Paper + feasibility audit
Separated paper-described components from implementation choices before writing model code.
Scope lockedDataset provenance audit
Found the proposed composite benchmark was source-confounded and treated it as a blocker.
Four-class merge rejectedProtocol A-Reduced
Streamed only selected frames, preserved lineage, and produced a balanced 300-sample protocol.
214 / 42 / 44Making it work on 8 GB
Serial decoding, local batch 1, zero workers, no RAM cache, CPU extraction, temporary free T4 training.
~573 MiB extraction peakFinding the signal
RetinaFace with CLAHE and Haar fallbacks generated 300 context-preserving 384×384 crops.
300 / 300 cropsStart simple: semantics
ConvNeXt-Tiny established a controlled semantic baseline.
0.590 Macro-F1What if texture helps?
A 64-D pixel-only texture path lifted PRINT F1 from .690 to .889.
0.679 Macro-F1What do edges add?
Fixed Sobel gradients added different corrections, but SCREEN F1 fell in this run.
0.640 Macro-F1Three experiments, one split
Compared aggregate metrics and failure transitions without test-set retuning.
5 failures persistedAnalysis + publication
Published safe code and aggregate evidence while excluding restricted document imagery.
Open research recordLessons, not slogans
Data first
A scientifically clean protocol mattered more than immediate training.
Branches shift failures
Texture and edge changed which samples failed—not simply how many.
Small data = humility
Forty-four test samples cannot support broad performance claims.