[devissaputra]

2026 methodology upgrade of Paper 1 (new empirical research; DRAFT)

Why this replaces the older model-ranking narrative

Recent work already studies source corrections, leakage and information matching (Polydoros 2026, Padi et al. 2026, Badran & Preisach 2026). The new question is a frozen history-budget sensitivity test: when a count-based logistic model, an order-aware logistic model, and a compact GRU all see the same maximum K prior interactions and are scored on exactly the same target events, how do their relative rankings change?

Completed real-data execution

• Official 2015 source SHA-256 pinned and verified.
• Stage-3 selected 4,000 learners with zero overlap against the previous two 5,000-person 2015 cohorts; 2,800/600/600 learner-disjoint splits.
• 582 evaluable test learners and 20,592 identical responses across all six model-budget combinations. Eighteen one-interaction test learners correctly have no t>=1 target.
• At K=1 Brier: count 0.177882, order-aware 0.177882, GRU 0.180003. GRU minus order-aware +0.002121, learner-cluster 95% interval [+0.001198,+0.003013].
• At K=20 Brier: count 0.176674, order-aware 0.173885, GRU 0.175140. GRU minus order-aware +0.001255 [0.000075,0.002426]; GRU minus count -0.001533 [-0.002761,-0.000279].
• The ordering of models depends on which history representation baseline is used: a meaningful sensitivity result, not causal architecture attribution. CIs are descriptive, uncorrected multiple comparisons and conditional on single split and seed.

Reproducibility and documents

• Frozen pre-test protocol
• Full executable experiment
• Actual machine-readable results
• Successful GitHub Actions 40310733288
• Integrated 13-page methodological LaTeX manuscript
• Executed scientific results / prior-literature comparison
• Reproducible generated figure script
• 41-check document / numeric QA
• Final PDF, full package and minimal arXiv source ZIP preserved under ChatGPT Library /Paper1_Methodological_2026.

QA history and scope

The initial run completed computational analyses but release CI erroneously equated 600 test-split members with 600 evaluated post-first-response learners. After recording this in the frozen protocol's erratum, the release-only assertion was fixed and all exact-event alignment guards passed. No models, features, hyperparameters, seeds, data, or result choices were changed.

Leave draft and layered on prior verified PR #2 until author reviews every claim, verifies AI-assistance disclosure and authorship, and approves possible arXiv submission. This is not peer reviewed, not a new KT algorithm, not modern Eedi validation and not demonstrated generalization to other education platforms.