{ "created_utc": "2026-06-03T12:59:30.225444+00:00", "paper_reference": "/home/unix/yektefai/.cache/spectra/runs/conch_generalizability_audit_20260603/source/CONCH/README.md", "model": "MahmoodLab/conch conch_ViT-B-16, source commit 141cc09c7d4ff33d8eda562bd75169b457f71a62", "dataset": "CRC-VAL-HE-7K from 1aurent/NCT-CRC-HE mirror of Zenodo 10.5281/zenodo.1214456; NCT-CRC-HE-100K streamed only for class-balanced prospective reference features", "target_scoring_status": "completed", "duplicate_policy": "exact MD5 and 64-bit dHash diagnostics computed; exact duplicate target images are retained only if unique in CRC-VAL, otherwise flagged before scoring", "axes": { "nct_color_texture_proximity": { "axis_col": "axis_nct_color_texture_similarity", "level_col": "level_nct_color_texture_proximity", "direction": "higher score means more similar to a class-balanced NCT-CRC-HE-100K color/texture reference; SPC is read as level increases and similarity decreases", "feature_family": "RGB/HSV/intensity histograms, tissue fraction, and edge/entropy features", "primary": true, "status": "frozen_before_target_scoring", "unit": "224x224 H&E ROI image patch", "n_levels": 6, "level_rule": "equal-frequency quantile bins over CRC-VAL by axis_nct_color_texture_similarity; level 1 is highest similarity", "threshold_quantiles_ascending": [ -663.7870483398438, -6.185784101486206, -4.49544095993042, -3.3102675676345825, -2.6528658866882324, -2.2217737436294556, -1.3983206748962402 ], "level_summary_before_scoring": [ { "level": 1, "n": 1197, "axis_mean": -1.9807237386703491, "axis_min": -2.2217416763305664, "axis_max": -1.3983206748962402, "label_counts": { "ADI": 12, "BACK": 3, "DEB": 69, "LYM": 135, "MUC": 161, "MUS": 118, "NORM": 261, "STR": 123, "TUM": 315 } }, { "level": 2, "n": 1197, "axis_mean": -2.424286127090454, "axis_min": -2.6528658866882324, "axis_max": -2.2218058109283447, "label_counts": { "ADI": 35, "BACK": 19, "DEB": 47, "LYM": 92, "MUC": 287, "MUS": 96, "NORM": 218, "STR": 73, "TUM": 330 } }, { "level": 3, "n": 1196, "axis_mean": -2.9441776275634766, "axis_min": -3.310204267501831, "axis_max": -2.6533102989196777, "label_counts": { "ADI": 46, "BACK": 32, "DEB": 27, "LYM": 90, "MUC": 339, "MUS": 151, "NORM": 154, "STR": 68, "TUM": 289 } }, { "level": 4, "n": 1197, "axis_mean": -3.856531858444214, "axis_min": -4.49544095993042, "axis_max": -3.310330867767334, "label_counts": { "ADI": 124, "BACK": 208, "DEB": 16, "LYM": 134, "MUC": 181, "MUS": 140, "NORM": 92, "STR": 121, "TUM": 181 } }, { "level": 5, "n": 1196, "axis_mean": -5.18813943862915, "axis_min": -6.184281349182129, "axis_max": -4.495564937591553, "label_counts": { "ADI": 679, "BACK": 160, "DEB": 1, "LYM": 173, "MUC": 41, "MUS": 24, "NORM": 15, "STR": 18, "TUM": 85 } }, { "level": 6, "n": 1197, "axis_mean": -33.85818099975586, "axis_min": -663.7870483398438, "axis_max": -6.187286853790283, "label_counts": { "ADI": 442, "BACK": 425, "DEB": 179, "LYM": 10, "MUC": 26, "MUS": 63, "NORM": 1, "STR": 18, "TUM": 33 } } ], "measured_similarity_decreases": true, "split_membership_uses_target_model": false, "split_membership_uses_heldout_labels": false, "predeclared_metric": "balanced accuracy by level, with accuracy, macro F1, per-class support, fixed baselines, and class-composition controls", "success_criterion": "claim-supporting only if target balanced accuracy shows an interpretable monotone, threshold, or localized degradation as similarity decreases, not materially explained by class composition or fixed