File size: 896 Bytes
58904bb | 1 2 3 4 5 6 7 8 9 10 11 12 | \begin{tabular}{llll}
\toprule
Experiment layer & What is run & Question answered & Statistical unit \\
\midrule
Static calibration & Per-regime dropout grid at fixed $(P,U,C)$ cells & Where does validation loss plateau over dropout? & 15--16 cells per regime \\
Coefficient fit & Per-regime regression from cell targets $p^\star$ & Can pressure variables predict useful dropout? & MAE / leave-out MAE \\
4M final-prefix locked stream & Per-regime continuing model over growing prefixes & Does decay beat the best tested fixed dropout? & 5 paired seeds per regime \\
L20 8M extrapolation & Per-regime 51.84M-param model with frozen coefficients & Does the law extend to larger model/data scale? & 3 paired seeds per regime \\
Stage-wise audit & Per-regime decay vs best static at every prefix & Is the schedule final-loss or per-prefix optimal? & Prefix-level deltas \\
\bottomrule
\end{tabular}
|