diff --git a/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/within_label_marker_check.csv b/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/within_label_marker_check.csv new file mode 100644 index 0000000000000000000000000000000000000000..195a4c15135706cd8c158b985489b327eb77fd5b --- /dev/null +++ b/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/within_label_marker_check.csv @@ -0,0 +1,18 @@ +side,arm,thin,t_cells,pt_label_cells,pt_label_pct_of_T,cycling_rule_pct_of_T,pct_of_pt_label_passing_cycling_rule,pct_of_nonpt_T_passing_cycling_rule,n_cycling_rule_T +observed native,,1.0,61162,444,0.7259409437232268,0.3776854909911383,37.387387387387385,0.10705227444909253,231 +generated,zs__raw,,8631,243,2.815432742440042,0.3707565751361372,11.522633744855968,0.047687172150691466,32 +observed thinned to arm,zs__raw,0.461,61162,444,0.7259409437232268,0.26323534220594486,29.504504504504503,0.04940874205342732,161 +generated,lora_n5__raw,,12396,224,1.8070345272668602,0.4356243949661181,22.767857142857142,0.02464673020046007,54 +observed thinned to arm,lora_n5__raw,1.0,61162,444,0.7259409437232268,0.3776854909911383,37.387387387387385,0.10705227444909253,231 +generated,scratch_n5__raw,,13611,135,0.9918448313863787,0.20571596502828596,18.51851851851852,0.02226179875333927,28 +observed thinned to arm,scratch_n5__raw,1.0,61162,444,0.7259409437232268,0.3776854909911383,37.387387387387385,0.10705227444909253,231 +generated,real_lora__raw,,14147,153,1.0815013783841096,0.31102000424118187,26.143790849673202,0.028583678719451194,44 +observed thinned to arm,real_lora__raw,0.934,61162,444,0.7259409437232268,0.3629704718616134,36.93693693693694,0.09552356796995949,222 +generated,real_full__raw,,13435,180,1.339784145887607,0.45403796055080015,27.77777777777778,0.08298755186721991,61 +observed thinned to arm,real_full__raw,0.882,61162,444,0.7259409437232268,0.3466204506065858,35.585585585585584,0.08893573569616918,212 +generated,scratch_full__raw,,13443,110,0.8182697314587518,0.15621513055121625,16.363636363636363,0.022500562514062853,21 +observed thinned to arm,scratch_full__raw,0.953,61162,444,0.7259409437232268,0.3678754782381217,37.387387387387385,0.09717052603840706,225 +generated,lora_recipe__raw,,13851,177,1.2778860732077106,0.33932568045628475,24.293785310734464,0.029252596167909903,47 +observed thinned to arm,lora_recipe__raw,0.923,61162,444,0.7259409437232268,0.3597004676106079,36.711711711711715,0.0938766099015119,220 +generated,scratch_recipe__raw,,13815,98,0.7093738689829895,0.16648570394498732,23.46938775510204,0.0,23 +observed thinned to arm,scratch_recipe__raw,0.927,61162,444,0.7259409437232268,0.35806546548510515,36.711711711711715,0.09222965183306434,219 diff --git a/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/within_spearman.csv b/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/within_spearman.csv new file mode 100644 index 0000000000000000000000000000000000000000..e78d7313c0f3ebddcc8c143b7a0f7cfc2c293b28 --- /dev/null +++ b/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/within_spearman.csv @@ -0,0 +1,113 @@ +comparison,kind,scratch_arm,pretrained_arm,state,role,informative,spearman_scratch,scratch_ci_lo,scratch_ci_hi,spearman_pretrained,pretrained_ci_lo,pretrained_ci_hi,diff_pretrained_minus_scratch,diff_ci_lo,diff_ci_hi,n_boot_valid,verdict +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,C1Q-high monocyte (reference),rare state,True,0.409314,-0.176266,0.825466,0.152335,-0.404309,0.669611,-0.256979,-0.764943,0.111725,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,Cycling T/NK,rare state,True,-0.107277,-0.620664,0.457555,0.178922,-0.257259,0.585544,0.286199,-0.296023,0.788569,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,Cycling activated T (CD38/HLA-DR),rare state,True,-0.54902,-0.861401,-0.061166,-0.245701,-0.674488,0.31247,0.303319,-0.122906,0.727918,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,Proliferating T (label),rare state,True,0.431373,-0.046304,0.754077,0.465686,-0.036174,0.823722,0.034314,-0.583906,0.630467,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,Plasmablast (label),rare state,True,0.38619,-0.172637,0.776613,0.275865,-0.287344,0.757497,-0.110325,-0.657444,0.419166,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,Plasmablast (markers),rare state,True,0.42205,-0.081131,0.751902,0.100737,-0.453442,0.664051,-0.321313,-0.86052,0.184232,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,MS1 monocyte (IL1R2+),rare state,True,0.781863,0.451369,0.940858,0.83784,0.520542,0.967269,0.055978,-0.236997,0.375034,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,Cycling monocyte (Ye 2026),rare state,True,0.177777,-0.343521,0.630829,0.219214,-0.396144,0.674118,0.041436,-0.783896,0.81648,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,ISG-high monocyte (positive control),positive control,True,-0.218137,-0.76152,0.351511,-0.091968,-0.536932,0.422605,0.126169,-0.549977,0.785217,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,ISG-high T,rare state,True,0.007371,-0.523075,0.565881,0.146826,-0.342649,0.597525,0.139455,-0.757775,0.995997,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,DN2 / ABC B,rare state,True,0.32598,-0.263048,0.770043,0.375,-0.18,0.755858,0.04902,-0.314766,0.40256,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,HSPC,rare state,True,0.134742,-0.36178,0.600455,0.102639,-0.459899,0.648495,-0.032103,-0.781511,0.770415,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,Cytotoxic CD4 T,rare state,True,-0.394608,-0.773109,0.149594,0.742647,0.372272,0.945075,1.137255,0.461466,1.603059,2000,pretrained tracks donors better +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,Granulocyte-like (LDG),rare state,True,0.403464,-0.129541,0.774269,0.488552,-0.001575,0.80894,0.085088,-0.548289,0.702574,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,Platelet signal in monocytes (negative control),negative control,True,-0.375,-0.814138,0.238491,0.218137,-0.270962,0.629771,0.593137,-0.004963,1.145422,2000,not significant +5 pairs: scratch vs pretrained LoRA,main,scratch_n5__raw,lora_n5__raw,MEAN over 13 informative rare states,aggregate,True,0.186863,,,0.295428,,,0.108565,-0.079967,0.293607,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,C1Q-high monocyte (reference),rare state,True,0.508601,-0.020174,0.825016,0.490761,-0.06768,0.832847,-0.01784,-0.403535,0.34011,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,Cycling T/NK,rare state,True,-0.171229,-0.667974,0.377413,0.181484,-0.41307,0.690599,0.352713,-0.276513,0.884306,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,Cycling activated T (CD38/HLA-DR),rare state,True,-0.568627,-0.861434,-0.0746,-0.27451,-0.729537,0.268413,0.294118,-0.223623,0.806006,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,Proliferating T (label),rare state,True,0.487745,-0.034179,0.837337,0.485294,-0.051114,0.77205,-0.002451,-0.607195,0.514346,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,Plasmablast (label),rare state,True,0.432004,-0.098627,0.829081,0.013539,-0.5988,0.598845,-0.418465,-1.041086,0.283147,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,Plasmablast (markers),rare state,True,0.199877,-0.324989,0.609872,0.12285,-0.48229,0.64234,-0.077027,-0.739008,0.563217,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,MS1 monocyte (IL1R2+),rare state,True,0.91355,0.719944,0.977695,0.811359,0.552959,0.939199,-0.102191,-0.371746,0.071832,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,Cycling monocyte (Ye 2026),rare state,True,0.190203,-0.364751,0.63276,-0.357115,-0.723796,0.152473,-0.547318,-1.173526,0.25681,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,ISG-high monocyte (positive control),positive control,True,0.645399,0.19796,0.873457,0.69042,0.264595,0.909364,0.045021,-0.28218,0.396471,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,ISG-high T,rare state,True,0.411942,-0.08371,0.749667,0.71401,0.35352,0.897252,0.302068,-0.144566,0.774303,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,DN2 / ABC B,rare state,True,0.490196,-0.026274,0.890999,0.375,-0.158324,0.765892,-0.115196,-0.636944,0.365785,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,HSPC,rare state,True,-0.068036,-0.565619,0.466769,0.237823,-0.285065,0.660663,0.305859,-0.182755,0.7394,1999,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,Cytotoxic CD4 T,rare state,True,0.578431,0.06475,0.8933,0.761673,0.376274,0.937894,0.183242,-0.14298,0.59168,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,Granulocyte-like (LDG),rare state,True,0.158444,-0.378294,0.642618,0.250709,-0.228443,0.668442,0.092266,-0.351515,0.567024,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,Platelet signal in monocytes (negative control),negative control,True,0.54902,-0.038416,0.876418,0.004902,-0.520207,0.543458,-0.544118,-1.035795,0.015168,2000,not significant +58 pairs: scratch vs pretrained LoRA,main,scratch_full__raw,real_lora__raw,MEAN over 13 informative rare states,aggregate,True,0.274085,,,0.293298,,,0.019214,-0.12139,0.186036,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,C1Q-high monocyte (reference),rare state,True,0.508601,-0.020174,0.825016,0.568976,0.035869,0.893465,0.060375,-0.324048,0.475662,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,Cycling T/NK,rare state,True,-0.171229,-0.667974,0.377413,0.465686,0.032164,0.72426,0.636915,0.007546,1.111006,2000,pretrained tracks donors better +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,Cycling activated T (CD38/HLA-DR),rare state,True,-0.568627,-0.861434,-0.0746,0.014706,-0.493812,0.56469,0.583333,-0.057005,1.15492,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,Proliferating T (label),rare state,True,0.487745,-0.034179,0.837337,0.735294,0.25333,0.93009,0.247549,-0.377179,0.797129,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,Plasmablast (label),rare state,True,0.432004,-0.098627,0.829081,0.373006,-0.171561,0.786727,-0.058998,-0.628844,0.570228,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,Plasmablast (markers),rare state,True,0.199877,-0.324989,0.609872,0.077348,-0.470003,0.608514,-0.122529,-0.720601,0.462367,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,MS1 monocyte (IL1R2+),rare state,True,0.91355,0.719944,0.977695,0.909871,0.706177,0.975186,-0.003679,-0.16399,0.152412,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,Cycling monocyte (Ye 2026),rare state,True,0.190203,-0.364751,0.63276,0.154134,-0.422447,0.644451,-0.036069,-0.747053,0.636275,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,ISG-high monocyte (positive control),positive control,True,0.645399,0.19796,0.873457,0.367647,-0.127738,0.713936,-0.277752,-0.720404,0.161919,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,ISG-high T,rare state,True,0.411942,-0.08371,0.749667,0.448837,-0.052807,0.77361,0.036895,-0.585046,0.637833,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,DN2 / ABC B,rare state,True,0.490196,-0.026274,0.890999,0.485294,-0.139055,0.898561,-0.004902,-0.701558,0.686296,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,HSPC,rare state,True,-0.068036,-0.565619,0.466769,0.062972,-0.408687,0.540971,0.131008,-0.449664,0.703284,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,Cytotoxic CD4 T,rare state,True,0.578431,0.06475,0.8933,0.815727,0.505528,0.946034,0.237296,-0.154074,0.719686,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,Granulocyte-like (LDG),rare state,True,0.158444,-0.378294,0.642618,0.458021,-0.091978,0.836419,0.299578,-0.171477,0.780701,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,Platelet signal in monocytes (negative control),negative control,True,0.54902,-0.038416,0.876418,0.171674,-0.397619,0.681118,-0.377346,-1.058418,0.271872,2000,not significant +58 pairs: scratch vs pretrained full FT,main,scratch_full__raw,real_full__raw,MEAN over 13 informative rare states,aggregate,True,0.274085,,,0.428452,,,0.154367,-0.008216,0.31307,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,C1Q-high monocyte (reference),rare