baselines" }, "nct_color_proximity": { "axis_col": "axis_nct_color_similarity", "level_col": "level_nct_color_proximity", "direction": "higher score means more similar to the NCT reference in color/stain statistics only", "feature_family": "RGB/HSV/intensity summary and histogram features", "primary": false, "status": "frozen_before_target_scoring", "unit": "224x224 H&E ROI image patch", "n_levels": 6, "level_rule": "equal-frequency quantile bins over CRC-VAL by axis_nct_color_similarity; level 1 is highest similarity", "threshold_quantiles_ascending": [ -663.7427978515625, -5.299793243408203, -3.8028407096862793, -3.026949644088745, -2.42405104637146, -2.014843702316285, -1.2132200002670288 ], "level_summary_before_scoring": [ { "level": 1, "n": 1197, "axis_mean": -1.7763118743896484, "axis_min": -2.014312267303467, "axis_max": -1.2132200002670288, "label_counts": { "ADI": 5, "BACK": 1, "DEB": 63, "LYM": 136, "MUC": 198, "MUS": 114, "NORM": 260, "STR": 120, "TUM": 300 } }, { "level": 2, "n": 1197, "axis_mean": -2.209824800491333, "axis_min": -2.42405104637146, "axis_max": -2.0153751373291016, "label_counts": { "ADI": 31, "BACK": 19, "DEB": 51, "LYM": 80, "MUC": 261, "MUS": 96, "NORM": 231, "STR": 73, "TUM": 355 } }, { "level": 3, "n": 1196, "axis_mean": -2.698131799697876, "axis_min": -3.026761531829834, "axis_max": -2.42411470413208, "label_counts": { "ADI": 56, "BACK": 26, "DEB": 28, "LYM": 95, "MUC": 321, "MUS": 153, "NORM": 148, "STR": 65, "TUM": 304 } }, { "level": 4, "n": 1197, "axis_mean": -3.3934972286224365, "axis_min": -3.8028407096862793, "axis_max": -3.0271377563476562, "label_counts": { "ADI": 438, "BACK": 53, "DEB": 13, "LYM": 85, "MUC": 159, "MUS": 119, "NORM": 78, "STR": 105, "TUM": 147 } }, { "level": 5, "n": 1196, "axis_mean": -4.380153656005859, "axis_min": -5.290450096130371, "axis_max": -3.803377151489258, "label_counts": { "ADI": 435, "BACK": 298, "DEB": 5, "LYM": 194, "MUC": 68, "MUS": 45, "NORM": 23, "STR": 37, "TUM": 91 } }, { "level": 6, "n": 1197, "axis_mean": -33.34444808959961, "axis_min": -663.7427978515625, "axis_max": -5.309136390686035, "label_counts": { "ADI": 373, "BACK": 450, "DEB": 179, "LYM": 44, "MUC": 28, "MUS": 65, "NORM": 1, "STR": 21, "TUM": 36 } } ], "measured_similarity_decreases": true, "split_membership_uses_target_model": false, "split_membership_uses_heldout_labels": false, "predeclared_metric": "balanced accuracy by level, with accuracy, macro F1, per-class support, fixed baselines, and class-composition controls", "success_criterion": "claim-supporting only if target balanced accuracy shows an interpretable monotone, threshold, or localized degradation as similarity decreases, not materially explained by class composition or fixed baselines" }, "nct_tissue_fraction_proximity": { "axis_col": "axis_nct_tissue_fraction_similarity", "level_col": "level_nct_tissue_fraction_proximity", "direction": "higher score means tissue/white/dark/saturation fractions closer to NCT reference median", "feature_family": "tissue, white, dark, and high-saturation fractions", "primary": false, "status": "frozen_before_target_scoring", "unit": "224x224 H&E ROI image patch", "n_levels": 6, "level_rule": "equal-frequency quantile bins over CRC-VAL by axis_nct_tissue_fraction_similarity; level 1 is highest similarity", "threshold_quantiles_ascending": [ -1.3519560663221, -0.5370219781274885, -0.2329905217522512, -0.126035891873561, -0.07572624918340282, -0.05588390117843652, -0.0017131845275781992 ], "level_summary_before_scoring": [ { "level": 1, "n": 1197, "axis_mean": -0.0412261724727763, "axis_min": -0.05588324195122028, "axis_max": -0.0017131845275781992, "label_counts": { "ADI": 71, "BACK": 143, "DEB": 77, "LYM": 4, "MUC": 441, "MUS": 147, "NORM": 43, "STR": 122, "TUM": 149 } }, { "level": 2, "n": 1197, "axis_mean": -0.06518621235445655, "axis_min": -0.07572624918340282, "axis_max": -0.05588456040565276, "label_counts": { "ADI": 75, "BACK": 360, "DEB": 33, "LYM": 5, "MUC": 347, "MUS": 136, "NORM": 52, "STR": 74, "TUM": 115 } }, { "level": 3, "n": 1196, "axis_mean": -0.09508928347085352, "axis_min": -0.12600927531452527, "axis_max": -0.07577552223254742, "label_counts": { "ADI": 6, "BACK": 322, "DEB": 41, "LYM": 51, "MUC": 123, "MUS": 139, "NORM": 179, "STR": 92, "TUM": 243 } }, { "level": 4, "n": 1197, "axis_mean": -0.17152745621365806, "axis_min": -0.2329905217522512, "axis_max": -0.12606250843259678, "label_counts": { "ADI": 4, "BACK": 17, "DEB": 14, "LYM": 218, "MUC": 82, "MUS": 130, "NORM": 332, "STR": 83, "TUM": 317 } }, { "level": 5, "n": 1196, "axis_mean": -0.35005610840966705, "axis_min": -0.536730652498139, "axis_max": -0.23314876669746273, "label_counts": { "ADI": 204, "BACK": 5, "DEB": 6, "LYM": 352, "MUC": 41, "MUS": 33, "NORM": 134, "STR": 47, "TUM": 374 } }, { "level": 6, "n": 1197, "axis_mean": -0.8232089180332494, "axis_min": -1.3519560663221, "axis_max": -0.537313303756838, "label_counts": { "ADI": 978, "DEB": 168, "LYM": 4, "MUC": 1, "MUS": 7, "NORM": 1, "STR": 3, "TUM": 35 } } ], "measured_similarity_decreases": true, "split_membership_uses_target_model": false, "split_membership_uses_heldout_labels": false, "predeclared_metric": "balanced accuracy by level, with accuracy, macro F1, per-class support, fixed baselines, and class-composition controls", "success_criterion": "claim-supporting only if target balanced accuracy shows an interpretable monotone, threshold, or localized degradation as similarity decreases, not materially explained by class composition or fixed baselines" }, "nct_texture_complexity_proximity": { "axis_col": "axis_nct_texture_complexity_similarity", "level_col": "level_nct_texture_complexity_proximity", "direction": "higher score means edge and intensity-entropy statistics closer to NCT reference median", "feature_family": "edge mean/std, gray std, gray entropy", "primary": false, "status": "frozen_before_target_scoring", "unit": "224x224 H&E ROI image patch", "n_levels": 6, "level_rule": "equal-frequency quantile bins over CRC-VAL by axis_nct_texture_complexity_similarity; level 1 is highest similarity", "threshold_quantiles_ascending": [ -3.0238595986715215, -1.5056876030814872, -0.4763609921835658, -0.3312275347553513, -0.22168787981848512, -0.11897188624419303, -0.006895138046239978 ], "level_summary_before_scoring": [ { "level": 1, "n": 1197, "axis_mean": -0.06871213827735315, "axis_min": -0.11896121132312383, "axis_max": -0.006895138046239978, "label_counts": { "ADI": 27, "DEB": 77, "LYM": 80, "MUC": 316, "MUS": 163, "NORM": 16, "STR": 133, "TUM": 385 } }, { "level": 2, "n": 1197, "axis_mean": -0.17002409779046082, "axis_min": -0.22168787981848512, "axis_max": -0.1189825611652622, "label_counts": { "ADI": 32, "DEB": 73, "LYM": 120, "MUC": 286, "MUS": 150, "NORM": 51, "STR": 79, "TUM": 406 } }, { "level": 3, "n": 1196, "axis_mean": -0.27564838915569956, "axis_min": -0.33117522595723753, "axis_max": -0.22174500107893977, "label_counts": { "ADI": 48, "DEB": 76, "LYM": 184, "MUC": 213, "MUS": 105, "NORM": 200, "STR": 71, "TUM": 299 } }, { "level": 4, "n": 1197, "axis_mean": -0.3930598592409681, "axis_min": -0.4763609921835658, "axis_max": -0.3312798435534651, "label_counts": { "ADI": 64, "DEB": 67, "LYM": 227, "MUC": 139, "MUS": 57, "NORM": 445, "STR": 68, "TUM": 130 } }, { "level": 5, "n": 1196, "axis_mean": -0.879775890651379, "axis_min": -1.505275898955185, "axis_max": -0.4764164847986502, "label_counts": { "ADI": 801, "BACK": 20, "DEB": 42, "LYM": 23, "MUC": 81, "MUS": 117, "NORM": 29, "STR": 70, "TUM": 13 } }, { "level": 6, "n": 1197, "axis_mean": -2.192245776761918, "axis_min": -3.0238595986715215, "axis_max": -1.5060993072077893, "label_counts": { "ADI": 366, "BACK": 827, "DEB": 4 } } ], "measured_similarity_decreases": true, "split_membership_uses_target_model": false, "split_membership_uses_heldout_labels": false, "predeclared_metric": "balanced accuracy by level, with accuracy, macro F1, per-class support, fixed baselines, and class-composition controls", "success_criterion": "claim-supporting only if target balanced accuracy shows an interpretable monotone, threshold, or localized degradation as similarity decreases, not materially explained by class composition or fixed baselines" } }, "outcome_informed_hypotheses": { "tum_tissue_fraction_confirmation": { "created_utc": "2026-06-03T13:25:21.785440+00:00", "status": "frozen_before_confirmation_target_scoring", "discovery_source": "CRC-VAL all-scored evidence showed TUM recall decreased across the predeclared NCT tissue-fraction proximity axis", "hypothesis": "For true colorectal adenocarcinoma epithelium ROI patches, CONCH zero-shot TUM recall decreases as prospective tissue/white/dark/saturation fraction similarity to the NCT reference decreases.", "confirmation_dataset": "NCT-CRC-HE-100K-NONORM streamed TUM shards from the public HF/Zenodo mirror", "inclusion_criteria": "label 8/TUM patches from three NONORM TUM shards; label defines the class-localized claim boundary, not the split levels", "axis_col": "axis_nct_tissue_fraction_similarity", "level_col": "level_nct_tissue_fraction_proximity", "level_rule": "six equal-frequency bins over the confirmation panel; level 1 is highest NCT tissue-fraction similarity and level 6 is lowest", "threshold_quantiles_ascending": [ -1.0668714124515266, -0.5995464715663411, -0.27277102653030927, -0.13077265411993577, -0.07603596031594981, -0.0552486357712001, -0.001742380475234945 ], "level_summary_before_scoring": [ { "level": 1, "n": 300, "axis_mean": -0.041712291016583455, "axis_min": -0.05518255849348028, "axis_max": -0.001742380475234945 }, { "level": 2, "n": 300, "axis_mean": -0.06470081222295941, "axis_min": -0.07598835159911965, "axis_max": -0.055261851226744055 }, { "level": 3, "n": 300, "axis_mean": -0.10078102736432884, "axis_min": -0.13072730308987887, "axis_max": -0.07605976467436487 }, { "level": 4, "n": 300, "axis_mean": -0.18679127091707304, "axis_min": -0.2727552997061788, "axis_max": -0.1308180051499927 }, { "level": 5, "n": 300, "axis_mean": -0.43971989120966887, "axis_min": -0.5995366144056653, "axis_max": -0.2728024801785702 }, { "level": 6, "n": 300, "axis_mean": -0.768699645623449, "axis_min": -1.0668714124515266, "axis_max": -0.59959575736972 } ], "split_membership_uses_target_model": false, "split_membership_uses_prediction_errors": false, "confirmation_sample_size": 1800, "success_criterion": "TUM recall shows a monotone, threshold, or large localized drop as level increases; fixed feature baselines and target CONCH are reported by the same frozen levels.", "limitations": "Confirmation