Pavelevich/llm-checker v3.3.0

on GitHub

v3.3.0 Highlights

• Added a Calibration Quick Start path in the main README for first-time setup.
• Added calibration fixtures and deterministic end-to-end coverage for:
  • calibrate --policy-out ... -> recommend --calibrated ...
• Expanded calibrated-routing docs and precedence behavior:
  • --policy takes precedence over --calibrated
  • default discovery at ~/.llm-checker/calibration-policy.{yaml,yml,json}
• Updated command docs for generated calibration artifacts:
  • calibration-result.json
  • calibration-policy.yaml
• Updated ml-model/README.md quick-start and CLI examples to match current commands.
• Fixed training artifact output path to reliably write into ml-model/trained regardless of working directory.
• Hardened Jetson CUDA detection to avoid false CPU-only fallback and added regression tests.

Known Limitations

• calibrate --mode full currently supports --runtime ollama only.
• recommend/ai-run fall back to deterministic selection when calibrated policy is missing/invalid or route models are unavailable.
• Calibration suite quality checks (checks) are optional in dry-run and contract-only modes.