state,True,0.489678,0.036009,0.806611,0.547913,0.020978,0.87288,0.058236,-0.387577,0.538573,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,Cycling T/NK,rare state,True,0.221952,-0.297923,0.691203,0.389706,-0.098504,0.747816,0.167753,-0.461223,0.728025,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,Cycling activated T (CD38/HLA-DR),rare state,True,-0.189893,-0.618376,0.422991,0.066176,-0.470271,0.648544,0.256069,-0.280013,0.822484,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,Proliferating T (label),rare state,True,0.620098,0.158294,0.837907,0.639706,0.19552,0.875776,0.019608,-0.37721,0.439247,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,Plasmablast (label),rare state,True,0.16718,-0.420097,0.69607,0.131531,-0.4926,0.670867,-0.035649,-0.745337,0.704692,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,Plasmablast (markers),rare state,True,0.192994,-0.327719,0.638266,0.321473,-0.265466,0.80691,0.128478,-0.485315,0.781547,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,MS1 monocyte (IL1R2+),rare state,True,0.896759,0.664217,0.972667,0.874696,0.627327,0.970077,-0.022064,-0.283503,0.217679,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,Cycling monocyte (Ye 2026),rare state,True,-0.032839,-0.609575,0.507255,0.261804,-0.291588,0.706862,0.294643,-0.431089,1.021391,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,ISG-high monocyte (positive control),positive control,True,0.312686,-0.263248,0.821483,0.523343,0.0037,0.833771,0.210657,-0.327734,0.80692,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,ISG-high T,rare state,True,0.441439,-0.073143,0.806311,0.526058,-0.035541,0.86381,0.084619,-0.464724,0.569055,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,DN2 / ABC B,rare state,True,0.594596,0.17649,0.869137,0.39951,-0.156523,0.841589,-0.195087,-0.75359,0.369643,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,HSPC,rare state,True,0.382202,-0.09303,0.700314,-0.161023,-0.59686,0.309649,-0.543225,-1.178511,0.23281,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,Cytotoxic CD4 T,rare state,True,0.392157,-0.202033,0.830373,0.845211,0.53701,0.982633,0.453055,-0.010331,1.00128,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,Granulocyte-like (LDG),rare state,True,0.464206,-0.082955,0.833167,0.33603,-0.199494,0.73937,-0.128176,-0.654505,0.447147,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,Platelet signal in monocytes (negative control),negative control,True,0.129412,-0.427902,0.659777,0.063725,-0.460237,0.577653,-0.065686,-0.767428,0.534753,2000,not significant +58 pairs + recipes: scratch vs pretrained LoRA,main,scratch_recipe__raw,lora_recipe__raw,MEAN over 13 informative rare states,aggregate,True,0.356964,,,0.398368,,,0.041405,-0.105909,0.217255,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,C1Q-high monocyte (reference),rare state,True,0.409314,-0.176266,0.825466,0.315467,-0.313989,0.69194,-0.093846,-0.802385,0.541012,1999,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,Cycling T/NK,rare state,True,-0.107277,-0.620664,0.457555,0.244022,-0.30131,0.718601,0.351299,-0.521522,1.108512,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,Cycling activated T (CD38/HLA-DR),rare state,True,-0.54902,-0.861401,-0.061166,-0.196561,-0.660121,0.392801,0.352459,-0.232785,0.974573,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,Proliferating T (label),rare state,True,0.431373,-0.046304,0.754077,0.286765,-0.265843,0.698446,-0.144608,-0.818505,0.53553,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,Plasmablast (label),rare state,True,0.38619,-0.172637,0.776613,0.161506,-0.421376,0.688283,-0.224684,-0.913885,0.482829,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,Plasmablast (markers),rare state,True,0.42205,-0.081131,0.751902,0.249856,-0.343736,0.710688,-0.172194,-0.766592,0.350218,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,MS1 monocyte (IL1R2+),rare state,True,0.781863,0.451369,0.940858,0.790422,0.437414,0.921487,0.008559,-0.341161,0.397112,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,Cycling monocyte (Ye 2026),rare state,True,0.177777,-0.343521,0.630829,0.387466,-0.221576,0.794537,0.209688,-0.583474,0.952905,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,ISG-high monocyte (positive control),positive control,True,-0.218137,-0.76152,0.351511,0.20098,-0.390353,0.71655,0.419118,0.02711,0.825833,2000,pretrained tracks donors better +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,ISG-high T,rare state,True,0.007371,-0.523075,0.565881,-0.123307,-0.719379,0.465917,-0.130678,-0.969434,0.744279,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,DN2 / ABC B,rare state,True,0.32598,-0.263048,0.770043,-0.068627,-0.588636,0.542418,-0.394608,-0.89009,0.208482,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,HSPC,rare state,True,0.134742,-0.36178,0.600455,0.427994,-0.107695,0.817319,0.293252,-0.511511,1.030454,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,Cytotoxic CD4 T,rare state,True,-0.394608,-0.773109,0.149594,0.262661,-0.234652,0.68543,0.657269,-0.056902,1.250828,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,Granulocyte-like (LDG),rare state,True,0.403464,-0.129541,0.774269,0.259428,-0.370754,0.73837,-0.144036,-0.780788,0.37775,2000,not significant +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,Platelet signal in monocytes (negative control),negative control,True,-0.375,-0.814138,0.238491,0.247549,-0.330033,0.725624,0.622549,0.030263,1.166367,2000,pretrained tracks donors better +reference: scratch 5 pairs vs pretrained zero-shot (no pairs),reference,scratch_n5__raw,zs__raw,MEAN over 13 informative rare states,aggregate,True,0.186863,,,0.230545,,,0.043682,-0.158806,0.225553,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,C1Q-high monocyte (reference),rare state,True,0.508601,-0.020174,0.825016,0.315467,-0.313989,0.69194,-0.193134,-0.631487,0.212557,1999,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,Cycling T/NK,rare state,True,-0.171229,-0.667974,0.377413,0.244022,-0.30131,0.718601,0.415251,-0.479639,1.239827,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,Cycling activated T (CD38/HLA-DR),rare state,True,-0.568627,-0.861434,-0.0746,-0.196561,-0.660121,0.392801,0.372067,-0.308963,1.07994,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,Proliferating T (label),rare state,True,0.487745,-0.034179,0.837337,0.286765,-0.265843,0.698446,-0.20098,-0.983922,0.63522,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,Plasmablast (label),rare state,True,0.432004,-0.098627,0.829081,0.161506,-0.421376,0.688283,-0.270498,-0.734806,0.204577,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,Plasmablast (markers),rare state,True,0.199877,-0.324989,0.609872,0.249856,-0.343736,0.710688,0.049979,-0.525362,0.581143,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,MS1 monocyte (IL1R2+),rare state,True,0.91355,0.719944,0.977695,0.790422,0.437414,0.921487,-0.123129,-0.443883,0.120046,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,Cycling monocyte (Ye 2026),rare state,True,0.190203,-0.364751,0.63276,0.387466,-0.221576,0.794537,0.197263,-0.519778,0.826515,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,ISG-high monocyte (positive control),positive control,True,0.645399,0.19796,0.873457,0.20098,-0.390353,0.71655,-0.444419,-1.083235,0.236437,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,ISG-high T,rare state,True,0.411942,-0.08371,0.749667,-0.123307,-0.719379,0.465917,-0.535249,-1.331125,0.353289,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,DN2 / ABC B,rare state,True,0.490196,-0.026274,0.890999,-0.068627,-0.588636,0.542418,-0.558824,-1.289289,0.346436,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,HSPC,rare state,True,-0.068036,-0.565619,0.466769,0.427994,-0.107695,0.817319,0.49603,-0.296885,1.153662,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,Cytotoxic CD4 T,rare state,True,0.578431,0.06475,0.8933,0.262661,-0.234652,0.68543,-0.31577,-0.794684,0.160358,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,Granulocyte-like (LDG),rare state,True,0.158444,-0.378294,0.642618,0.259428,-0.370754,0.73837,0.100984,-0.564993,0.71432,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,Platelet signal in monocytes (negative control),negative control,True,0.54902,-0.038416,0.876418,0.247549,-0.330033,0.725624,-0.301471,-0.971582,0.394139,2000,not significant +reference: scratch 58 pairs vs pretrained zero-shot (no pairs),reference,scratch_full__raw,zs__raw,MEAN over 13 informative rare states,aggregate,True,0.274085,,,0.230545,,,-0.043539,-0.295031,0.197929,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,C1Q-high monocyte (reference),rare state,True,0.489678,0.036009,0.806611,0.315467,-0.313989,0.69194,-0.17421,-0.709314,0.265254,1999,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,Cycling T/NK,rare state,True,0.221952,-0.297923,0.691203,0.244022,-0.30131,0.718601,0.02207,-0.728295,0.768438,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,Cycling activated T (CD38/HLA-DR),rare state,True,-0.189893,-0.618376,0.422991,-0.196561,-0.660121,0.392801,-0.006668,-0.616215,0.573105,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,Proliferating T (label),rare state,True,0.620098,0.158294,0.837907,0.286765,-0.265843,0.698446,-0.333333,-0.935793,0.34253,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,Plasmablast (label),rare state,True,0.16718,-0.420097,0.69607,0.161506,-0.421376,0.688283,-0.005674,-0.48833,0.517519,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,Plasmablast (markers),rare state,True,0.192994,-0.327719,0.638266,0.249856,-0.343736,0.710688,0.056862,-0.553615,0.662044,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,MS1 monocyte (IL1R2+),rare state,True,0.896759,0.664217,0.972667,0.790422,0.437414,0.921487,-0.106338,-0.484231,0.207588,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,Cycling monocyte (Ye 2026),rare state,True,-0.032839,-0.609575,0.507255,0.387466,-0.221576,0.794537,0.420305,-0.317737,1.146483,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,ISG-high monocyte (positive control),positive control,True,0.312686,-0.263248,0.821483,0.20098,-0.390353,0.71655,-0.111705,-0.901268,0.703001,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,ISG-high T,rare state,True,0.441439,-0.073143,0.806311,-0.123307,-0.719379,0.465917,-0.564746,-1.1204,0.005323,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,DN2 / ABC B,rare state,True,0.594596,0.17649,0.869137,-0.068627,-0.588636,0.542418,-0.663224,-1.200943,0.133641,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,HSPC,rare state,True,0.382202,-0.09303,0.700314,0.427994,-0.107695,0.817319,0.045792,-0.437369,0.455927,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,Cytotoxic CD4 T,rare state,True,0.392157,-0.202033,0.830373,0.262661,-0.234652,0.68543,-0.129496,-0.632175,0.382309,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,Granulocyte-like (LDG),rare state,True,0.464206,-0.082955,0.833167,0.259428,-0.370754,0.73837,-0.204778,-0.861137,0.469232,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,Platelet signal in monocytes (negative control),negative control,True,0.129412,-0.427902,0.659777,0.247549,-0.330033,0.725624,0.118137,-0.474911,0.705647,2000,not significant +reference: scratch 58 + recipes vs pretrained zero-shot (no pairs),reference,scratch_recipe__raw,zs__raw,MEAN over 13 informative rare states,aggregate,True,0.356964,,,0.230545,,,-0.126418,-0.336522,0.117174,2000,not significant diff --git a/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/within_state_table.csv b/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/within_state_table.csv new file mode 100644 index 0000000000000000000000000000000000000000..0dd5593d333c519481a118c4ad97a7e8cdf30c83 --- /dev/null +++ b/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/within_state_table.csv @@ -0,0 +1,16 @@ +state,role,obs_parent_cells,obs_native_state_cells,obs_thinned_state_cells_min,obs_thinned_state_cells_max,informative,obs_donors_with_state,obs_top_donor_share,obs_top_donor,gen_over_obs__zs__raw,abs_log2__zs__raw,spearman__zs__raw,gen_state_cells__zs__raw,gen_parent_cells__zs__raw,obs_thinned_state_cells__zs__raw,thin__zs__raw,gen_over_obs__lora_n5__raw,abs_log2__lora_n5__raw,spearman__lora_n5__raw,gen_state_cells__lora_n5__raw,gen_parent_cells__lora_n5__raw,obs_thinned_state_cells__lora_n5__raw,thin__lora_n5__raw,gen_over_obs__scratch_n5__raw,abs_log2__scratch_n5__raw,spearman__scratch_n5__raw,gen_state_cells__scratch_n5__raw,gen_parent_cells__scratch_n5__raw,obs_thinned_state_cells__scratch_n5__raw,thin__scratch_n5__raw,gen_over_obs__real_lora__raw,abs_log2__real_lora__raw,spearman__real_lora__raw,gen_state_cells__real_lora__raw,gen_parent_cells__real_lora__raw,obs_thinned_state_cells__real_lora__raw,thin__real_lora__raw,gen_over_obs__real_full__raw,abs_log2__real_full__raw,spearman__real_full__raw,gen_state_cells__real_full__raw,gen_parent_cells__real_full__raw,obs_thinned_state_cells__real_full__raw,thin__real_full__raw,gen_over_obs__scratch_full__raw,abs_log2__scratch_full__raw,spearman__scratch_full__raw,gen_state_cells__scratch_full__raw,gen_parent_cells__scratch_full__raw,obs_thinned_state_cells__scratch_full__raw,thin__scratch_full__raw,gen_over_obs__lora_recipe__raw,abs_log2__lora_recipe__raw,spearman__lora_recipe__raw,gen_state_cells__lora_recipe__raw,gen_parent_cells__lora_recipe__raw,obs_thinned_state_cells__lora_recipe__raw,thin__lora_recipe__raw,gen_over_obs__scratch_recipe__raw,abs_log2__scratch_recipe__raw,spearman__scratch_recipe__raw,gen_state_cells__scratch_recipe__raw,gen_parent_cells__scratch_recipe__raw,obs_thinned_state_cells__scratch_recipe__raw,thin__scratch_recipe__raw,closest_trained_arm,closest_scratch_arm,closest_pretrained_trained_arm +C1Q-high monocyte (reference),rare state,31471,1822,1590,1822,True,17,0.320527,Abd-S123,0.024986,5.322736,0.315467,12,9506,1590,0.416,1.07489,0.104188,0.152335,447,7183,1822,1.0,1.005146,0.007405,0.409314,357,6241,1791,0.893,1.053362,0.075002,0.490761,227,3891,1743,0.767,0.916807,0.125309,0.568976,208,4120,1733,0.73,0.466973,1.098588,0.508601,106,4036,1770,0.84,0.625773,0.676288,0.547913,136,3965,1725,0.729,0.5325,0.909147,0.489678,108,3598,1774,0.826,scratch_n5__raw,scratch_n5__raw,real_lora__raw +Cycling T/NK,rare state,71619,288,185,288,True,17,0.125,Abd-PS19,2.516672,1.331517,0.244022,69,10614,185,0.425,1.15404,0.206694,0.178922,62,13360,288,1.0,0.50825,0.97639,-0.107277,32,15657,288,1.0,0.931543,0.102305,0.181484,57,15651,280,0.936,1.336003,0.417923,0.465686,76,14815,275,0.883,0.415901,1.265687,-0.171229,25,15598,276,0.925,0.987329,0.018397,0.389706,57,15201,272,0.923,0.479032,1.061807,0.221952,29,15999,271,0.904,lora_recipe__raw,scratch_n5__raw,lora_recipe__raw +Cycling activated T (CD38/HLA-DR),rare state,61162,679,362,679,True,17,0.114875,Abd-S66,0.64599,0.630416,-0.196561,33,8631,362,0.461,0.443262,1.17377,-0.245701,61,12396,679,1.0,0.403693,1.308668,-0.54902,61,13611,679,1.0,0.408872,1.290279,-0.27451,61,14147,645,0.934,0.578323,0.790053,0.014706,78,13435,614,0.882,0.314981,1.666662,-0.568627,45,13443,650,0.953,0.45704,1.129607,0.066176,65,13851,628,0.923,0.278936,1.841993,-0.189893,39,13815,619,0.927,real_full__raw,scratch_n5__raw,real_full__raw +Proliferating T (label),rare state,61162,444,444,444,True,17,0.175676,HC07,3.878322,1.955433,0.286765,243,8631,444,1.0,2.489231,1.3157,0.465686,224,12396,444,1.0,1.366289,0.450262,0.431373,135,13611,444,1.0,1.489793,0.575111,0.485294,153,14147,444,1.0,1.845583,0.884076,0.735294,180,13435,444,1.0,1.127185,0.172724,0.487745,110,13443,444,1.0,1.760317,0.815835,0.639706,177,13851,444,1.0,0.977178,0.033306,0.620098,98,13815,444,1.0,scratch_recipe__raw,scratch_recipe__raw,real_lora__raw +Plasmablast (label),rare state,131629,716,716,716,True,17,0.177374,Abd-S66,1.506762,0.591452,0.161506,209,25500,716,1.0,1.874441,0.90646,0.275865,260,25500,716,1.0,1.124664,0.169495,0.38619,156,25500,716,1.0,3.474925,1.796982,0.013539,482,25500,716,1.0,3.280271,1.713815,0.373006,455,25500,716,1.0,1.240015,0.310357,0.432004,172,25500,716,1.0,2.299795,1.201505,0.131531,319,25500,716,1.0,1.17513,0.232821,0.16718,163,25500,716,1.0,scratch_n5__raw,scratch_n5__raw,lora_n5__raw +Plasmablast (markers),rare state,131629,3407,1618,3407,True,17,0.125624,Abd-S66,0.781626,0.35545,0.249856,245,25500,1618,0.503,0.503011,0.991339,0.100737,332,25500,3407,1.0,0.349986,1.514629,0.42205,231,25500,3407,1.0,1.039898,0.056441,0.12285,609,25500,3023,0.898,1.154475,0.207237,0.077348,643,25500,2875,0.859,0.642013,0.639326,0.199877,373,25500,2999,0.89,0.710737,0.492612,0.321473,411,25500,2985,0.893,0.734132,0.445888,0.192994,415,25500,2918,0.862,real_lora__raw,scratch_recipe__raw,real_lora__raw +MS1 monocyte (IL1R2+),rare state,31471,7470,5568,7470,True,16,0.271218,Skin-S140,0.083836,3.576279,0.790422,141,9506,5568,0.416,0.151909,2.71872,0.83784,259,7183,7470,1.0,0.095614,3.386628,0.781863,138,6241,7278,0.893,0.205964,2.279533,0.811359,178,3891,6990,0.767,0.361097,1.469541,0.909871,325,4120,6875,0.73,0.404617,1.305372,0.91355,372,4036,7169,0.84,0.301019,1.732076,0.874696,261,3965,6882,0.729,0.4621,1.113723,0.896759,377,3598,7136,0.826,scratch_recipe__raw,scratch_recipe__raw,real_full__raw +Cycling monocyte (Ye 2026),rare state,31471,1684,992,1684,True,17,0.270784,Skin-S140,0.093446,3.419729,0.387466,28,9506,992,0.416,0.111874,3.160047,0.219214,43,7183,1684,1.0,0.061377,4.026153,0.177777,19,6241,1561,0.893,0.089557,3.481044,-0.357115,16,3891,1445,0.767,0.113928,3.133807,0.154134,21,4120,1408,0.73,0.086753,3.526941,0.190203,17,4036,1528,0.84,0.096243,3.377178,0.261804,17,3965,1402,0.729,0.086946,3.523731,-0.032839,15,3598,1509,0.826,real_full__raw,scratch_recipe__raw,real_full__raw +ISG-high monocyte (positive control),positive control,31471,4195,1907,4195,True,17,0.246007,IC23,2.626643,1.39322,0.20098,1513,9506,1907,0.416,0.324813,1.62232,-0.091968,311,7183,4195,1.0,0.324734,1.622669,-0.218137,246,6241,3820,0.893,0.435961,1.19773,0.69042,181,3891,3358,0.767,0.576362,0.794952,0.367647,241,4120,3194,0.73,0.306379,1.706609,0.645399,142,4036,3614,0.84,0.597133,0.743876,0.523343,243,3965,3230,0.729,0.263355,1.924919,0.312686,108,3598,3587,0.826,lora_recipe__raw,scratch_n5__raw,lora_recipe__raw +ISG-high T,rare state,61162,864,241,864,True,16,0.263889,Res-S139,2.70515,1.435709,-0.123307,92,8631,241,0.461,0.456853,1.130198,0.146826,80,12396,864,1.0,0.145625,2.77967,0.007371,28,13611,864,1.0,0.373078,1.42245,0.71401,68,14147,788,0.934,0.449465,1.15372,0.448837,70,13435,709,0.882,0.169555,2.560174,0.411942,30,13443,805,0.953,0.532931,0.907981,0.526058,91,13851,754,0.923,0.183512,2.446057,0.441439,32,13815,772,0.927,lora_recipe__raw,scratch_recipe__raw,lora_recipe__raw +DN2 / ABC B,rare state,21022,822,607,822,True,17,0.237226,Abd-PS23,1.267906,0.342448,-0.068627,162,4425,607,0.681,1.106915,0.146544,0.375,173,3997,822,1.0,1.63464,0.708973,0.32598,186,2910,822,1.0,0.350657,1.511867,0.375,66,4903,807,0.988,0.372326,1.425363,0.485294,76,5233,820,0.993,0.633688,0.658155,0.490196,115,5053,755,0.902,0.28267,1.822809,0.39951,57,5157,822,1.0,0.459289,1.122527,0.594596,81,5051,734,0.88,lora_n5__raw,scratch_full__raw,lora_n5__raw +HSPC,rare state,131629,140,117,140,True,17,0.15,HC01,0.88238,0.180528,0.427994,20,25500,117,0.503,0.958643,0.060935,0.102639,26,25500,140,1.0,0.479321,1.060935,0.134742,13,25500,140,1.0,0.573547,0.802017,0.237823,15,25500,135,0.898,0.782109,0.354558,0.062972,20,25500,132,0.859,1.338276,0.420376,-0.068036,35,25500,135,0.89,0.710699,0.492689,-0.161023,19,25500,138,0.893,0.955911,0.065051,0.382202,25,25500,135,0.862,lora_n5__raw,scratch_recipe__raw,lora_n5__raw +Cytotoxic CD4 T,rare state,29300,413,342,413,True,16,0.234867,HC07,0.467215,1.097841,0.262661,19,3484,342,0.699,1.793379,0.842681,0.742647,161,6369,413,1.0,1.515945,0.600217,-0.394608,159,7441,413,1.0,0.895231,0.159668,0.761673,81,6419,413,1.0,1.322166,0.402903,0.815727,111,5956,413,0.997,0.800203,0.321562,0.578431,76,6738,413,1.0,1.233716,0.30301,0.845211,109,6268,413,1.0,1.187554,0.247993,0.392157,113,6800,410,0.987,real_lora__raw,scratch_recipe__raw,real_lora__raw +Granulocyte-like (LDG),rare state,131629,4732,2589,4732,True,17,0.275359,Abd-S39,0.139565,2.840988,0.259428,70,25500,2589,0.503,0.054543,4.19647,0.488552,50,25500,4732,1.0,0.128721,2.957683,0.403464,118,25500,4732,1.0,0.038254,4.708239,0.250709,32,25500,4318,0.898,0.041325,4.596826,0.458021,33,25500,4122,0.859,0.043096,4.536308,0.158444,36,25500,4312,0.89,0.025286,5.305527,0.33603,21,25500,4287,0.893,0.058573,4.093615,0.464206,47,25500,4142,0.862,scratch_n5__raw,scratch_n5__raw,lora_n5__raw +Platelet signal in monocytes (negative control),negative control,31471,6486,3221,6486,True,17,0.214924,PIC04,0.359741,1.474969,0.247549,350,9506,3221,0.416,0.835598,0.259119,0.218137,1237,7183,6486,1.0,0.59832,0.741011,-0.375,708,6241,5967,0.893,0.547251,0.869727,0.004902,363,3891,5365,0.767,0.699329,0.515957,0.171674,466,4120,5090,0.73,0.392348,1.349795,0.54902,286,4036,5684,0.84,0.544771,0.876279,0.063725,351,3965,5114,0.729,0.445627,1.16609,0.129412,285,3598,5594,0.826,lora_n5__raw,scratch_n5__raw,lora_n5__raw diff --git a/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/workflow_result.json b/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/workflow_result.json new file mode 100644 index 0000000000000000000000000000000000000000..b9d916e37c91238f26fe0adaa82ff492f5296d64 --- /dev/null +++ b/paper-releases/pbisc-diffusion/updates/20261005T142225Z-rare-states-scratch-vs-pretrained/diffusion/dataset/server_1/model8/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/workflow_result.json @@ -0,0 +1,837 @@ +{ + "synth": { + "headline": "Given the same sepsis training pairs, from-scratch models reproduce rare-state frequencies about as well as pretrained LoRA or full FT on the pre-specified 13-state average. There is no significant difference at 5 or 58 pairs, but equivalence is not shown. Pretrained models are better on proliferating and interferon T/NK states, keep donor-specific composition with 5 pairs, and train 15-19x faster. Scratch is closer mainly where the pretrained models over-call fold-0 scANVI labels. With no sepsis pairs in training, every fine-tuned model, scratch or pretrained, fails the sepsis monocyte states (C1Q-high <=0.09x, MS1 <=0.27x, against about 0.5x and 0.7x for own-cohort LoRA). After the depth correction, scratch ties cross-cohort LoRA on average. It is clearly worse on cycling T/NK (0.05x vs 0.15-0.46x, confirmed) and trails full FT and the training-free calibrated pretrained model (xcal).", + "within_cohort": "Setting: fold 0, 17 held-out donors, the same 5 or 58 sepsis pairs for both initialisations.\n\nTraining time: scratch takes 61-86 min vs 3.2-5.4 min for pretrained, i.e. 15-19x longer.\n\nAggregate (pre-specified 13-state mean |log2 gen/obs|): no significant scratch-pretrained difference.\n- 5 pairs: scratch 1.52 vs LoRA 1.30, D +0.22 [-0.08, 0.41] (BC [0.01, 0.46]).\n- 58 pairs: scratch 1.41 vs LoRA 1.39 (D +0.02 [-0.35, 0.25]) and vs full FT 1.27 (+0.14 [-0.14, 0.41]).\n- 58 pairs with recipes: scratch 1.31 vs LoRA 1.39 (-0.09 [-0.38, 0.18]).\n- State-level tests: 3 of 52 favour pretrained, 7 favour scratch, 42 n.s.\n- This is non-significance, not equivalence: the intervals span about +-0.3 log2, roughly 1.2-fold.\n\nThe average hides a lineage split.\n- Pretrained is consistently closer on T/NK cycling and interferon states: cycling T/NK 0.93-1.34x vs 0.42-0.51x; ISG-high T 0.37-0.53x vs 0.15-0.18x. The 5-state T/NK block is significantly closer at 58 pairs (vs LoRA +0.48 [0.07, 0.88], significant in all 7 scoring variants; vs full FT +0.45; with recipes +0.48).\n- Scratch's ISG deficit matches its weaker ISG positive control, so it is a general interferon gap.\n- Scratch is closer where the pretrained models over-call scANVI labels: proliferating T 1.49-2.49x vs 0.98-1.37x; plasmablast 1.87-3.47x vs 1.12-1.24x. This is specific to fold 0 (pooled LoRA-58 0.78x and 0.76x) and is not better cycling biology.