lin-0159
12.5 Blockwise admission
The follow-up experiment was registered separately with fresh seeds, four fixed training and validation probes, pretrained block normalization, and lexicographic guards. It met all criteria.
|
Audit |
Pretrained |
Blockwise LINCS |
Change |
Result |
|---|---|---|---|---|
|
Task RMSE |
\(0.05629\) |
\(0.05633\) |
\(+0.06\% \) |
inside task guard |
|
Base Kan RMS |
\(0.94531\) |
\(0.88325\) |
\(-6.56\% \) |
improved |
|
Tangent Kan RMS |
\(1.87822\) |
\(1.73857\) |
\(-7.43\% \) |
improved in \(5/5\) seeds |
|
Perturbed RMSE |
\(0.05743\) |
\(0.05730\) |
\(-0.23\% \) |
preserved |
|
Length-7 RMSE |
\(0.10792\) |
\(0.10758\) |
\(-0.31\% \) |
preserved |
The result does not prove that blockwise admission always outperforms a joint score. It establishes a stronger contract: every admitted repair satisfies the declared task, base, tangent, and normalized-score guards.