| name | leaderboard-dynamics-analysis |
| description | Analyze leaderboard score distributions, compression, selective reporting |
| execution | subagent |
| prompt | ./prompt.md |
| input | leaderboard_data (model, score, date) |
| used-by | benchmark-archaeology |
Leaderboard Dynamics Analysis SOP
Analyze the dynamics of benchmark leaderboards: score distributions, compression over time, selective reporting patterns, and statistical artifacts.
Input
- leaderboard_data: Structured data with (model_name, score, date, optional: model_size, paper_reference)
Procedure
- Compute score distribution statistics over time
- Detect score compression (shrinking gap between top models)
- Identify selective reporting patterns (cherry-picking, subset selection)
- Analyze relationship between model size and score
- Detect anomalies (withdrawn results, unreplicated claims)
Output
Leaderboard health assessment with statistical evidence.