Reproduction logbook — paper-E3C6mg6v4b

Semi-Supervised Gaze Estimation via Disentangled Subspace Contrastive Learning (DSCL) — arXiv 2605.27080v1.

Five claims falsified, one verified. None inconclusive.

• 1 — falsified. In the claim's own 5% column, 8.42° of the 8.79° needs no semi-supervised learning at all, and the same DSCL module averages 1.89° on the five other backbones there.
• 2 — verified. γ=1.0 drives the Jacobian's off-dominant column mass 0.699 → 0.550 (p = 0.005); exclusive control is exact.
• 3 — falsified. In that exact cell PureGaze starts 2.09° behind the paper's own plain ResNet-18 baseline and finishes 0.53° behind it once DSCL is applied to both.
• 4 — falsified. The L_J ablation, actually run: +0.175% relative (p = 0.46), against a claimed 6.24%.
• 5 — falsified. Over the same three settings, Table 1's RT-Gene reaches 13.67° with no unlabelled data — DSCL's margin is 0.07° — and head to head DSCL loses to CLSS at all three fractions.
• 6 — falsified. f(y)=Σyk is a counterexample; 0 violations in 16 040 025 comparable pairs.

The logbook is in pages/; the runnable evidence is in code/ and outputs/. Everything runs on CPU.