\n- At 58 pairs scratch is also closer on DN2 (0.46-0.63x vs 0.28-0.37x, robust to depth) and on MS1 vs LoRA (0.40x vs 0.21x, BC only; fold 0 is LoRA's worst MS1 fold, 0.71x pooled).\n- C1Q-high favours pretrained at 58 pairs (1.05x and 0.92x vs 0.47x), but only by BC and after composition standardisation, and fold 0 is LoRA's best C1Q fold (0.50x pooled).\n- Both fail LDG (0.03-0.13x) and cycling monocytes (0.06-0.11x) (confirmed).\n\nWith 5 pairs, scratch is donor-collapsed (composition r 0.17 vs 0.70), so its pooled ratios overstate it (C1Q 1.01x). The borderline 5-pair pretrained advantage is a depth/composition artifact of the deeper LoRA-5 arm: at common depth D is +0.10.\n\nDonor tracking: no detectable difference, and both are poor. Mean Spearman, pretrained vs scratch: 0.30 vs 0.19 (5 pairs), 0.29 vs 0.27 (LoRA-58), 0.43 vs 0.27 (full FT, BC-significant only), 0.40 vs 0.36 (recipes). A perfectly tracking generator would reach about 0.8.\n\nPretrained zero-shot is the least faithful arm (1.76). Both 58-pair scratch arms beat it, mainly through C1Q-high. So the cohort's own pairs recover C1Q-high under either initialisation.", + "cross_cohort": "Setting: all 82 sepsis donors held out; models trained on ImMicro, Imperial, Giroux and Liu pairs only.\n\nTraining time: scratch 127 min vs LoRA r16 20 min, r64 44 min, full FT 93 min; xcal needs no training.\n\nSepsis monocyte states fail in every fine-tuned model, scratch or pretrained:\n- C1Q-high: 0.03-0.09x.\n- MS1: 0.10-0.27x.\n- Own-cohort LoRA reaches 0.50-0.53x and 0.64-0.71x.\n- xcal (pretrained, cross-cohort calibration only, no sepsis pairs) recovers MS1 to 0.72x [0.55, 1.05] but not C1Q-high (0.07x).\n\nConfirmed: the scratch model almost never generates cycling T/NK (0.05x; 35 of 72,018 T/NK cells), against 0.15-0.46x for every pretrained cross-cohort arm. The gap is significant in all 4 comparisons and survives the depth correction. Cycling activated T and cycling monocytes show the same direction.\n\nConfirmed: under the screen's scoring, scratch is further from observed than every pretrained cross-cohort arm. Mean |log2| is 2.62 for scratch vs r16 2.32, r64 2.21, full FT 1.92 and xcal 1.53. Part of this is a depth artifact: scratch cells have 20,606-UMI libraries with only ~0.1% on the marker genes, and they were scored against unthinned observed cells. At matched depth:\n- The gap to LoRA vanishes: D -0.11 [-0.24, 0.11] vs r16 and -0.04 vs r64.\n- Deficits remain vs full FT (+0.25 [0.03, 0.54], marginal, carried by the cycling and MS1 states) and vs xcal (+0.68 [0.55, 0.89]).\n\nPer state:\n- Scratch is closer only where the pretrained arms over-generate: HSPC 1.31x vs 2.96x (r64 and full FT); cytotoxic CD4 T 0.94x vs 1.41-2.02x; plasmablast label vs full FT (2.21x), a fragile call.\n- Plasmablast markers favour LoRA and full FT robustly (0.78x, 0.69x and 1.47x vs 0.34x).\n- ISG-high T is a tie once depth is matched (scratch 1.00x at r16 depth).\n\nLonger scratch training makes the result worse. The final s162000 checkpoint is further from observed than s50000 (+0.29 [0.15, 0.43]):\n- C1Q, MS1, cycling monocytes, cycling activated T and LDG moved away; plasmablasts and proliferating T (label) improved.\n- At matched depth, s50000 beats r16 (-0.39) while s162000 only ties it.\n- Over the same steps the library inflated from 13,891 to 20,606 UMIs and benchmark coverage fell from 0.063 to 0.032. This is a co-trend, not a tested cause.\n\nDonor tracking does not differ from LoRA r16/r64 (0 of 13 states), but the test has low power.", + "contrasts": "Within the cohort the contrasts are anecdotal: there are only 2 adult and 1 paediatric healthy controls.\n- 25 eligible (state, contrast) pairs. Directions kept: LoRA-5 23 vs scratch-5 17; LoRA-58 20 and full FT-58 20 vs scratch-58 19; LoRA+recipes 23 vs scratch+recipes 16; zero-shot 21.\n- Only the recipe gap is significant: discordant pairs 7 vs 0, sign test p = 0.016, uncorrected and not independent. At 5 pairs it is 8 vs 2 (p = 0.11); at 58 pairs it is 2 vs 1 and 3 vs 2 (p = 1).\n- At 58 pairs, every scratch loss is a near-zero generated difference (below 0.5 pp). Measured against each arm's own generated level, though, 3 of scratch-58's losses and 2 of scratch+recipe's are real reversals (LoRA-58 1, full FT 2).\n- At 5 pairs, scratch has 3 strong reversals: DN2 adult (+4.74 vs observed -3.16 pp) and cytotoxic CD4 T adult (+1.70 vs -3.11) and all-sepsis (+1.00 vs -0.87). LoRA-5 has 2: C1Q-high adult (-0.72 vs +4.72) and proliferating T label adult.\n- Eight of the 25 pairs rest on a single paediatric control.\n\nAcross cohorts the control groups are larger (7, 4 and 21 donors).\n- 21 common pairs. Directions kept: scratch 14 (s50000 15), LoRA r16 11, r64 10, full FT 16, xcal 19, zero-shot 15, own-cohort LoRA-5 17, LoRA-58 19.\n- Scratch, r16 and r64 are not distinguishable from coin flipping (scratch 14 of 21, p = 0.095).\n- Every cross-cohort arm compresses the effects it keeps. The median kept pair reproduces 4% of the observed difference for scratch (7% at r16 depth), 13-15% for LoRA r16/r64 and full FT, 23% for xcal and 34% for own-cohort LoRA-58.\n- On pairs kept by both arms, at the comparator's depth, scratch is only about 1.3x more compressed than LoRA (n.s.). Its compression relative to full FT (0.07 vs 0.26), xcal (about 6x) and own-cohort LoRA (about 4.6x) is robust.\n- MS1, adult sepsis minus adult HC: scratch reproduces 6% of the observed +18.97 pp, r16 10%, full FT 25%, xcal 50%, own-cohort LoRA-58 62%.\n- C1Q-high: scratch reproduces 1-4% (paediatric reversed), r16 reverses all three contrasts, own-cohort LoRA-58 reproduces 30-34%.\n- Cytotoxic CD4 T adult and all-vs-all are reversed by scratch, r16, r64 and full FT, and kept by xcal and own-cohort LoRA.", + "table_within": [ + { + "state": "C1Q-high monocyte (reference)", + "role": "rare state", + "zs": 0.02, + "lora_5": 1.07, + "scratch_5": 1.01, + "lora_58": 1.05, + "full_58": 0.92, + "scratch_58": 0.47, + "lora_recipe_58": 0.63, + "scratch_recipe_58": 0.53, + "verdict": "pretrained", + "note": "Weak. 5 pairs: tie (scratch-5 is donor-collapsed; composition-standardised 0.61 vs 0.85). 58 pairs: pretrained closer (D +1.01 vs LoRA, +0.97 vs full FT), significant only by BC and after composition standardisation, robust to depth scheme. Fold 0 is LoRA-58's best C1Q fold (0.50x pooled over 5 folds)." + }, + { + "state": "MS1 monocyte (IL1R2+)", + "role": "rare state", + "zs": 0.08, + "lora_5": 0.15, + "scratch_5": 0.1, + "lora_58": 0.21, + "full_58": 0.36, + "scratch_58": 0.4, + "lora_recipe_58": 0.3, + "scratch_recipe_58": 0.46, + "verdict": "scratch", + "note": "Weak. About 0.1x for both at 5 pairs. 58 pairs: scratch closer than LoRA (D -0.97 [-1.44, 0.04], BC-significant in every depth scheme); tie vs full FT; n.s. with recipes. Fold 0 is LoRA-58's worst MS1 fold (0.71x pooled)." + }, + { + "state": "Cycling monocyte (Ye 2026)", + "role": "rare state", + "zs": 0.09, + "lora_5": 0.11, + "scratch_5": 0.06, + "lora_58": 0.09, + "full_58": 0.11, + "scratch_58": 0.09, + "lora_recipe_58": 0.1, + "scratch_recipe_58": 0.09, + "verdict": "both-fail", + "note": "Confirmed: 0.06-0.11x in every trained arm; no significant difference." + }, + { + "state": "Cycling T/NK", + "role": "rare state", + "zs": 2.52, + "lora_5": 1.15, + "scratch_5": 0.51, + "lora_58": 0.93, + "full_58": 1.34, + "scratch_58": 0.42, + "lora_recipe_58": 0.99, + "scratch_recipe_58": 0.48, + "verdict": "pretrained", + "note": "D > 0 in 4 of 4; significant vs LoRA-58 (+1.15 [0.20, 1.78]) and after T-subtype composition standardisation. Fold 0 has the lowest observed level of the 5 folds; LoRA-58 is 0.34-0.66x in the other folds." + }, + { + "state": "Cycling activated T (CD38/HLA-DR)", + "role": "rare state", + "zs": 0.65, + "lora_5": 0.44, + "scratch_5": 0.4, + "lora_58": 0.41, + "full_58": 0.58, + "scratch_58": 0.31, + "lora_recipe_58": 0.46, + "scratch_recipe_58": 0.28, + "verdict": "tie", + "note": "All arms under-generate. Pretrained is closer under the screen scoring (significant vs full FT). After T-subtype composition standardisation, scratch is significantly closer at 5 pairs and the full-FT advantage disappears." + }, + { + "state": "Proliferating T (label)", + "role": "rare state", + "zs": 3.88, + "lora_5": 2.49, + "scratch_5": 1.37, + "lora_58": 1.49, + "full_58": 1.85, + "scratch_58": 1.13, + "lora_recipe_58": 1.76, + "scratch_recipe_58": 0.98, + "verdict": "scratch", + "note": "Pretrained over-generates the scANVI label; scratch is significantly closer in 3 of 4. Specific to fold 0 (LoRA-58 0.49-1.09x in other folds, 0.78x pooled). Not better biology: fewer scratch label cells pass the cycling-marker rule (16-23% vs 23-28%; observed 36-37%)." + }, + { + "state": "ISG-high T", + "role": "rare state", + "zs": 2.71, + "lora_5": 0.46, + "scratch_5": 0.15, + "lora_58": 0.37, + "full_58": 0.45, + "scratch_58": 0.17, + "lora_recipe_58": 0.53, + "scratch_recipe_58": 0.18, + "verdict": "pretrained", + "note": "D > 0 in 4 of 4; significant with recipes (+1.52 [0.08, 2.49]); the ordering holds under depth and composition checks. Both under-generate. Scratch also under-generates the ISG positive control, so this is a general interferon gap." + }, + { + "state": "Cytotoxic CD4 T", + "role": "rare state", + "zs": 0.47, + "lora_5": 1.79, + "scratch_5": 1.52, + "lora_58": 0.9, + "full_58": 1.32, + "scratch_58": 0.8, + "lora_recipe_58": 1.23, + "scratch_recipe_58": 1.19, + "verdict": "tie", + "note": "0.80-1.79x in every trained arm; no fidelity difference. Scratch-5 loses donor tracking (Spearman -0.40 vs 0.74) and reverses 2 contrasts." + }, + { + "state": "Plasmablast (label)", + "role": "rare state", + "zs": 1.51, + "lora_5": 1.87, + "scratch_5": 1.12, + "lora_58": 3.47, + "full_58": 3.28, + "scratch_58": 1.24, + "lora_recipe_58": 2.3, + "scratch_recipe_58": 1.18, + "verdict": "scratch", + "note": "Pretrained over-generates plasma cells (1.25-1.89% vs 0.64% observed). Scratch is closer in 4 of 4, but significant only vs full FT, and that is lost under donor-mean weighting. Specific to fold 0 (LoRA-58 0.76x pooled)." + }, + { + "state": "Plasmablast (markers)", + "role": "rare state", + "zs": 0.78, + "lora_5": 0.5, + "scratch_5": 0.35, + "lora_58": 1.04, + "full_58": 1.15, + "scratch_58": 0.64, + "lora_recipe_58": 0.71, + "scratch_recipe_58": 0.73, + "verdict": "tie", + "note": "At 58 pairs pretrained is closer under the screen scoring, but once depth is matched it overshoots (1.49-1.90x) and scratch becomes closer. At 5 pairs pretrained is robustly closer." + }, + { + "state": "DN2 / ABC B", + "role": "rare state", + "zs": 1.27, + "lora_5": 1.11, + "scratch_5": 1.63, + "lora_58": 0.35, + "full_58": 0.37, + "scratch_58": 0.63, + "lora_recipe_58": 0.28, + "scratch_recipe_58": 0.46, + "verdict": "scratch", + "note": "58 pairs: scratch closer in every depth scheme; significant with recipes (-0.70 [-1.21, -0.08]). 