v0.3.0 — Smarter Detection: Configurable Drift & Anomaly Scoring

v0.3.0 - Improved drift/anomaly detection, configurable parameters

Highlights:
- Anomaly detection: 20% → 80% detection rate
- Drift detection: 0% → 60% with realistic form revisions
- Scoring rebalanced: 60% spatial + 40% name
- All parameters configurable via field_memory.yaml
- FieldMemoryConfig class + TemplateMemory.from_config()
- STRESS_TEST_REPORT.md (90 complex PDFs, real Textract)
- 158 unit tests passing

v0.2.0 — Real OCR Support & Improved Docs

What's New

Field Name Normalization (Critical Fix)

• Library now works correctly with real Textract/OCR output
• Handles trailing colons, punctuation, and whitespace in field names (e.g., "Employee Name:" → matches "Employee Name")
• Normalization happens at lookup time only — stored data preserves OCR fidelity
• Validated with real AWS Textract API calls: all fields score 0.999

Improved README

• Added "The Problem" section with before/after diagrams
• Added "Beyond Document Processing" — applicability to UI testing, manufacturing, medical imaging, retail
• Added architecture diagram (Mermaid)
• Added FAQ, badges, "Try it in 30 Seconds", decision tree
• Added non-production disclaimer and PII compliance note

Other

• Fixed all PCSR Guardian feedback (F-1 through F-5)
• License headers corrected to MIT-0 across all files
• Benchmarks directory added (gitignored — run locally)

Proven Results (Real Textract)

• Template identification: 99.9% accuracy
• Spatial field location: 99.9% confidence
• Token savings: ~$200 per 10K documents (estimated)

Initial Release

First public release of textract-field-memory.

Features:

• Template learning and spatial field matching
• Document type identification by layout
• Anomaly and drift detection
• Template health analytics and field stability scoring
• Batch processing with error isolation
• Document cluster tracking
• Export/import (JSON/CSV)
• Interactive Streamlit dashboard
• Zero runtime dependencies (pure Python stdlib)
• 158 tests passing