data are fresh relative to CRC-VAL discovery but not an independent hospital/site cohort; NONORM patches share the public Kather/NCT source family." }, "tum_tissue_fraction_expansion": { "created_utc": "2026-06-03T14:36:00.238017+00:00", "status": "expanded_scored_and_analyzed", "parent_hypothesis": "tum_tissue_fraction_confirmation", "purpose": "Densify and expand the same prospective TUM tissue-fraction axis after the six-level confirmation showed a low-similarity tail drop.", "axis_col": "axis_nct_tissue_fraction_similarity", "level_col": "level_nct_tissue_fraction_proximity", "level_rule": "10 equal-frequency bins over an unused expansion panel; level 1 is highest NCT tissue-fraction similarity and level 10 is lowest", "features": [ "tissue_fraction", "white_fraction", "dark_fraction", "saturation_high_fraction" ], "direction": "performance is assessed as axis_nct_tissue_fraction_similarity decreases and level increases", "confirmation_dataset": "Unused NCT-CRC-HE-100K-NONORM TUM rows from public TUM shards 00017-00020", "exclusion_rule": "Exclude all prior confirmation stream keys and exact decoded-image MD5 matches before target-model scoring.", "threshold_quantiles_ascending": [ -1.0710634722132715, -0.7256840602490031, -0.538265826258121, -0.3243584221237248, -0.18833336464338185, -0.1256348591839639, -0.09033238299010123, -0.07103837658121823, -0.05937209258738387, -0.044864270231615516, -0.0038558588596034255 ], "level_summary_before_scoring": [ { "level": 1, "n": 400, "axis_mean": -0.032156738191543174, "axis_min": -0.044813727746743914, "axis_max": -0.0038558588596034255, "tissue_fraction_mean": 0.9648188376913265, "white_fraction_mean": 0.03332389987244898, "dark_fraction_mean": 0.004395677216198979, "saturation_high_fraction_mean": 0.06252416493941328, "gray_mean_mean": 0.5730506090819836, "gray_std_mean": 0.15543254060670733, "gray_entropy_mean": 3.3206060326099394, "edge_mean_mean": 0.09301598191261291 }, { "level": 2, "n": 400, "axis_mean": -0.05249815196113376, "axis_min": -0.0593673719885826, "axis_max": -0.04486988606326791, "tissue_fraction_mean": 0.9606418407206633, "white_fraction_mean": 0.03750682597257653, "dark_fraction_mean": 0.002980159837372449, "saturation_high_fraction_mean": 0.04341064453125, "gray_mean_mean": 0.5956369005143642, "gray_std_mean": 0.14850821977481246, "gray_entropy_mean": 3.2514441430568697, "edge_mean_mean": 0.0958783174585551 }, { "level": 3, "n": 400, "axis_mean": -0.06491176392463456, "axis_min": -0.07099504183478178, "axis_max": -0.05937327273708419, "tissue_fraction_mean": 0.9583699378188776, "white_fraction_mean": 0.03963010204081632, "dark_fraction_mean": 0.0019863081951530612, "saturation_high_fraction_mean": 0.026604253029336736, "gray_mean_mean": 0.6253101798892021, "gray_std_mean": 0.13726445017382502, "gray_entropy_mean": 3.135664712190628, "edge_mean_mean": 0.09220357747748494 }, { "level": 4, "n": 400, "axis_mean": -0.0792351934053692, "axis_min": -0.09027883708378559, "axis_max": -0.07105694861540529, "tissue_fraction_mean": 0.9439887595663266, "white_fraction_mean": 0.05335698341836734, "dark_fraction_mean": 0.003259177694515306, "saturation_high_fraction_mean": 0.046797522720025506, "gray_mean_mean": 0.616130146831274, "gray_std_mean": 0.14342215364798905, "gray_entropy_mean": 3.1603490328788757, "edge_mean_mean": 0.0893395973648876 }, { "level": 5, "n": 400, "axis_mean": -0.10685911272265092, "axis_min": -0.1255533129635206, "axis_max": -0.0903680802609784, "tissue_fraction_mean": 