Supplementary material and archived generation outputs for "LithoGPT: Baseline Floors, Depth Conditioning, and Coherence at Range in Autoregressive Well-Log Generation"

Companion record to SPE conference paper SPE-234177-MS, "LithoGPT: Baseline Floors, Depth Conditioning, and Coherence at Range in Autoregressive Well-Log Generation". SPE conference papers do not carry supplementary files. The protocol and estimator definitions, the data provenance and the guard-metric tables were moved into the paper itself as Appendix A, and the seven further limitations became items 13 to 19 of its Limitations section. This record holds the supporting detail that did not need to travel with the manuscript. Contents. DEPOSITED_SUPPLEMENT.md: model, arm and fitting detail; the full seven-claim reconciliation with the accepted abstract; tokenizer refit robustness across three random states; per-well depth slope of NPHI; full teacher-forced reliability tables; full run-length statistics; chance-corrected token agreement at every lag; dispersion at three named estimands; and spatial continuity under the original study's definition. FigS9_continuity.png and FigS9_continuity.pdf: the spatial-continuity figure referenced by that section. atce_ablation_v3_raw_results_2026-07-26.json: the archived raw generation outputs cited in the paper's Data and Code Availability statement. 149,074,228 bytes, SHA-256 ac963f2c56ec27ac8f2fd837faa90108dc580a53560dbeb3c50a41c35f0110ed. The well-log data underlying the study are the 2020 FORCE Machine Learning Contest release (Bormann et al. 2020), measurements under NLOD 2.0 and lithofacies labels under CC BY 4.0.