5 pairs: LoRA closer (BC only), and scratch-5 strongly reverses the adult contrast." + }, + { + "state": "HSPC", + "role": "rare state", + "zs": 0.88, + "lora_5": 0.96, + "scratch_5": 0.48, + "lora_58": 0.57, + "full_58": 0.78, + "scratch_58": 1.34, + "lora_recipe_58": 0.71, + "scratch_recipe_58": 0.96, + "verdict": "tie", + "note": "Only 13-35 generated cells per arm; wide intervals; no significant difference." + }, + { + "state": "Granulocyte-like (LDG)", + "role": "rare state", + "zs": 0.14, + "lora_5": 0.05, + "scratch_5": 0.13, + "lora_58": 0.04, + "full_58": 0.04, + "scratch_58": 0.04, + "lora_recipe_58": 0.03, + "scratch_recipe_58": 0.06, + "verdict": "both-fail", + "note": "Confirmed: 0.03-0.13x in every trained arm. The significant scratch-closer calls (5 pairs, recipes) compare two failures." + }, + { + "state": "ISG-high monocyte (positive control)", + "role": "control", + "zs": 2.63, + "lora_5": 0.32, + "scratch_5": 0.32, + "lora_58": 0.44, + "full_58": 0.58, + "scratch_58": 0.31, + "lora_recipe_58": 0.6, + "scratch_recipe_58": 0.26, + "verdict": "tie", + "note": "Control; not counted. No significant fidelity difference; scratch loses its disease directions." + }, + { + "state": "Platelet signal in monocytes (negative control)", + "role": "control", + "zs": 0.36, + "lora_5": 0.84, + "scratch_5": 0.6, + "lora_58": 0.55, + "full_58": 0.7, + "scratch_58": 0.39, + "lora_recipe_58": 0.54, + "scratch_recipe_58": 0.45, + "verdict": "pretrained", + "note": "Control; not counted. Technical signal. Pretrained closer in 4 of 4; significant vs full FT (+0.83 [0.45, 1.19])." + } + ], + "table_cross": [ + { + "state": "C1Q-high monocyte (reference)", + "role": "rare state", + "own_lora_58": 0.5, + "own_lora_5": 0.53, + "zs": 0.03, + "xcal": 0.07, + "lora_r16": 0.03, + "lora_r64": 0.06, + "full_108k": 0.09, + "scratch_162k": 0.07, + "scratch_50k": 0.09, + "scratch_162k_at_r16_depth": 0.08, + "verdict": "both-fail", + "note": "Every cross-cohort arm is at 0.03-0.09x. Scratch is significantly closer than r16, but both fail. Among the tested arms, only own-cohort LoRA (0.50-0.53x) recovers it." + }, + { + "state": "MS1 monocyte (IL1R2+)", + "role": "rare state", + "own_lora_58": 0.71, + "own_lora_5": 0.64, + "zs": 0.08, + "xcal": 0.72, + "lora_r16": 0.12, + "lora_r64": 0.11, + "full_108k": 0.27, + "scratch_162k": 0.1, + "scratch_50k": 0.16, + "scratch_162k_at_r16_depth": 0.16, + "verdict": "both-fail", + "note": "Scratch equals LoRA r16/r64 (0.14x vs 0.13x after marker-depth matching). Full FT is modestly but robustly higher. xcal (no training, no sepsis pairs) reaches 0.72x [0.55, 1.05], as high as own-cohort LoRA." + }, + { + "state": "Cycling monocyte (Ye 2026)", + "role": "rare state", + "own_lora_58": 0.23, + "own_lora_5": 0.3, + "zs": 0.1, + "xcal": 0.25, + "lora_r16": 0.06, + "lora_r64": 0.07, + "full_108k": 0.11, + "scratch_162k": 0.02, + "scratch_50k": 0.06, + "scratch_162k_at_r16_depth": 0.04, + "verdict": "both-fail", + "note": "0.02-0.30x in every arm, own-cohort included. Scratch is lowest (21 of 20,808 monocytes). At matched depth, pretrained stays closer only vs full FT and xcal." + }, + { + "state": "Cycling T/NK", + "role": "rare state", + "own_lora_58": 0.54, + "own_lora_5": 0.29, + "zs": 0.83, + "xcal": 0.46, + "lora_r16": 0.15, + "lora_r64": 0.17, + "full_108k": 0.36, + "scratch_162k": 0.05, + "scratch_50k": 0.06, + "scratch_162k_at_r16_depth": 0.06, + "verdict": "pretrained", + "note": "Confirmed: scratch makes 35 cycling cells among 72,018 T/NK cells. Pretrained is significantly closer in all 4 comparisons, at matched and detection-matched depth, and in every leave-one-donor-out run." + }, + { + "state": "Cycling activated T (CD38/HLA-DR)", + "role": "rare state", + "own_lora_58": 0.3, + "own_lora_5": 0.16, + "zs": 0.36, + "xcal": 0.25, + "lora_r16": 0.1, + "lora_r64": 0.14, + "full_108k": 0.23, + "scratch_162k": 0.05, + "scratch_50k": 0.08, + "scratch_162k_at_r16_depth": 0.07, + "verdict": "both-fail", + "note": "Every fine-tuned arm is at 0.23x or less. Pretrained is less bad, and significantly closer at matched depth vs r64, full FT and xcal (same direction as the confirmed cycling result)." + }, + { + "state": "Proliferating T (label)", + "role": "rare state", + "own_lora_58": 0.78, + "own_lora_5": 0.62, + "zs": 1.96, + "xcal": 0.64, + "lora_r16": 0.79, + "lora_r64": 1.04, + "full_108k": 1.13, + "scratch_162k": 0.53, + "scratch_50k": 0.44, + "scratch_162k_at_r16_depth": 0.53, + "verdict": "pretrained", + "note": "Weak: significant only vs r16 (D +0.57 [0.09, 0.80]). This is a label-frequency result only: about 5% of label cells carry 2+ cycling genes in scratch or r16, vs 37% observed." + }, + { + "state": "ISG-high T", + "role": "rare state", + "own_lora_58": 0.57, + "own_lora_5": 0.39, + "zs": 1.74, + "xcal": 1.59, + "lora_r16": 1.66, + "lora_r64": 1.93, + "full_108k": 2.05, + "scratch_162k": 0.6, + "scratch_50k": 0.7, + "scratch_162k_at_r16_depth": 1, + "verdict": "tie", + "note": "Pretrained arms over-generate and scratch under-generates at screen depth (1.00x at r16 depth, 0.67x detection-matched). All differences are n.s." + }, + { + "state": "Cytotoxic CD4 T", + "role": "rare state", + "own_lora_58": 1.35, + "own_lora_5": 1.06, + "zs": 0.43, + "xcal": 2.02, + "lora_r16": 1.41, + "lora_r64": 1.7, + "full_108k": 1.61, + "scratch_162k": 0.94, + "scratch_50k": 0.86, + "scratch_162k_at_r16_depth": 1.07, + "verdict": "scratch", + "note": "Scratch is closer than xcal (significant) and than r64 and full FT at matched depth. But scratch, r16, r64 and full FT all reverse the adult and all-vs-all contrasts, and scratch tracks donors worse than full FT." + }, + { + "state": "Plasmablast (label)", + "role": "rare state", + "own_lora_58": 0.76, + "own_lora_5": 0.41, + "zs": 0.39, + "xcal": 0.78, + "lora_r16": 1.38, + "lora_r64": 1.01, + "full_108k": 2.21, + "scratch_162k": 0.87, + "scratch_50k": 0.49, + "scratch_162k_at_r16_depth": 0.87, + "verdict": "scratch", + "note": "Weak: scratch is about as close as r64. It is significantly closer only than the over-generating full FT, and that call holds in 61 of 82 leave-one-donor-out runs." + }, + { + "state": "Plasmablast (markers)", + "role": "rare state", + "own_lora_58": 0.33, + "own_lora_5": 0.19, + "zs": 0.28, + "xcal": 0.37, + "lora_r16": 0.78, + "lora_r64": 0.69, + "full_108k": 1.47, + "scratch_162k": 0.34, + "scratch_50k": 0.26, + "scratch_162k_at_r16_depth": 0.53, + "verdict": "pretrained", + "note": "LoRA and full FT are closer, robust to marker-depth matching (scratch 0.47x vs 0.58-1.21x). xcal is no better than scratch, and own-cohort LoRA-58 is also low." + }, + { + "state": "DN2 / ABC B", + "role": "rare state", + "own_lora_58": 0.31, + "own_lora_5": 0.42, + "zs": 1.49, + "xcal": 0.3, + "lora_r16": 0.17, + "lora_r64": 0.18, + "full_108k": 0.28, + "scratch_162k": 0.16, + "scratch_50k": 0.16, + "scratch_162k_at_r16_depth": 0.24, + "verdict": "both-fail", + "note": "All cross-cohort arms are at 0.16-0.30x (own-cohort 0.31-0.42x). The scratch-pretrained ordering flips with the depth reference." + }, + { + "state": "HSPC", + "role": "rare state", + "own_lora_58": 0.4, + "own_lora_5": 0.71, + "zs": 0.55, + "xcal": 1.68, + "lora_r16": 1.62, + "lora_r64": 2.96, + "full_108k": 2.97, + "scratch_162k": 1.31, + "scratch_50k": 0.96, + "scratch_162k_at_r16_depth": 1.47, + "verdict": "scratch", + "note": "Scratch is significantly closer than r64 and full FT (both about 2.96x) at both depth references; n.s. vs r16 and xcal. Depends on the checkpoint (s100000 was 2.58x)." + }, + { + "state": "Granulocyte-like (LDG)", + "role": "rare state", + "own_lora_58": 0.07, + "own_lora_5": 0.04, + "zs": 0.24, + "xcal": 0.06, + "lora_r16": 0.01, + "lora_r64": 0.02, + "full_108k": 0.02, + "scratch_162k": 0.01, + "scratch_50k": 0.04, + "scratch_162k_at_r16_depth": 0.02, + "verdict": "both-fail", + "note": "Fine-tuned cross-cohort arms are at 0.01-0.02x (scratch: 35 of 123,000 cells); xcal 0.06x, own-cohort 0.04-0.07x." + }, + { + "state": "ISG-high monocyte (positive control)", + "role": "control", + "own_lora_58": 0.61, + "own_lora_5": 0.55, + "zs": 1.8, + "xcal": 1.42, + "lora_r16": 1.31, + "lora_r64": 1.45, + "full_108k": 1.77, + "scratch_162k": 0.47, + "scratch_50k": 0.53, + "scratch_162k_at_r16_depth": 0.95, + "verdict": "tie", + "note": "Control; not counted. n.s. Scratch is 0.95x at r16 depth and closer to 1 than pretrained after marker-depth matching." + }, + { + "state": "Platelet signal in monocytes (negative control)", + "role": "control", + "own_lora_58": 0.56, + "own_lora_5": 0.57, + "zs": 0.28, + "xcal": 0.19, + "lora_r16": 0.18, + "lora_r64": 0.18, + "full_108k": 0.36, + "scratch_162k": 0.08, + "scratch_50k": 0.13, + "scratch_162k_at_r16_depth": 0.17, + "verdict": "pretrained", + "note": "Control; not counted. Technical signal. Pretrained is closer, robust to depth matching (scratch 0.17x vs 0.23-0.35x)." + } + ], + "confirmed_claims": [ + "within-5: Both initialisations fail the hardest states within the cohort. LDG is at 0.03-0.13x and cycling monocytes at 0.06-0.11x in every trained arm. The significant 'scratch closer' LDG calls compare two failures (5 pairs D -1.23 [-2.07, -0.33]; recipes -1.19 [-2.18, -0.36]), and no cycling-monocyte comparison is significant.", + "cross-2: Across cohorts, pretraining clearly helps for cycling states. The final scratch model generates cycling T/NK at 0.05x (35 of 72,018 T/NK cells) vs 0.15-0.46x for every pretrained cross-cohort arm. Pretrained is significantly closer in all 4 comparisons (D +1.62 to +3.22), at matched depth (+1.23 [0.61, 1.91] vs r16), with detection-matched depth and in all 82 leave-one-donor-out runs. Cycling activated T (0.05x vs 0.10-0.25x) and cycling monocytes (0.02x vs 0.06-0.25x) go the same way; at matched depth they stay significant vs full FT and xcal (and vs r64 for activated T). The label-based proliferating T state should not be counted as cycling evidence.", + "cross-3: Over the 13 rare states, the final scratch model is further from observed than every pretrained cross-cohort arm under the screen's scoring (mean |log2| 2.62 vs 2.32 r16, 2.21 r64, 1.92 full FT, 1.53 xcal; D +0.30 to +1.08). The gap to LoRA r16/r64 is entirely depth-driven and vanishes at matched depth (-0.11 [-0.24, 0.11]; -0.04 [-0.21, 0.13]). Deficits remain vs full FT (+0.25 [0.03, 0.54], marginal and carried by the cycling and MS1 states) and vs xcal (+0.68 [0.55, 0.89]). The result is specific to the final checkpoint: the 50k scratch checkpoint beats r16 at matched depth." + ], + "weakened_or_refuted": [ + "within-1 (weakened): 'Scratch is about as faithful as pretrained' holds only as non-significance on the pre-specified 13-state mean at 58 pairs and with recipes; equivalence is not shown (CI +-0.3 log2). The mean hides a pretrained advantage on the T/NK block, which is significant at 58 pairs in every scoring variant vs LoRA. Over the 11 marker states, full FT is significantly closer (+0.35 [0.02, 0.62]). The 5-pair BC-borderline result is a depth/composition artifact. Fold 0 is LoRA-5's best fold.", + "within-2 (weakened): The pretrained advantage holds robustly in direction for cycling T/NK and ISG-high T. 'In every matched comparison' fails for cycling activated T: after T-subtype composition standardisation, scratch is significantly closer at 5 pairs. The three significant calls each depend on the scoring scheme, and the near-1x cycling T/NK level is specific to fold 0.", + "within-3 (weakened): Scratch's label-state wins rest on pretrained over-labelling, which is specific to fold 0 (pooled LoRA-58: proliferating T 0.78x, plasmablast 0.76x). The within-label cycling-marker deficit of scratch is significant only at 58 pairs. A seventh scratch-closer call (DN2 with recipes) was omitted from the claim.", + "within-4 (weakened): 'Neither initialisation wins consistently on the sepsis monocyte states' survives. The per-state magnitudes are fold- and model-specific: C1Q is LoRA-58's best fold and MS1 its worst. 'Both reach observed C1Q at 5 pairs' is confounded by CD16-monocyte over-generation in scratch-5 (composition-standardised 0.61x vs 0.85x).", + "within-6 (weakened): Pretrained keeps more disease directions significantly only with recipes (7 vs 0, p = 0.016 uncorrected). There is no difference at 58 pairs. Relative to each arm's own generated level, some 'weak' scratch losses are real reversals. All within-cohort contrasts rest on 1-2 control donors.", + "within-7 (weakened): 'Similar donor tracking' becomes 'no detectable difference, both poor, with a small consistent tilt towards pretrained'. vs full FT the tilt is BC-significant (+0.15, BC [0.003, 0.33]).", + "within-8 (weakened): Zero-shot is the least faithful arm, and own-pair training recovers C1Q-high robustly. However: the 5-pair scratch advantage over zero-shot is n.s.; MS1 recovery is partial; 'zero-shot closer for cycling activated T' is a depth/composition artifact; and the zero-shot thin constants are 0.42-0.70, not 0.42-0.50.", + "cross-1 (weakened; causal sentence refuted for MS1): Every cross-cohort fine-tuned arm fails C1Q-high and MS1, and only own-cohort LoRA recovers C1Q-high. But 'target-cohort pairs on the pretrained model, not pretraining alone' is false for MS1: xcal reaches 0.72x with no sepsis pairs and no training. Within the cohort, scratch with 58 own pairs also reaches C1Q 0.47x and MS1 0.40-0.46x, so the pretrained model is not shown to be necessary. The MS1 range also omits full FT 20k (0.29x).", + "cross-4 (weakened): The depth mismatch is confirmed. Its correction factor depends on the reference: 1.1-2.0x at r16 library depth, up to 3.2x with detection matching. 