48 lines
1.2 KiB
Markdown
48 lines
1.2 KiB
Markdown
# Weighted Evidence Model
|
|
|
|
## Purpose
|
|
|
|
The earlier evidence engine treated all evidence items equally. This revision adds a more realistic model with:
|
|
|
|
- evidence-type weights
|
|
- recency weighting
|
|
- dimension-level rubric storage
|
|
- confidence estimates based on weighted support
|
|
|
|
## Evidence weighting
|
|
|
|
Default weights:
|
|
- explanation: 1.0
|
|
- problem: 1.5
|
|
- transfer: 2.0
|
|
- project: 2.5
|
|
|
|
## Recency policy
|
|
|
|
Each evidence item can be marked `is_recent`. Recent items receive a multiplier. This allows weak recent performance to matter more than stale success, which is useful for resurfacing fragile concepts.
|
|
|
|
## Confidence
|
|
|
|
Confidence is currently derived from total weighted evidence mass using a saturating function:
|
|
|
|
`confidence = total_weight / (total_weight + 1.0)`
|
|
|
|
This is simple, monotonic, and interpretable.
|
|
|
|
## Current mastery rule
|
|
|
|
A concept is mastered if:
|
|
- weighted mean score >= mastery threshold
|
|
- confidence >= confidence threshold
|
|
|
|
A previously mastered concept resurfaces if:
|
|
- weighted mean score < resurfacing threshold
|
|
- and recent weak evidence drags its summary downward enough
|
|
|
|
## Future work
|
|
|
|
- per-dimension mastery thresholds
|
|
- decay by timestamp instead of a boolean recent flag
|
|
- Bayesian knowledge tracing
|
|
- separate competence vs fluency models
|