0.9144078344228316, "white_fraction_mean": 0.08099803691007652, "dark_fraction_mean": 0.006222048389668366, "saturation_high_fraction_mean": 0.06520931321747449, "gray_mean_mean": 0.6111874159425497, "gray_std_mean": 0.15934387601912023, "gray_entropy_mean": 3.281589385867119, "edge_mean_mean": 0.09117200749926269 }, { "level": 6, "n": 400, "axis_mean": -0.15345025465881343, "axis_min": -0.18804276059605826, "axis_max": -0.1257164054044072, "tissue_fraction_mean": 0.877804179288903, "white_fraction_mean": 0.11509860291772961, "dark_fraction_mean": 0.010056501116071427, "saturation_high_fraction_mean": 0.06885787029655611, "gray_mean_mean": 0.6289397599548101, "gray_std_mean": 0.16653991650789976, "gray_entropy_mean": 3.2811660021543503, "edge_mean_mean": 0.08731644692830741 }, { "level": 7, "n": 400, "axis_mean": -0.24530111849651184, "axis_min": -0.3240875201986453, "axis_max": -0.1887692707143675, "tissue_fraction_mean": 0.8313740632971939, "white_fraction_mean": 0.16137720224808674, "dark_fraction_mean": 0.02130889269770408, "saturation_high_fraction_mean": 0.09912194076849491, "gray_mean_mean": 0.6255396398156882, "gray_std_mean": 0.17740710940212012, "gray_entropy_mean": 3.293075656294823, "edge_mean_mean": 0.08087569048628211 }, { "level": 8, "n": 400, "axis_mean": -0.4344567379494981, "axis_min": -0.5382185353079448, "axis_max": -0.32499052661557776, "tissue_fraction_mean": 0.8480972476881377, "white_fraction_mean": 0.14100884885204082, "dark_fraction_mean": 0.10438386878188774, "saturation_high_fraction_mean": 0.31991724131058674, "gray_mean_mean": 0.5108736436069011, "gray_std_mean": 0.19790371922776104, "gray_entropy_mean": 3.326561543345451, "edge_mean_mean": 0.06901142910122872 }, { "level": 9, "n": 400, "axis_mean": -0.6397387986207269, "axis_min": -0.7256807463606972, "axis_max": -0.538454990058826, "tissue_fraction_mean": 0.9078415477519133, "white_fraction_mean": 0.07720757932079081, "dark_fraction_mean": 0.24455551458864797, "saturation_high_fraction_mean": 0.5909424326371173, "gray_mean_mean": 0.3819294899702072, "gray_std_mean": 0.2049444031342864, "gray_entropy_mean": 3.2194570595026017, "edge_mean_mean": 0.058612208594568074 }, { "level": 10, "n": 400, "axis_mean": -0.8467361235586389, "axis_min": -1.0710634722132715, "axis_max": -0.7257138852437561, "tissue_fraction_mean": 0.9636960399394132, "white_fraction_mean": 0.027579520089285717, "dark_fraction_mean": 0.38806052694515303, "saturation_high_fraction_mean": 0.7993637396364797, "gray_mean_mean": 0.28601037822663783, "gray_std_mean": 0.1644209673628211, "gray_entropy_mean": 2.872992829978466, "edge_mean_mean": 0.05007558848010376 } ], "split_membership_uses_target_model": false, "split_membership_uses_prediction_errors": false, "split_membership_uses_target_confidence": false, "confirmation_sample_size": 4000, "success_criterion": "The expanded panel supports a monotone, threshold, or practically large localized TUM-recall drop in the lowest similarity tail, with fixed feature baselines reported on the same frozen levels.", "limitations": "Still same public Kather/NCT source family rather than an independent hospital/site cohort.", "analysis_artifacts": [ "artifacts/tum_expansion/analysis.json", "artifacts/tum_expansion/expanded_metrics.json", "artifacts/tum_expansion/expanded_finding_report.md" ], "validity_decision": "valid_for_localized_public_crc_tum_tail_boundary__weak_for_independent_site_generalization", "target_curve_by_level": [ { "level": 1, "n": 400, "tum_recall": 