'Corrected ISG states about 1.0x' holds only at the r16 reference (ISG-high T 0.67x with detection matching). Library thinning is also imperfect for the pretrained arms.", + "cross-5 (weakened; exclusivity refuted): Scratch is robustly closer for HSPC and cytotoxic CD4 T where the pretrained arms over-generate. Plasmablast label vs full FT is fragile (61 of 82 LODO). 'These are the only significant scratch wins apart from the failed states' is false: plasmablast markers vs xcal at matched depth is another. 'Uniformly failed' overstates DN2.", + "cross-6 (weakened): No scratch-LoRA difference in donor tracking, but this is a low-power null (median CI width 0.45; many donors with zero generated state cells). 'Own-cohort LoRA tracks best' holds only for LoRA-58 on average; xcal tracks MS1 better.", + "cross-7 (weakened; '3-4x more compressed than LoRA' refuted): Direction parity holds but is uninformative (scratch, r16 and r64 are not above coin flipping). On shared pairs at the comparator's depth, scratch is only about 1.3x more compressed than LoRA (n.s.). Compression relative to full FT, xcal and own-cohort LoRA is robust.", + "cross-8 (weakened; 'only plasmablasts improve' refuted): Longer scratch training makes the model worse overall (s162000 vs s50000, +0.29 [0.15, 0.43]). Proliferating T (label) also improves significantly, the C1Q step is fragile (57 of 82 LODO), and the link to library inflation and coverage loss is a 4-point co-trend, not a tested cause." + ], + "caveats": [ + "Cross-cohort depth confound (high severity). Scratch cells have 20,606-UMI median libraries (18-20x the pretrained arms, 10x observed) but only 236 detected genes and ~0.1% marker share. The screen scored them against unthinned observed cells (thin 1.0), while the pretrained arms were scored against cells thinned to 0.34-0.62. This biases scratch marker-state ratios down by 1.1-2.0x (up to 3.2x with detection matching). Primary cross-cohort columns are the screen's numbers; the r16-depth column and the matched-depth aggregates correct for this. Library thinning is also imperfect for the pretrained arms, so every absolute cross-cohort ratio carries a shared unknown bias.", + "Library-ratio common thinning (V1) is invalid for cross-cohort scratch: it removes ~95% of its marker counts and manufactures complete misses that should not be quoted. Marker-sum matching (V2) is the preferred depth match; median-of-ratios (V3) is only a stress test.", + "Cross-cohort scratch count profiles are unrealistic in ways the 73-marker screen cannot see. The library inflates with training (13,891 to 20,606 UMIs) and its spread is compressed (q90/q10 2.0 vs observed 7.3). This plausibly underlies its 0.03 coverage. The full generated matrices were not examined.", + "Scratch with 5 pairs is donor-collapsed (composition r 0.17 vs 0.70; marker between-donor CV 0.19-0.33x observed). Its pooled ratios, e.g. C1Q-high 1.01x, do not mean donor-level fidelity.", + "Within the cohort there is one fold, one trained model per arm and one generation seed, and the donor bootstrap ignores training and sampling variance. Fold 0 favours LoRA-5 overall and LoRA-58 on C1Q-high, and disfavours LoRA-58 on MS1; it also has the lowest observed cycling-T, proliferating-T-label and plasmablast levels. Scratch exists only for fold 0, so these fold effects cannot be checked for scratch.", + "Label states depend on scANVI calls on generated cells. Pretrained arms over-generate plasma cells and proliferating-labelled T cells, which also inflates their cycling activated T counts. Generated label cells under-express the cycling programme in every arm.", + "Controls: within the cohort, scratch under-generates the ISG positive control and loses its directions, so its ISG-high T deficit is a general interferon gap. Pretrained arms reproduce the platelet negative control more closely, so 'closer to observed' also rewards technical signal.", + "Zero-shot ISG overshoot is mostly a depth effect: ISG-high T goes from 2.71x to 0.76x within the cohort and from 1.74x to 0.59x across cohorts after marker-depth matching.", + "Statistics: 17 donors within the cohort; 52 within-cohort and 52 cross-cohort state-level tests with no multiplicity correction; percentile and bias-corrected intervals disagree on borderline states. Within-cohort contrasts rest on 1-2 healthy control donors per age group and are anecdotal.", + "Checkpoints: the cross-cohort scratch checkpoint was selected by validation loss, which ignores rare-state fidelity, and the 50k checkpoint is closer overall. The within-cohort 58-pair scratch run hit a 100k-step cap.", + "Verdict labels are judgments over the matched comparisons. 'both-fail' means scratch and every pretrained fine-tuned comparator are at <=0.30x; the per-row notes say which side is less bad. Ratios are pooled over donors; per-donor zero-generated-state cells are common in the cross-cohort arms.", + "REPORT.md was NOT written: the subagent harness rejected the report-file write. The full report content is returned in the report_markdown field for the coordinator to save if wanted. No other file was created or modified.", + "Archive: this subagent published no pbisc-diffusion delta. The synthesis is new derived evidence for the coordinator's single task-end archive update." + ], + "report_file": "NOT WRITTEN - the harness blocks subagent report files. Full content is in report_markdown. Intended path: /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/REPORT.md", + "report_markdown": "# Rare-state generation: from scratch vs pretrained (LoRA / full FT)\n\nPBISC-Diffusion, Model 8. Ye 2026 paired sepsis cohort (paediatric + adult PBMC bulk/sc). 2026-10-05.\n\nSources:\n- Screen outputs in `out/`: `sepsis_fold0_*.csv` and `sepsis_loco_*.csv` (method `tools/sepsis_states.py`, rules `tools/states.json`).\n- Derived analyses (`within_*`, `cross_*`), the confound audit (`confound_*`) and two verifier passes (`verify_recompute_*`, `verify_confound_*`), all in `out/scratch_vs_lora_20261005/`.\n\nThe synthesis ran on CPU only and changed no existing file.\n\n## 1. Question\n\nGiven the same training pairs, how well does a model trained from scratch generate rare cell states? \"From scratch\" means random initialisation with no corpus pretraining. The comparison is the pretrained model fine-tuned with LoRA or with full fine-tuning.\n\n## 2. Data, metric and arms\n\n**Metric.**\n- For each arm and state, ratio = generated state % of parent / observed state % of parent.\n- Observed cells are binomially thinned to the arm's generated parent median library, with one constant per arm and marker state. The two scANVI-label states are not thinned.\n- A ratio of 1.00 means the observed frequency. Fidelity is |log2 ratio|.\n- For a scratch-pretrained pair, D = fidelity(scratch) - fidelity(pretrained). D > 0 means pretrained is closer.\n- Brackets are 95% donor-bootstrap intervals (2,000 resamples, seed 20261005). \"BC\" is the bias-corrected sensitivity interval.\n- 15 states: 13 rare states, one positive control (ISG-high monocyte) and one negative control (platelet signal in monocytes).\n- All 15 are informative in both settings; the smallest observed-thinned count is 117 within the cohort and 629 across cohorts.\n- No arm generated zero cells of any state, so there are no complete misses.\n\n**Within the cohort.** Fold 0, with 17 held-out donors: 9 adult and 3 paediatric sepsis, 2 adult and 1 paediatric healthy control, and 1 adult and 1 paediatric ICU control. Each arm generated 1,500 cells per donor from real bulk, K = 1,000.\n\n| Arm | Initialisation | Pairs | Training time |\n|---|---|---|---|\n| zs__raw | pretrained, zero-shot | 0 | none |\n| lora_n5__raw | pretrained + LoRA | 5 | 3.2 min |\n| scratch_n5__raw | scratch | 5 | 61 min |\n| real_lora__raw | pretrained + LoRA | 58 | 5.4 min |\n| real_full__raw | pretrained + full FT | 58 | 4.7 min |\n| scratch_full__raw | scratch | 58 | 79 min (100k-step cap) |\n| lora_recipe__raw | pretrained + LoRA + recipes | 58 | 4.5 min |\n| scratch_recipe__raw | scratch + recipes | 58 | 86 min |\n\nScratch needs 15-19x the training time of the matched pretrained arm.\n\n**Across cohorts.** All 82 usable sepsis donors were held out: 46 adult and 15 paediatric sepsis, 7 adult and 4 paediatric healthy controls, and 6 adult and 4 paediatric ICU controls. Models were trained on ImMicro, Imperial, Giroux and Liu pairs only. Each arm generated 1,500 cells per donor.\n\n| Arm | Initialisation / method | Training time |\n|---|---|---|\n| zs__raw | pretrained, zero-shot | none |\n| xcalL_K1000 | pretrained zero-shot + cross-cohort calibration | none |\n| locoL_lora16sat_s26750_K1000 | pretrained + LoRA r16, step-saturated | 20 min |\n| locoL_lora64sat_s59000_K1000 | pretrained + LoRA r64 | 44 min |\n| locoL_fullsat_s108500_K1000 (and s20000) | pretrained + full FT, saturated at 108.5k steps | 93 min |\n| locoL_scratchsat_s162000_K1000 (milestones s50000/s100000/s150000) | scratch; s162000 is the selected final state (step-152,000 weights, chosen by inner-validation loss) | 127 min |\n| real_lora__raw, lora_n5__raw | own-cohort references: pretrained LoRA on the sepsis cohort's own pairs, cross-fitted 5-fold (~58 or 5 pairs per fold) | fold 0: 5.4 / 3.2 min |\n\nScratch needs 1.4-6.4x the training time of the pretrained cross-cohort arms.\n\n**Benchmark context.** These numbers come from the paired cross-cohort benchmark and were not re-derived here.\n- The cross-cohort scratch model has precision 1.00 but coverage 0.03-0.06. LoRA r16 has coverage 0.34 and own-cohort LoRA 0.83.\n- Within the cohort, scratch matched pretrained on coverage and precision. With 5 pairs, however, it lost donor-specific composition: across-donor composition r was 0.17 vs 0.70.\n\n**Verdict rule** (both result tables).\n- **pretrained** or **scratch**: that initialisation is closer in most matched comparisons. At least one difference is significant (percentile or BC interval), none is significant the other way by the percentile interval, and the ordering survives the depth and composition checks.\n- **both-fail**: scratch and every pretrained fine-tuned comparator are at 0.30x or less, so any ordering compares two failures.\n- **tie**: everything else.\n\nThe matched comparisons are:\n- Within the cohort: 5 pairs, 58 pairs vs LoRA, 58 pairs vs full FT, and 58 pairs with recipes.\n- Across cohorts: final scratch vs LoRA r16, LoRA r64, full FT 108.5k and xcal.\n\n## 3. Within-cohort result (same pairs, fold 0)\n\nValues are generated/observed ratios; 1.00 is the observed frequency.