RESULTS_MAP.md

Manuscript results map

This map links every numbered manuscript table and figure in the supplied LaTeX source to the curated evidence. Verified means the rounded manuscript values were checked against the stored CSV/JSON records.

Tables

Manuscript label	Content	Primary repository source	Status and notes
tab:dataset_overview	Nine scene dimensions, bands, classes, labelled support, role	results/main_tables/dataset_registry.csv	Verified. Trento is the primary holdout.
tab:frozen_pipeline	Shared and scene-specific primary settings	configs/primary.yaml; results/provenance/frozen_results_registry.json	Verified for all nine scenes.
tab:primary_results	Full/support region counts and Graph-RMS OA/BA/NMI/ARI	results/main_tables/master_results_tidy.csv; results/supporting_results/oversegmentation.csv	Verified. Nine-scene means 0.6358/0.5319/0.7245/0.6086.
tab:classical_complete	Graph-RMS plus four true-K classical baselines	results/main_tables/master_results_tidy.csv	Verified. Baseline entries are five-seed means.
tab:classical_statistics	Average ranks, Friedman, Holm-adjusted Wilcoxon	results/main_tables/classical_average_ranks.csv; friedman.csv; wilcoxon.csv	Verified. Inferential family excludes no-K and adapted modern methods.
tab:no_k_complete	Graph-RMS, Leiden, and fixed-default HDBSCAN	results/main_tables/no_k_aggregate.csv; Graph-RMS rows in master_results_tidy.csv	Verified. Leiden has five evaluation seeds; HDBSCAN one declared run.
tab:no_k_statistics	Win counts and exact paired Wilcoxon summaries	results/supporting_results/no_k_per_run.csv; results/main_tables/no_k_aggregate.csv	Verified against manuscript values.
tab:trento_configuration	Frozen primary Trento settings and region counts	results/per_dataset/trento/summary.json; configs/primary.yaml	Verified: T100, radius 0.20, tau 0.75, beta 0.5, 213 fine, 68 full, 41 on support.
tab:trento_comparison	Trento metrics, bootstrap CI, SLIC-KMeans, Leiden, HDBSCAN	master_results_tidy.csv; bootstrap_ci.csv; no_k_aggregate.csv	Verified. Bootstrap block size 16, 300 replicates.
tab:trento_diagnostics	AMI, homogeneity, completeness, purity, class coverage	results/supporting_results/graph_rms_information_metrics.csv; oversegmentation.csv	Verified.
tab:region_granularity	K, full/support region counts and ratios	results/supporting_results/oversegmentation.csv	Verified.
tab:granularity_controls	Trento post-hoc K=6 and count-matched MiniBatch-KMeans	results/supporting_results/oracle_k.csv; count_matched_minibatch_aggregate.csv; graph_rms_information_metrics.csv	Verified. Count-matched dashes reflect fields omitted from the stored summary.
tab:ablation	Eight Pavia component variants	results/supporting_results/ablation_summary.csv; ablation_runs_raw.csv	Verified, including 189,178 regions without fine grouping.
tab:sensitivity_summary	Selected findings from 51 runs	results/supporting_results/sensitivity_runs_raw.csv; sensitivity_summary.csv	Verified. The full raw grid is retained.
tab:selector_audit	Mean oracle gaps and rule-ablation consequences	results/supporting_results/selector_oracle_gaps.csv; selector_rule_ablation.csv	Verified.
tab:auto_selector_summary	Eight-development-scene primary versus automatic means	results/automatic_selector/development_master_table.csv; development_rule_search.csv; development_lock.json	Verified. Global calibration used development-scene NMI/ARI; deployment did not use target labels/K.
tab:perturbation_results	Salinas-A, Indian Pines, and adaptive Pavia perturbations	results/supporting_results/robustness_perturbations.csv; perturbation_runs_frozen_threshold.csv; pavia_adaptive_robustness.csv	Verified. Pavia pixel-replacement trials are abstentions.
tab:pavia_partition_agreement	Pavia perturbation ARI versus frozen partition	results/supporting_results/pavia_adaptive_robustness.csv	Verified.
tab:cross_device_repeatability	CPU/GPU cluster counts and partition ARI	results/supporting_results/cross_device_repeatability.csv	Verified.
tab:runtime_memory_complete	Three-run runtime and peak process/GPU memory	results/supporting_results/runtime_memory.csv; profile_runs_raw.csv; results/provenance/environment.json	Verified. Hardware A100-SXM4-80GB.

Figures

Manuscript filename	Curated file	Numerical/data source	Status
figures/graph.pdf	figures/graph.pdf	Author-supplied workflow artwork; method equations and configs/primary.yaml	Present; one-page vector PDF with embedded fonts.
figures/fig2_representative_cluster_maps.png	figures/fig2_representative_cluster_maps.png	Frozen NPY partitions for KSC, HongHu, Trento; reference maps required only for regeneration	Present; copied without visual redesign from the supplied evidence figure.
figures/fig3_fragmentation.png	figures/fig3_fragmentation.png	results/supporting_results/oversegmentation.csv	Present.
figures/fig4_sensitivity.png	figures/fig4_sensitivity.png	results/supporting_results/sensitivity_runs_raw.csv	Present.
figures/repeatability_and_perturbation_matched_style.png	same path under figures/	cross_device_repeatability.csv; pavia_adaptive_robustness.csv	Present.
figures/fig6_runtime_memory.png	figures/fig6_runtime_memory.png	runtime_memory.csv	Present.

Per-scene selected partitions

Every primary scene has:

results/per_dataset/<dataset>/summary.json
results/per_dataset/<dataset>/endpoint_curve.csv
results/per_dataset/<dataset>/selected_labels.npy

NPY arrays contain the complete scene partition in the original H x W shape. The portable registry points to these summaries. The original evidence source paths and checksums remain documented in the provenance files.

Interpretation boundaries

• master_results_tidy.csv also preserves adapted DLSS/S2DL outputs. Those rows are historical diagnostics and are not used in the manuscript's main superiority or inferential claims.
• The no-K table must remain separate from the true-K classical table.
• The automatic Trento audit must remain separate from the independent primary Trento holdout result.
• Full-scene region counts must not be replaced with labelled-support counts.