Frozen Protocol A-Reduced

Three experiments. One test set.

Explore the exact architectures, metrics, confusion matrices, and real training artifacts from M3, M4, and M5.

M4 · Frozen test split

Semantic + Texture

Accuracy68.18%

Macro-F10.679

Parameters28,355,317

Failures14/44

Learned fusion weights: Semantic 0.503779 · Texture 0.496221. Not causal feature-importance estimates.

61711214012

Rows: true · Columns: predicted

Side-by-side

Metric	M3 · Semantic	M4 · Semantic + Texture	M5 · Semantic + Edge
Accuracy	0.591	0.682	0.636
Macro-F1	0.590	0.679	0.640
BONA_FIDE F1	0.414	0.480	0.480
PRINT F1	0.690	0.889	0.846
SCREEN F1	0.667	0.667	0.595
Failures	18	14	16

ConvNeXt + Texture produced the highest observed score on this frozen test split. With only 44 test samples and one primary seed, the difference is descriptive rather than evidence of statistical superiority.