Repository navigation
Data-quality acceptance fixes
Failed explicitly requested integrity checks and missing required fields now reject records independently of aggregate score and tier. Previously a failed hash check could still produce a score of 86 and be accepted. Outcomes survive SQL/document storage and appear in decision reasons and validation_passed.
The scanner also detects altered derived reference errors and agreement scores. Policies reject malformed stored check flags rather than accepting truthy substitutes.
Performance and validation
Routine scoring no longer constructs explanation dictionaries or redundantly calculates confidence. The optimized score matches the explanation path across 300 deterministic cases covering all built-in profiles.
Local validation: 94 tests, strict documentation build, README example, package build, and Twine checks passed. A three-run local benchmark of 10,000 single-field scans per run measured median throughput of approximately 38,200 records/second versus 35,900 before the change (about 6.5% higher). This excludes file/database I/O; it is not a production performance guarantee.
Compatibility
New evidence snapshots contain optional required_fields_passed and integrity_checks_passed fields, which affect newly computed evidence hashes. Decision exports add validation_passed. Existing snapshots without check outcomes retain their prior behavior; re-scan historical records to apply these checks. Explicitly constructed evidence still requires justified inputs, and reference sources and tolerances must be validated for the intended use.
Install: pip install data-reliability-index==0.7.1.