Confidence measures evidence quality, not certainty
The score combines source completeness, sample strength, agreement between independent context sections, historical calibration, and contradiction penalties. A high score can still lose; a low-data opportunity is not upgraded by persuasive language.
Operational principles
- Coverage is disclosed.
- Missing evidence creates a penalty.
- Disagreement creates a penalty.
- Performance history is retained for calibration.
The implementation is reviewed against published results and is expected to change when new evidence justifies a documented revision.