[devissaputra]

Paper 1 — completed feasible scientific/technical work (22 October 2026)

Author-review working preprint, 11 pages. Draft PR; do not merge without the author reviewing claims and contributions.

1. Official corrected 2009 provenance: now directly established as NON-equivalent

Official maintainer-linked corrected ASSISTments 2009 release (SHA-256 162ef8d2d28bcbfea6591a282994062bd8d5eaa00636544292a0d268dca6e5da) has 4,163 learners and 283,105 skill-labeled response events. Frozen transformed Atomi mirror (SHA-256 2053b7ea63f1d5ea953834bf82ed89543e1f11bae1e3e7535c2f9d7f49d4dda7) contains 4,148 learners and 274,331 events.

All 4,148 mirror learner IDs exist in the corrected source. Among these, 3,331 have the same normalized skill-response sequence ordered by numeric order_id; 817 do not. Exact parity of the full source is therefore disproved. Student-level raw data have not been committed. Evidence: source comparison workflow 39773920168, aggregate report, runner.

2. Official corrected 2009 exploratory refit: completed

A new frozen, learner-disjoint 2009 corrected-source source-sensitivity experiment used 2,914 train / 624 validation / 625 test learners, 48,842 held-out interactions:

• Enriched PFA AUC 0.69508, Brier 0.19675, log loss 0.58086; LBFGS converged in 930 iterations.
• Extended GRU AUC 0.73243, Brier 0.18507, log loss 0.55124; best validation epoch 13.
• Paired pooled Brier GRU-PFA −0.0116811, 95% learner-cluster interval [−0.0137180, −0.0098625], 2,000 resamples.
• Exploratory because the 2009 environment and predictive direction were already studied; cannot prove architecture-specific causation.

Evidence: corrected source run 39774414969, source code, protocol, JSON.

3. ASSISTments 2015 independent-year replication: completed

A first nonconverged PFA pilot was retained as diagnostic/inconclusive. Stage 2 excluded its 5,000 learners, fit PFA and GRU on new training learners and evaluated 750 untouched held-out learners / 26,697 events:

• Converged enriched PFA Brier 0.18695, AUC 0.68129;
• Extended GRU Brier 0.18227, AUC 0.69968;
• Brier delta −0.0046813, bootstrap 95% interval [−0.0060730, −0.0033457].
  Evidence: Actions 39754854426.

4. Final artifact integration and new QA

• Integrated 11-page LaTeX preprint with corrected-source parity and exploratory refit, the two 2015 stages, and explicit causal and novelty caveats.
• Final scholarly research audit.
• Machine-verifiable final 60/60 release checks, including clean-room two-pass LaTeX compile, bibliography, figure integrity, output-content consistency and SHA-256 checks.
• Proposed arXiv metadata and author-only checklist.
• Complete manuscript PDF and minimal arXiv source ZIP are available in the user's Library folder /Paper1_Final_20261022 and current ChatGPT delivery. The GitHub TeX source alone requires the three figures shipped in the arXiv source ZIP.

Remaining external gates, not delegated

1. Author personally verifies and approves contribution, raw source inference and disclosures.
2. Optional independent human scholarly review (arXiv moderation is not academic peer review). Not claimed as performed.
3. Author's arXiv account/endorsement/license/preview and final submission. Not performed.

Limits: The 2009 strict parity is refuted. Findings compare unequal information representations and model capacities, do not identify a recurrence-only effect or teaching efficacy, and do not establish wide institutional transportability. These limitations are explicit in the manuscript.