\n\n| State | zero-shot | LoRA-5 | scratch-5 | LoRA-58 | full FT-58 | scratch-58 | LoRA+recipe-58 | scratch+recipe-58 | Verdict |\n|---|---|---|---|---|---|---|---|---|---|\n| C1Q-high monocyte (reference) | 0.02 | 1.07 | 1.01 | 1.05 | 0.92 | 0.47 | 0.63 | 0.53 | pretrained |\n| MS1 monocyte (IL1R2+) | 0.08 | 0.15 | 0.10 | 0.21 | 0.36 | 0.40 | 0.30 | 0.46 | scratch |\n| Cycling monocyte (Ye 2026) | 0.09 | 0.11 | 0.06 | 0.09 | 0.11 | 0.09 | 0.10 | 0.09 | both-fail |\n| Cycling T/NK | 2.52 | 1.15 | 0.51 | 0.93 | 1.34 | 0.42 | 0.99 | 0.48 | pretrained |\n| Cycling activated T (CD38/HLA-DR) | 0.65 | 0.44 | 0.40 | 0.41 | 0.58 | 0.31 | 0.46 | 0.28 | tie |\n| Proliferating T (label) | 3.88 | 2.49 | 1.37 | 1.49 | 1.85 | 1.13 | 1.76 | 0.98 | scratch |\n| ISG-high T | 2.71 | 0.46 | 0.15 | 0.37 | 0.45 | 0.17 | 0.53 | 0.18 | pretrained |\n| Cytotoxic CD4 T | 0.47 | 1.79 | 1.52 | 0.90 | 1.32 | 0.80 | 1.23 | 1.19 | tie |\n| Plasmablast (label) | 1.51 | 1.87 | 1.12 | 3.47 | 3.28 | 1.24 | 2.30 | 1.18 | scratch |\n| Plasmablast (markers) | 0.78 | 0.50 | 0.35 | 1.04 | 1.15 | 0.64 | 0.71 | 0.73 | tie |\n| DN2 / ABC B | 1.27 | 1.11 | 1.63 | 0.35 | 0.37 | 0.63 | 0.28 | 0.46 | scratch |\n| HSPC | 0.88 | 0.96 | 0.48 | 0.57 | 0.78 | 1.34 | 0.71 | 0.96 | tie |\n| Granulocyte-like (LDG) | 0.14 | 0.05 | 0.13 | 0.04 | 0.04 | 0.04 | 0.03 | 0.06 | both-fail |\n| (control) ISG-high monocyte, positive | 2.63 | 0.32 | 0.32 | 0.44 | 0.58 | 0.31 | 0.60 | 0.26 | tie |\n| (control) Platelet signal in monocytes, negative | 0.36 | 0.84 | 0.60 | 0.55 | 0.70 | 0.39 | 0.54 | 0.45 | pretrained |\n| Mean abs log2, 13 rare states | 1.76 | 1.30 | 1.52 | 1.39 | 1.27 | 1.41 | 1.39 | 1.31 | |\n\n### Per-state notes\n\n**C1Q-high (pretrained, weak).**\n- At 5 pairs both reach the observed level (1.01x vs 1.07x). Scratch-5, however, is donor-collapsed and over-generates CD16 monocytes; composition-standardised, the ratios are 0.61x vs 0.85x.\n- At 58 pairs scratch falls to 0.47x, against LoRA 1.05x (D +1.01 [-0.26, 1.34]) and full FT 0.92x (D +0.97 [-0.10, 1.25]).\n- These differences are significant only by BC and after composition standardisation (+0.82 [0.24, 1.20]; +0.83 [0.35, 1.33]). They are robust to the depth scheme.\n- Fold 0 is this LoRA's best C1Q fold: 1.04x vs 0.27-0.76x in folds 1-4, and 0.50x pooled. A fold-0 LoRA trained on 40 pairs gives only 0.44x.\n\n**MS1 (scratch, weak).**\n- Both are near 0.1x at 5 pairs.\n- At 58 pairs scratch is 0.40x vs LoRA 0.21x (D -0.97 [-1.44, 0.04]); this is BC-significant in every depth scheme.\n- Scratch ties full FT (0.36x). With recipes it is 0.46x vs 0.30x, n.s.\n- Fold 0 is this LoRA's worst MS1 fold: 0.72-1.21x elsewhere, 0.71x pooled.\n\n**Cycling monocyte (both-fail).** 0.06-0.11x in every trained arm. No difference is significant.\n\n**Cycling T/NK (pretrained).**\n- Pretrained 0.93-1.34x vs scratch 0.42-0.51x. D > 0 in 4 of 4 comparisons.\n- Significant vs 58-pair LoRA (+1.15 [0.20, 1.78]), and still significant after T-subtype composition standardisation (+1.00 [0.15, 1.59]).\n- Fold 0 has the lowest observed level of the five folds. The same LoRA gives 0.34-0.66x in the other folds.\n\n**Cycling activated T (tie).**\n- Every arm under-generates (0.28-0.58x).\n- Under the screen's scoring, pretrained is closer and significantly so vs full FT (+0.87 [0.24, 1.40]).\n- However, 82-94% of the pretrained arms' cycling-rule T cells sit inside their over-generated proliferating-T label (observed 72-74%).\n- After composition standardisation, scratch is significantly closer at 5 pairs (-0.59 [-1.03, -0.17]) and the full-FT advantage is gone (+0.36 [-0.14, 0.75]).\n\n**Proliferating T, label (scratch, specific to fold 0).**\n- The pretrained models over-generate this label: 1.49-2.49x vs 0.98-1.37x for scratch. Scratch is significantly closer in 3 of 4 comparisons.\n- This is not better biology. Fewer scratch label cells pass the cycling-marker rule (16-23% vs 23-28%; observed 36-37%). That gap is significant only at 58 pairs.\n- In folds 1-4 the same 58-pair LoRA under-generates the label (0.49-1.09x; 0.78x pooled).\n\n**ISG-high T (pretrained).**\n- Pretrained 0.37-0.53x vs scratch 0.15-0.18x. D > 0 in 4 of 4; significant with recipes (+1.52 [0.08, 2.49]).\n- The ordering holds in every depth scheme and after composition standardisation.\n- Both under-generate. Scratch also under-generates the ISG positive control and loses its disease directions, so this is a general interferon gap in scratch.\n\n**Cytotoxic CD4 T (tie).**\n- 0.80-1.79x in every trained arm, with no fidelity difference.\n- Scratch-5 loses donor tracking (Spearman -0.40 vs 0.74) and reverses two contrasts.\n\n**Plasmablast, label (scratch, specific to fold 0).**\n- Pretrained over-generates (1.87-3.47x vs 1.12-1.24x) because it makes too many plasma cells: 1.25-1.89% of cells vs 0.64% observed, against 0.61-0.67% for scratch.\n- The difference is significant only vs full FT (-1.40 [-2.02, -0.13]), and that significance is lost under donor-mean weighting.\n- Pooled over all 82 donors, the 58-pair LoRA gives 0.76x.\n\n**Plasmablast, markers (tie).**\n- Under the screen's scoring, pretrained is closer at 58 pairs (1.04-1.15x vs 0.64x).\n- Once depth is matched, pretrained overshoots (1.49-1.90x) and scratch becomes closer.\n- At 5 pairs pretrained is robustly closer.\n\n**DN2 / ABC B (scratch at 58 pairs).**\n- At 58 pairs scratch is 0.46-0.63x vs pretrained 0.28-0.37x. This is robust in every depth scheme and significant with recipes (-0.70 [-1.21, -0.08]).\n- At 5 pairs LoRA is closer (BC only), and scratch-5 strongly reverses the adult contrast.\n\n**HSPC (tie).** Each arm generated only 13-35 cells, so the intervals are wide.\n\n**LDG (both-fail).** 0.03-0.13x in every trained arm. The two significant \"scratch closer\" calls compare two failures.\n\n**Controls.**\n- ISG-high monocyte: no significant difference.\n- Platelet signal: pretrained is closer in 4 of 4, significantly vs full FT (+0.83 [0.45, 1.19]). This is a technical signal, so being closer is not a biological win.\n\n### Aggregate\n\nThe pre-specified 13-state mean shows no significant difference:\n- 5 pairs: D +0.22 [-0.08, 0.41] (BC [0.01, 0.46]).\n- 58 pairs: +0.02 [-0.35, 0.25] vs LoRA and +0.14 [-0.14, 0.41] vs full FT.\n- Recipes: -0.09 [-0.38, 0.18].\n- State-level tests: 3 of 52 favour pretrained, 7 favour scratch and 42 are n.s.; with BC intervals it is 9 vs 11.\n\nThis is weaker than \"equally faithful\", for four reasons:\n- **Non-significance, not equivalence.** The intervals span about +-0.3 log2.\n- **The mean hides a T/NK advantage for pretrained.** Over the five T/NK states, pretrained is significantly closer at 58 pairs:\n - vs LoRA +0.48 [0.07, 0.88], significant in all 7 scoring variants;\n - vs full FT +0.45 [0.01, 1.02];\n - with recipes +0.48 [0.01, 0.91].\n Over the 11 marker states, full FT is significantly closer (+0.35 [0.02, 0.62]) in 4 of 7 variants.\n- **The 5-pair BC result is a depth/composition artifact.** At common depth D is +0.10 [-0.23, 0.35]; after composition standardisation it is -0.00.\n- **Fold 0 is LoRA-5's best fold.** Its mean abs log2 is 1.26-1.30 here vs 1.53-2.94 in folds 1-4.\n\nZero-shot (1.76) is the least faithful arm. The two 58-pair scratch arms beat it (D -0.35 [-0.66, -0.13] and -0.46 [-0.73, -0.24]), and the gain comes mostly from C1Q-high. Scratch-5 does not beat it significantly (D -0.24 [-0.52, 0.01]).\n\n**Donor tracking.** Mean Spearman, pretrained vs scratch:\n\n| Comparison | Pretrained | Scratch | Difference |\n|---|---|---|---|\n| 5 pairs | 0.30 | 0.19 | n.s. |\n| LoRA-58 | 0.29 | 0.27 | n.s. |\n| Full FT-58 | 0.43 | 0.27 | +0.15, BC [0.003, 0.33] |\n| Recipes | 0.40 | 0.36 | n.s. |\n\nNo difference is detectable, both track poorly, and there is a small tilt towards pretrained. A perfect generator would reach 0.78-0.85.\n\n## 4. Cross-cohort result (no sepsis pairs in training)\n\nValues are generated/observed ratios. The column \"scratch 162k at r16 depth\" scores the same scratch cells against observed cells thinned for LoRA r16.\n\n| State | own LoRA-58 | own LoRA-5 | zero-shot | xcal | LoRA r16 | LoRA r64 | full FT 108.5k | scratch 162k | scratch 162k at r16 depth | scratch 50k | Verdict |\n|---|---|---|---|---|---|---|---|---|---|---|---|\n| C1Q-high monocyte (reference) | 0.50 | 0.53 | 0.03 | 0.07 | 0.03 | 0.06 | 0.09 | 0.07 | 0.08 | 0.09 | both-fail |\n| MS1 monocyte (IL1R2+) | 0.71 | 0.64 | 0.08 | 0.72 | 0.12 | 0.11 | 0.27 | 0.10 | 0.16 | 0.16 | both-fail |\n| Cycling monocyte (Ye 2026) | 0.23 | 0.30 | 0.10 | 0.25 | 0.06 | 0.07 | 0.11 | 0.02 | 0.04 | 0.06 | both-fail |\n| Cycling T/NK | 0.54 | 0.29 | 0.83 | 0.46 | 0.15 | 0.17 | 0.36 | 0.05 | 0.06 | 0.06 | pretrained |\n| Cycling activated T (CD38/HLA-DR) | 0.30 | 0.16 | 0.36 | 0.25 | 0.10 | 0.14 | 0.23 | 0.05 | 0.07 | 0.08 | both-fail |\n| Proliferating T (label) | 0.78 | 0.62 | 1.96 | 0.64 | 0.79 | 1.04 | 1.13 | 0.53 | 0.53 | 0.44 | pretrained |\n| ISG-high T | 0.57 | 0.39 | 1.74 | 1.59 | 1.66 | 1.93 | 2.05 | 0.60 | 1.00 | 0.70 | tie |\n| Cytotoxic CD4 T | 1.35 | 1.06 | 0.43 | 2.02 | 1.41 | 1.70 | 1.61 | 0.94 | 1.07 | 0.86 | scratch |\n| Plasmablast (label) | 0.76 | 0.41 | 0.39 | 0.78 | 1.38 | 1.01 | 2.21 | 0.87 | 0.87 | 0.49 | scratch |\n| Plasmablast (markers) | 0.33 | 0.19 | 0.28 | 0.37 | 0.78 | 0.69 | 1.47 | 0.34 | 0.53 | 0.26 | pretrained |\n| DN2 / ABC B | 0.31 | 0.42 | 1.49 | 0.30 | 0.17 | 0.18 | 0.28 | 0.16 | 0.24 | 0.16 | both-fail |\n| HSPC | 0.40 | 0.71 | 0.55 | 1.68 | 1.62 | 2.96 | 2.97 | 1.31 | 1.47 | 0.96 | scratch |\n| Granulocyte-like (LDG) | 0.07 | 0.04 | 0.24 | 0.06 | 0.01 | 0.02 | 0.02 | 0.01 | 0.02 | 0.04 | both-fail |\n| (control) ISG-high monocyte, positive | 0.61 | 0.55 | 1.80 | 1.42 | 1.31 | 1.45 | 1.77 | 0.47 | 0.95 | 0.53 | tie |\n| (control) Platelet signal in monocytes, negative | 0.56 | 0.57 | 0.28 | 0.19 | 0.18 | 0.18 | 0.36 | 0.08 | 0.17 | 0.13 | pretrained |\n| Mean abs log2, 13 rare states | 1.29 | 1.54 | 1.82 | 1.53 | 2.32 | 2.21 | 1.92 | 2.62 | 2.21 | 2.33 | |\n\n### Per-state notes\n\n**C1Q-high (both-fail).** Every cross-cohort arm is at 0.03-0.09x, while own-cohort LoRA reaches 0.50-0.53x. Scratch (0.07x) is significantly closer than r16 (0.03x), but both fail.\n\n**MS1 (both-fail).**\n- Scratch (0.10x), r16 (0.12x) and r64 (0.11x) do not differ; after marker-depth matching it is 0.14x vs 0.13x.\n- Full FT is modestly but robustly higher (0.27x).\n- xcal reaches 0.72x [0.55, 1.05] with no sepsis pairs and no training. That is as high as own-cohort LoRA.\n\n**Cycling monocyte (both-fail).**\n- 0.02-0.30x in every arm, own-cohort LoRA included.\n- Scratch is lowest: 21 of 20,808 monocytes.\n- At matched depth, pretrained stays closer only vs full FT and xcal.\n\n**Cycling T/NK (pretrained; confirmed).**\n- Scratch makes 35 cycling cells among 72,018 T/NK cells (0.05x), vs 0.15-0.46x in every pretrained arm.\n- Pretrained is significantly closer in all 4 comparisons. This survives matched depth (vs r16, D +1.23 [0.61, 1.91]), detection-matched depth and every leave-one-donor-out run.\n\n**Cycling activated T (both-fail; pretrained less bad).** Scratch is at 0.05x vs 0.10-0.25x for the pretrained arms. At matched depth, the pretrained arms are significantly closer vs r64, full FT and xcal.\n\n**Proliferating T, label (pretrained, weak).**\n- Scratch 0.53x vs r16 0.79x (D +0.57 [0.09, 0.80]). Differences vs r64, full FT and xcal are n.s.\n- About 5% of label cells carry 2+ cycling genes in scratch or r16, vs 37% of observed cells. This is a label-frequency result, not cycling evidence.\n\n**ISG-high T (tie).**\n- Pretrained over-generates (1.35-2.05x) and scratch is at 0.60x.\n- At r16 depth scratch is 1.00x; detection-matched it is 0.67x.\n- All differences are n.s.\n\n**Cytotoxic CD4 T (scratch).**\n- Scratch (0.94x) is significantly closer than xcal (2.02x), and at matched depth closer than r64 (1.70x) and full FT (1.61x).\n- But scratch, r16, r64 and full FT all reverse the adult and all-vs-all contrasts.\n- Scratch also tracks donors worse than full FT (Spearman 0.40 vs 0.62).\n\n**Plasmablast, label (scratch, weak).**\n- Scratch (0.87x) is as close as r64 (1.01x).\n- It is significantly closer only than the over-generating full FT (2.21x), and that call holds in only 61 of 82 leave-one-donor-out runs.\n\n**Plasmablast, markers (pretrained).**\n- Scratch 0.34x vs r16 0.78x, r64 0.69x and full FT 1.47x. This survives marker-depth matching (0.47x vs 0.58-1.21x).\n- xcal (0.37x) is no better than scratch, and own-cohort LoRA-58 is also low (0.33x).\n\n**DN2 / ABC B (both-fail).** All cross-cohort arms are at 0.16-0.30x (own-cohort 0.31-0.42x). The ordering flips with the depth reference.\n\n**HSPC (scratch).**\n- Scratch (1.31x) is significantly closer than r64 and full FT (both about 2.96x) at both depth references.