[devissaputra]

Draft verification branch. Adds a real-data robustness runner (25-epoch GRU, richer PFA baseline, paired learner-cluster bootstrap) and a GitHub Actions job to rerun both original and extended experiments. Do not merge or use newly generated outcomes in the manuscript until CI data verification passes and the author reviews the evidence. Dataset mirror provenance remains an independent release gate.

[devissaputra]

Completed empirical verification (22 October 2026)

GitHub Actions run: https://github.com/devissaputra/knowledge_tracing_benchmark/actions/runs/39655888620
Raw numerical artifact: https://github.com/devissaputra/knowledge_tracing_benchmark/actions/runs/39655888620/artifacts/12370764135

Pass: Automated tests, end-to-end original real-data experiment, extended 25-epoch-budget GRU/PFA robustness experiment, and output validation.

• Original dataset mirror SHA-256 reproduced: 2053b7ea63f1d5ea953834bf82ed89543e1f11bae1e3e7535c2f9d7f49d4dda7. 4,148 learners; 274,331 interactions; 623 test learners; 41,228 test responses.
• Original six-model results agree with archived values to reported precision (max absolute ROC-AUC drift < 4e-8).
• Enriched cross-skill-history PFA: ROC-AUC 0.71355; Brier 0.19293 (standard PFA 0.69823 / 0.19761).
• Extended GRU seeds 13/42/73: AUC 0.74705 / 0.74908 / 0.74983, Brier 0.17934 / 0.17905 / 0.17891; best validation epochs 13 / 13 / 11 (stopped before 25-epoch cap).
• Paired learner-cluster bootstrap (2,000 resamples), pooled-interaction Brier difference, extended GRU 42 minus enriched PFA: -0.013874 with 95% percentile interval [-0.017231, -0.011049]. Corresponding intervals exclude zero for the other GRU seeds.

Release caveats: All robustness tests reuse a previously inspected fixed test partition, so findings are exploratory rather than confirmatory. The enriched PFA gets cross-skill summary counts, not the ordered full history available to GRU; architecture is not isolated. The pinned mirror has not been reconciled with official corrected ASSISTments rows. No independent dataset/generalization evaluation. These are verified empirical outputs, not by themselves evidence of arXiv readiness.

Please retain this PR as draft until source provenance and manuscript claims are checked by the author.