0.875 }, { "level": 2, "n": 400, "tum_recall": 0.95 }, { "level": 3, "n": 400, "tum_recall": 0.9425 }, { "level": 4, "n": 400, "tum_recall": 0.9275 }, { "level": 5, "n": 400, "tum_recall": 0.9125 }, { "level": 6, "n": 400, "tum_recall": 0.89 }, { "level": 7, "n": 400, "tum_recall": 0.83 }, { "level": 8, "n": 400, "tum_recall": 0.64 }, { "level": 9, "n": 400, "tum_recall": 0.7 }, { "level": 10, "n": 400, "tum_recall": 0.7025 } ], "main_contrast_levels_1_7_minus_8_10": { "difference": 0.22309523809523812, "ci95": [ 0.19455063553739568, 0.25163984065308054 ], "z": 17.50545536021469, "p_two_sided_normal": 1.3018178732663923e-68 } } }, "primary_claim_axis": "tum_tissue_fraction_confirmation", "closure_gate": { "status": "claim_valid_with_boundary", "closed_utc": "2026-06-03T13:32:29.935549+00:00", "finding": "Localized threshold degradation: CONCH zero-shot TUM recall remains high for TUM ROI patches with high-to-mid NCT tissue-fraction similarity, but drops in the low-similarity tail.", "claim_boundary": "Public CRC ROI tile classification using CONCH CRC prompts; strongest support is for true TUM / colorectal adenocarcinoma epithelium patches under tissue/white/dark/saturation fraction shift. Not a hospital/site, WSI, non-CRC, or all-class generalizability claim.", "primary_axis_validity": { "prospective": true, "frozen_before_confirmation_target_scoring": true, "measured_order_decreases": true, "confirmation_levels": 6, "confirmation_n_per_level": 300, "densified_deciles": 10, "fixed_baselines_run": true, "label_composition_confound_for_primary_local_claim": "controlled by constant true TUM label in confirmation panel; label defines class-local claim boundary, not level membership", "externality_limit": "confirmation is fresh relative to CRC-VAL discovery but shares the public Kather/NCT source family and is not an independent site cohort" }, "key_numbers": { "crc_val_overall_conch_balanced_accuracy": 0.785362682130508, "crc_val_overall_conch_accuracy": 0.7926183844011142, "crc_val_tum_six_level_recall_discovery": [ 0.7785234899328859, 0.6521739130434783, 0.6296296296296297, 0.5741324921135647, 0.39037433155080214, 0.11428571428571428 ], "nonorm_tum_confirmation_six_level_recall": [ 0.9033333333333333, 0.93, 0.93, 0.8833333333333333, 0.6966666666666667, 0.73 ], "nonorm_tum_confirmation_decile_recall": [ 0.8888888888888888, 0.9222222222222223, 0.9444444444444444, 0.9055555555555556, 0.9444444444444444, 0.8666666666666667, 0.85, 0.6777777777777778, 0.7333333333333333, 0.7222222222222222 ], "nonorm_tum_confirmation_centroid_baseline_six_level_recall": [ 0.53, 0.22333333333333333, 0.27, 0.19666666666666666, 0.056666666666666664, 0.0 ], "nonorm_tum_confirmation_logistic_baseline_six_level_recall": [ 0.35333333333333333, 0.15333333333333332, 0.14, 0.023333333333333334, 0.0, 0.0 ] } }, "negative_or_nonprimary_axes": { "nct_color_texture_proximity": { "status": "not_primary_due_to_class_composition_confounding", "target_bacc_by_level": [ 0.774364150956012, 0.7777338513768242, 0.7958834674632657, 0.7683016467797807, 0.7663481140247389, 0.6582762900694766 ], "target_acc_by_level": [ 0.7451963241436925, 0.7644110275689223, 0.7775919732441472, 0.7518796992481203, 0.8687290969899666, 0.847953216374269 ], "interpretation": "balanced accuracy decreases in the low-similarity tail, but raw accuracy rises because low-similarity levels are dominated by easy ADI/BACK/DEB composition; class-adjusted and stable-class checks do not support this as the primary broad axis." } } }