unchanged
• No raw data or validation outputs are tracked

Important limitation

This is public financial dataset validation, not production validation. It does not use customer data, does not make credit eligibility decisions, and does not prove regulatory compliance.