Online Resource 3: Reference-ranking sensitivity and updated comparison figures
Article: TraceRIFT: Reference-preserving ensemble selection for discrete fracture networks from incomplete traces
Journal: Earth Science Informatics
Authors: Shuo Wang, Chen Guo, Bingxin Yang, Keran Hao, Yu Zhang, Fuqiang Zhang
Affiliation: School of Information Engineering, Chang'an University, Xi'an 710064, China
Correspondence: Chen Guo, chenguo@chd.edu.cn
Prepared: 2026-10-10

This resource reports a retrospective, fixed-protocol sensitivity analysis on the existing test cohort. It supplies all three Reference rankings and all nine Reference/selector combinations, without selecting ranks by their observed results. Each ranking defines its own top ten and core six. Candidate count 64, retained count 10, external budget 2, A13 predictive scenarios and training-fold scales are unchanged. No generator or predictor is retrained. This is not cross-generator or independently untouched validation. Sequential-minus-TraceRIFT intervals include zero under all three rankings.

The source release is https://github.com/SureWangchd/TraceRIFT-CG/releases/tag/v0.1.1 . Obtain its registered statistical assets as instructed there. Large original candidate/predictor/truth caches are not duplicated in this ZIP. The exact asset and upstream-source SHA256 values are in protocol_frozen.json; prepared_inputs.json records derived input hashes. Code uses NumPy and the existing TraceRIFT-CG source. The frozen original and equivalent selectors were not edited.

Extract this ZIP next to the release source folder, keeping these sibling directories:
  TraceRIFT-CG/
  esi_reference_robustness_20261010/
  esi_external_baselines_20261010/optimization/fast_selector.py
  esi_figure_sync_20261010/
Use the original release's cache hierarchy under TraceRIFT-CG. The recorded protocol deliberately checks its original absolute source paths and exact Python executable/version. Do not edit these provenance receipts to make another machine pass. To perform a new reproduction elsewhere, create a NEW sibling folder such as reference_reproduction next to TraceRIFT-CG. Copy only common.py, select_policies.py and evaluate.py into it; keep the supplied optimization/fast_selector.py at the sibling location above. Do not copy the recorded protocol_frozen.json, prepared_inputs.json, inputs or results into the new run folder. The prepare command then records a new environment/path freeze while leaving the delivered run intact. For an optional fresh independent audit, also copy audit/independent_audit.py and the superseded_freeze_01_missing_torch provenance directory into that new folder; do not copy the recorded final audit reports.

With Python 3.11.4 / NumPy 2.4.6, from that NEW reproduction folder run:
  python -B select_policies.py prepare
  python -B select_policies.py development --workers 4
  python -B select_policies.py heldout --workers 4
  python -B evaluate.py
The first three commands never load query truth. Evaluate verifies saved policy hashes before loading truth. Independent audit code and recorded reports are under audit/. The new preparation locates sources through the sibling layout and freezes the current absolute paths; its subsequent stages enforce that new freeze. A superseded preparation-only attempt is retained for provenance and is not an execution entry point. A new environment may change floating-point or timing behavior; the exact archived results remain the delivered evidence, not overwritten receipts.

For the updated Figure 6/7 alone, return to the ZIP extraction root (the parent of the NEW reproduction folder), then use the bundled CSVs without the large scientific caches:
  python -B esi_figure_sync_20261010/plot_figures.py
The plotting script requires NumPy, pandas and matplotlib plus a local Times New Roman font, and uses included style/source_data files. Vector PDF, editable SVG, EPS and 600-dpi PNG exports plus export/figure QA are included. Source-data extraction is separate and optional. The figure folder's README and QA document its candidate-delivery stage; the final main manuscript embeds those PDFs at their native 119 mm width, giving 8--9 pt lettering. The eight-method plots refer to the primary Borda comparison; MMR and farthest-first were not tested with the alternative References in this resource.

All paired effects use first minus second as named in paired_comparisons.csv; positive values favor the second method. Primary means average the two degraded conditions within each parent. The 10,000 parent-bootstrap intervals are pointwise. Reference itself and complete observations do not undergo the independent check. Results include absolute regret, total loss, support loss, recalls, fold/condition summaries and full selection policies. No parent/query records were excluded. Existing Online Resource 2 is preserved separately and supplies the external-baseline and timing analyses.
