# Changelog

All notable changes to this project are documented here.

## [Unreleased]

### Changed
- **inference.py**: Set `TEMPERATURE = 0.0` for reproducible baseline (was 0.3)
- **inference.py**: Added explicit `ValueError` if `HF_TOKEN` environment variable is missing
- **inference.py**: Clarified score clamping to `(0.01, 0.99)` for normalization epsilon
- **README.md**: Updated Tasks table with "Est. Frontier Score" column for each difficulty tier
  - Clinical Note Review (Easy): 0.70–0.85
  - Medical Calculation (Medium): 0.45–0.65
  - Diagnostic Workup (Hard): 0.30–0.50
- **README.md**: Corrected dataset sizes to reflect actual file counts
  - MEDEC: 55,394 total (training 36,766 / validation 9,129 / test 9,499)
  - MedCalc-Bench: ~138,000 total (numeric-answer rows only; train 136,969 / test 16,521)
  - MedCaseReasoning: 766 structured clinical cases
- **README.md**: Updated headline from "15,000+ real cases" to "194,000+ real cases"
- **README.md**: Added note: "Each inference run picks one random case per difficulty level"
- **README.md**: Added "Prior Work & Novelty" section with SDBench (Nori et al., 2025) citation
- **README.md**: Expanded novelty claims with three key additions:
  1. Multi-skill integration (note review + calculation + diagnosis)
  2. Developer-native CLI interaction (same as Claude Code / aider / Codex)
  3. Scale (194,000+ real cases, not curated vignettes)
- **README.md**: Added Pydantic class definitions to Action/Observation space section
- **README.md**: Added JSON action format examples to Action space
- **README.md**: Added compliance note to Baseline Scores section on runtime and stdout format
- **GitHub**: Created public repository at https://github.com/pratinavseth/claude-code-for-health

### Fixed
- **inference.py**: Reverted score clamping from `(0.0, 1.0)` back to `(0.01, 0.99)` (epsilon required for normalization)
- **README.md**: Corrected all dataset row counts (external analysis had incorrect numbers)
- **README.md**: Clarified that 194,000+ is the full data pool, not per-run consumption

### Notes
- All changes maintain backward compatibility with OpenEnv spec
- Baseline inference completes in <5 minutes on standard hardware
- Docker build and HuggingFace Space deployment remain compatible