Deep learning–based knowledge tracing, 2015–2025: derived bibliometric and taxonomy tables (Scientometrics submission)

v4.1.0 adds the analysis layer on the locked N=926 codes (codes themselves are unchanged). What changed vs v4.0.0 (2026-08-16):- review_reindex_n335/: Horvitz–Thompson / Hájek reindex of A3 consensus onto the current TAK residual 335. Transferable UNCERTAIN: transport 20.7% [17.6, 23.8]; conservative 21.9% [18.9, 24.8]; two-arm on NL2 human DETERMINED-arm rates 28.2% [22.8, 34.0].- Same-record AI-draft vs human tables on 145 records (gross 86/145; UNCERTAIN 9 vs 67).- Concordance split by origin: NL2 human 0.653 (n=75) vs mixed 0.536 (n=97) vs AI-draft 0.136 (n=22).- Scripts that regenerate those tables. A3 consensus files in a3_residual_validation/ remain the 2026-08-03 sample. Their HT columns (15.1% / 30.4%) are an archived-residual comparator, not current-label shares. Not included: licensed Scopus raw exports; copyrighted full-text PDFs; coordinator_only/ AI-draft labels; coder working sheets. Claim boundaries: TITLE-ABS-KW codes, not a full-text census of N=926; A3 is a stratified residual audit (n=194 AVAILABLE of 200), reindexed not redrawn; two-arm 28.2% uses NL2 human DETERMINED-arm rates (n=75), not the mixed arm. Companion manuscript: Scientometrics submission. Concept DOI (all versions): 10.5281/zenodo.21635880.