Where IDSentinel fails

Errors are part of the result.

Prediction-level document imagery is withheld from the public site because redistribution permission is uncertain. The verified transition counts and error structure remain visible.

M3 · Semantic

18 failures

Dominant observed transition: 6 SCREEN → BONA_FIDE.

M4 · + Texture

14 failures

Large observed improvement for PRINT; SCREEN errors shifted.

M5 · + Edge

16 failures

Improved over baseline overall, with lower SCREEN F1.

M3 failure transitions

BOTH7

Corrected by texture and edge

TEXTURE ONLY4

Corrected only by M4

EDGE ONLY2

Corrected only by M5

PERSISTENT5

Wrong across all three

Transition analysis shows complementarity and regressions that aggregate accuracy hides. It does not establish why an individual image failed.

What can be stated

BONA_FIDE remains difficult

Its F1 was .414 in M3 and .480 in both branch models.

Texture helped PRINT

PRINT F1 rose from .690 to .889 in the observed M4 run.

Edge changed the tradeoff

PRINT improved to .846 while SCREEN fell to .595.

Why no failure gallery?

The local research artifacts contain real identity-document crops. Their public redistribution status is uncertain, so the repository and website deliberately omit those images rather than compromise dataset licensing.