Improvement: Temporal Score Normalization Across Cohorts

[sharmaanchita]

Problem
Age weighting exists, but scores are not normalized across time, making historical comparisons unreliable.

Basis of issue

1. Temporal normalization layer across cohorts
2. Score decay or relevance adjustment over time
3. Cross-cohort score alignment
4. Separation of raw vs normalized scores

Importance

1. Paper criterion Prevent validators from censoring prompts after the fact. #6: longitudinal normalized evaluation
2. Model capabilities improve over time
3. Raw scores become misleading without normalization

Current implementation gap

1. Linear / exponential age weighting only
2. No true temporal normalization

Implementation checklist

1. Normalized score metric independent of evaluation date
2. Cohort-aware score calibration
3. Historical scores remain comparable
4. Clear distinction between raw and normalized scores