docs/workspace_audit.md

This is the original GitHub source-selection audit. The HF payload adds the full graph set and explicitly indexed Viewer tables; current HF coverage is recorded in the root README and provenance/hf_source_snapshot.json.

Selection of paper artifacts

The source workspace was inventoried by filename, size, and cloud-placeholder status before copying. Its research branches are not all part of the paper. No docking, acquisition campaign, selector training, or new scientific comparison was run during packaging. Original workspace files were not edited.

Source area	Decision and reason
MolAS_AL_restart/AL_MolAS/mean_ucb_cost_5fold_paper	Primary acquisition implementation candidate and six-policy, five-fold, three-acquisition-seed archive. Preserve historical source; make separately documented portable release changes. The downstream paper uses acquisition seed 0.
port_as_results_260829/{six paper policies}	Preserve downstream CSVs, including per-complex predictions, fold summaries, and aggregate tables. Missing cloud contents remain inventoried.
PORT-AS_cost_aware_bare.git and original cost-aware remote	Versioned acquisition matrix/query-log evidence. Pin the exact historical commit and content identities; do not replace it with a moving branch.
camera_ready_final_2026-10-31 and user-supplied SRS_2026_AL_MolAS_Docking-8.pdf	Preserve the audited source, style, bibliography, eight figures, and supplied 12-page camera-ready PDF.
MolAS_AL/data/raw	Readable original raw tables; PB/runtime agree byte-for-byte with restart copies.
MolAS_AL legacy strict-score output and restart ground_truth	Preserve both existing target matrices with explicit provenance. The restart original was subsequently recovered; its values and the legacy representation were checked against the raw data without conflating their distinct file identities.
Downloads/data/MolAS	Original resident ligand/protein features. Copy needed structural inputs into the local data area; retain hashes and separate large-data publication requirements.
Downloads/cost_aware_relaxed_six_methods.zip	Complete original six-policy archive. Verify all recovered CSV identities against the pinned Git tree; losslessly deduplicate cumulative prefixes in the release.
Downloads/MolAS_AL-main	Recovered historical MolAS training adapter, launcher, model/data modules, embedding-generation source, and analysis notebooks. Preserve unchanged with the supplied license and hashes; it is an unversioned downloaded tree.
Original paper download's NeurIPS2026/scripts and Downloads/main_results	Additional plotting source and rendered analysis assets; only demonstrated relationships are used in the figure-lineage map.
cost_aware_design2 and dal_toolbox_design2	Only the feature-cache lineage and original feature-building source are relevant dependencies. Their unrelated AL campaigns are not paper baselines.
mean_ucb_cost_balance_100checkpoints	Earlier design; not substituted for the paper-specific six-policy archive.
budget_projected_gain_hybrid, hybrid_b1, hybrid_b50, PORT-AS_cost_aware_log_upload, PORT-AS_cost_aware_upload, incomplete upload copy	Development/hybrid runs outside the evaluated six-policy study; excluded.
design1_raw_metric_reconstruction, design2_score_matrix_reconstruction, paper_query_svd_design2, dl_design2_qin2025, cost_aware_predicted_runtime_design2	Earlier experiments and alternative formulations. Their existence does not establish controlled baseline comparisons in the paper. Excluded from reported results.
Generic audit and runtime_audit	Earlier-design audits; not mislabeled as the paper's final runtime-probe or plotting implementation.
Early MolAS_AL literature/pilot reports, top-level library catalog, docs, scripts/fetch_github_meta.py, data/repositories.yml	Background/development materials rather than dependencies of the reported study.
deliverables	Earlier research report, abstract, poster-planning materials; not final paper artifacts.
Earlier camera-ready folders, compilation scratch, auxiliary LaTeX files	Superseded drafts/build products; excluded.
Virtual environments, vendored unrelated libraries, caches, nested Git metadata, .DS_Store	Not copied into the publication source tree.

The audit includes the full accessible workspace's metadata, not a claim that every cloud-only file was read. A cloud placeholder contains a name and logical size but may not provide readable bytes. Missing files are explicitly identified in artifact manifests rather than represented by empty files.

Evidence boundaries

The acquisition implementation is supported by formula checks against saved records. A historical code commit and exact dependency lock were not archived with every reported run. The recovered downstream source and launch script add substantial protocol evidence, but launcher settings and saved filenames alone do not prove every executed training configuration. The repository documents these distinctions and does not claim a new end-to-end reproduction run.

The original acquisition report's three-seed coverage must not be confused with independent selector-seed replication. The archived downstream folders are seed0, and the manuscript describes that scope.