\n- The difference is n.s. vs r16 (1.62x) and xcal (1.68x).\n- It depends on the checkpoint: s100000 was at 2.58x.\n\n**LDG (both-fail).** Fine-tuned cross-cohort arms are at 0.01-0.02x; scratch generated 35 of 123,000 cells.\n\n**Controls.**\n- ISG-high monocyte: n.s. Scratch is 0.95x at r16 depth.\n- Platelet signal: pretrained is closer, robust to depth matching. This is a technical signal.\n\n### Aggregate (confirmed)\n\nUnder the screen's scoring, scratch is further from observed than every pretrained cross-cohort arm:\n\n| Comparison | D, screen scoring | D, matched depth |\n|---|---|---|\n| vs LoRA r16 | +0.30 [0.09, 0.52] | -0.11 [-0.24, 0.11] |\n| vs LoRA r64 | +0.41 [0.17, 0.62] | -0.04 [-0.21, 0.13] |\n| vs full FT | +0.69 [0.42, 1.00] | +0.25 [0.03, 0.54] |\n| vs xcal | +1.08 [0.92, 1.29] | +0.68 [0.55, 0.89] |\n\n- At matched depth the gap to LoRA vanishes.\n- The deficit vs full FT is marginal and carried by the cycling and MS1 states.\n- xcal has the lowest mean error of any cross-cohort arm (1.53, vs 1.75-2.62 for the others).\n\n### Scratch training length\n\nThe final checkpoint is further from observed than the 50k checkpoint: D +0.29 [0.15, 0.43].\n- Longer training moved C1Q-high, MS1, cycling monocytes, cycling activated T and LDG further away.\n- It improved only the two plasmablast states and proliferating T (label).\n- At matched depth, s50000 beats r16 (D -0.39 [-0.54, -0.17]) while s162000 only ties it.\n- Over the same steps the library inflated from 13,891 to 20,606 UMIs and coverage fell from 0.063 to 0.032. This is a co-trend across four checkpoints, not a tested cause.\n\n### Donor tracking\n\n- No state differs between scratch and r16 or r64 (0 of 13 each), but this is a low-power null.\n- Full FT and xcal each track 2 states better than scratch.\n- Own-cohort LoRA-58 tracks best on average (mean rho 0.36), although xcal tracks MS1 better.\n\n## 5. Contrasts\n\n### Within the cohort (anecdotal: 2 adult and 1 paediatric healthy control)\n\nDirections kept, out of 25 eligible state-contrast pairs:\n\n| Setting | Pretrained | Scratch |\n|---|---|---|\n| 5 pairs | LoRA-5: 23 | 17 |\n| 58 pairs | LoRA-58: 20; full FT-58: 20 | 19 |\n| 58 pairs + recipes | 23 | 16 |\n\nZero-shot keeps 21.\n\n- Only the recipe gap is significant: 7 vs 0 discordant pairs, p = 0.016, uncorrected and not independent.\n- At 58 pairs the gap is 2 vs 1 and 3 vs 2 discordant pairs (p = 1).\n- At 5 pairs, scratch has 3 strong reversals: DN2 adult (+4.74 vs observed -3.16 pp) and cytotoxic CD4 T adult and all-sepsis. LoRA-5 has 2.\n- Eight of the 25 pairs rest on a single paediatric control.\n\n### Across cohorts (control groups of 7, 4 and 21)\n\nDirections kept, out of 21 common pairs:\n\n| Arm | Kept |\n|---|---|\n| Scratch | 14 |\n| LoRA r16 | 11 |\n| LoRA r64 | 10 |\n| Full FT | 16 |\n| xcal | 19 |\n| Own-cohort LoRA-58 | 19 |\n\nScratch, r16 and r64 are not distinguishable from coin flipping.\n\nKept effects are compressed. The share of the observed difference that the median kept pair reproduces is:\n\n| Arm | Share reproduced |\n|---|---|\n| Scratch | 4% (7% at r16 depth) |\n| LoRA and full FT | 13-15% |\n| xcal | 23% |\n| Own-cohort LoRA-58 | 34% |\n\n- On pairs kept by both arms, at matched depth, scratch is only about 1.3x more compressed than LoRA (n.s.).\n- Compression relative to full FT, xcal and own-cohort LoRA is robust.\n- MS1, adult sepsis vs healthy control: scratch reproduces 6%, r16 10%, full FT 25%, xcal 50% and own-cohort LoRA-58 62%.\n- C1Q-high: scratch reproduces 1-4%, r16 reverses all three contrasts, and own-cohort LoRA-58 reproduces 30-34%.\n\n## 6. Confounds and caveats\n\n1. **Cross-cohort depth confound (high).**\n - Scratch cells carry 20,606-UMI libraries but only 236 detected genes, and about 0.1% of their counts fall on marker genes. Their marker detection matches the ~1,000-UMI pretrained arms.\n - The screen compared scratch with unthinned observed cells, so its marker-state ratios are biased down by 1.1-2.0x (up to 3.2x with detection matching).\n - Library thinning is also imperfect for the pretrained arms. Every absolute cross-cohort ratio therefore carries a shared bias.\n2. **V1 thinning is invalid for cross-cohort scratch.** It removes ~95% of marker counts. Its \"complete misses\" are artifacts and should not be quoted.\n3. **Cross-cohort scratch count profiles are unrealistic.** Libraries inflate, cell-to-cell spread is compressed (q90/q10 2.0 vs 7.3 observed) and coverage is 0.03. The full generated matrices were not examined.\n4. **Scratch-5 is donor-collapsed** (composition r 0.17 vs 0.70). Its pooled ratios overstate donor-level fidelity.\n5. **One fold, one model and one seed within the cohort.**\n - Fold 0 favours LoRA-5, and LoRA-58 on C1Q-high, but disfavours LoRA-58 on MS1.\n - It also has the lowest observed cycling-T and plasmablast levels.\n - Scratch exists only for fold 0.\n6. **Label states depend on scANVI calls.** The pretrained arms over-generate plasma cells and proliferating-labelled T cells.\n7. **Controls.**\n - Scratch's ISG deficit reflects a general interferon gap.\n - \"Closer to observed\" also rewards technical signal, as the platelet control shows.\n8. **Zero-shot ISG overshoot is mostly depth.** ISG-high T goes from 2.71x to 0.76x after matching.\n9. **Statistics.**\n - 17 donors within the cohort.\n - 104 state-level tests with no multiplicity correction.\n - Within-cohort contrasts rest on 1-2 controls.\n10. **Checkpoints.** The scratch checkpoint was chosen by validation loss, and the 50k checkpoint is closer. The within-cohort 58-pair scratch run hit its 100k-step cap.\n\n### Statements dropped after verification\n\n- **\"Pairs on the pretrained model, not pretraining alone, recover the monocyte states.\"** False for MS1, since xcal reaches 0.72x. The pretrained model is also not shown to be necessary: within the cohort, scratch with 58 pairs reaches C1Q 0.47x and MS1 0.40-0.46x.\n- **\"Scratch compression is 3-4x worse than LoRA.\"** It is about 1.3x on shared pairs at matched depth, and n.s.\n- **\"Only plasmablasts improve with longer scratch training.\"** Proliferating T (label) improves too.\n- **\"The only significant scratch wins are HSPC, cytotoxic CD4 T and plasmablast label.\"** Plasmablast markers vs xcal at matched depth is another.\n- **\"Zero-shot is closer for cycling activated T.\"** This is a depth/composition artifact.\n- **\"Pretrained is better on T cycling and interferon states in every comparison.\"** This fails for cycling activated T after composition standardisation.\n- **\"Donor tracking is similar.\"** The data support only \"no detectable difference\".\n- **\"Corrected ISG states are about 1.0x.\"** True only at the r16 reference.\n\n## 7. Verdict\n\n**With the same target-cohort pairs, scratch matches pretrained on the pre-specified 13-state average, but not everywhere.**\n- The difference on the average is not significant at 5 or 58 pairs, and equivalence is not shown.\n- Pretrained is better on proliferating and interferon T/NK states: cycling T/NK and ISG-high T, with the T/NK block significantly closer at 58 pairs.\n- Scratch collapses donors with 5 pairs and needs 15-19x the training time.\n- Scratch is closer mainly where the pretrained models over-call fold-0 scANVI labels, and on DN2 and the fold-specific MS1 result at 58 pairs.\n- Both fail LDG and cycling monocytes.\n\n**Without target-cohort pairs, initialisation is not the main factor.**\n- Every cross-cohort fine-tuned model, scratch or pretrained, fails the sepsis monocyte states: C1Q-high at 0.09x or less and MS1 at 0.27x or less. Own-cohort LoRA reaches about 0.5x and 0.7x.\n- After the depth correction, scratch ties LoRA on the aggregate.\n- Scratch is clearly worse on cycling T/NK (0.05x vs 0.15-0.46x; confirmed). It also trails full FT and xcal; xcal has the lowest error of any cross-cohort arm and brings MS1 to 0.72x.\n- Scratch is closer only where the pretrained arms over-generate (HSPC, cytotoxic CD4 T).\n- The scratch generator drifts with longer training: libraries inflate, coverage reaches 0.03 and the monocyte states get worse.\n\n**Bottom line.**\n- Pretraining alone does not produce the sepsis monocyte states: zero-shot gives C1Q-high 0.03x and MS1 0.08x.\n- Within the cohort, its own pairs recover C1Q-high under either initialisation.\n- Across cohorts, only own-cohort pairs recover C1Q-high. MS1 is also recovered by calibrating the pretrained model.\n- For rare states, pretraining adds four things:\n - better proliferating T/NK states;\n - donor-specific composition with few pairs;\n - 15-19x shorter training;\n - when there are no target pairs, a training-free calibrated model with the lowest average error of any cross-cohort arm.\n\n**Open points.** No new runs were made. These remain untested:\n- scratch on folds 1-4 and with a second seed;\n- scratch checkpoint selection by rare-state fidelity;\n- scratch combined with the xcal calibration.\n\n**Archive.** No pbisc-diffusion delta was published by this subagent." + }, + "judged": [ + { + "id": "within-1", + "claim": "Within the cohort, a from-scratch model trained on the same pairs generates rare states about as faithfully as the pretrained model; there is no significant overall pretraining advantage. Mean |log2(gen/obs)| over the 13 informative rare states is 1.524 (scratch) vs 1.300 (LoRA) at 5 pairs. At 58 pairs it is 1.412 (scratch) vs 1.393 (LoRA) and 1.275 (full FT). With recipes it is 1.306 (scratch) vs 1.395 (LoRA). The paired differences D are +0.224 [-0.078, 0.406], +0.019 [-0.350, 0.255], +0.137 [-0.138, 0.406] and -0.089 [-0.379, 0.178]. Of the 52 state-level tests, 3 favour pretrained, 7 favour scratch and 42 are not significant. The only borderline case is 5 pairs, where the bias-corrected sensitivity CI is [0.010, 0.458].", + "evidence": "within_bootstrap.csv rows with state='MEAN over 13 informative rare states' and kind=main: D, D_ci_lo, D_ci_hi, abs_log2_scratch, abs_log2_pretrained, D_bc_ci_lo, D_bc_ci_hi. within_bootstrap_counts.csv (states='rare states'): pretrained_closer 0/1/1/1, scratch_closer 2/0/2/3. Under the BC sensitivity the totals are 9 pretrained vs 11 scratch of 52, still balanced.", + "how_to_recompute": "Use out/sepsis_fold0_per_donor.csv. For each arm and state, sum gen_state, gen_parent, obsthin_state and obs_parent over the 17 donors. Compute r = ((sum gen_state + 0.5)/sum gen_parent) / ((sum obsthin_state + 0.5)/sum obs_parent). Fidelity is |log2 r|. D = fidelity(scratch arm) - fidelity(pretrained arm), averaged over all states except 'ISG-high monocyte (positive control)' and 'Platelet signal in monocytes (negative control)'. For the CI: sort donors by name and set idx = np.random.default_rng(20261005).integers(0, 17, (2000, 17)). Recompute the four sums on donors[idx[b]] for both arms with the same idx, average D over the 13 states per replicate, and take the 2.5 and 97.5 percentiles. Script: within_analysis.py, run as CUDA_VISIBLE_DEVICES='' PYTHONDONTWRITEBYTECODE=1 /home/sj_server_1/miniforge3/envs/cxg_corpus/bin/python within_analysis.py.", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "weakened", + "reason": "Every stated number reproduces exactly from sepsis_fold0_per_donor.csv. I used pooled +0.5 ratios and the same 2000-rep donor idx (seed 20261005, donors sorted). That covers the arm means, the four D values with percentile CIs, the 5-pair BC CI [0.010, 0.458], the 3/7/42 split and the 9 vs 11 split under BC. The reading 'about as faithfully / no overall advantage' is weaker than stated, for three reasons. First, it rests on non-significance: CI half-widths are about 0.3 log2, so equivalence is not shown. Second, it holds only for the pre-specified 13-state mean. That mean nets a pretrained advantage on marker-defined states against scratch wins on the two scANVI-label states, and within-3 argues those label wins are not better biology. Third, dropping the 2 label states makes pretrained ahead: 5 pairs D=+0.410 [-0.001, 0.649] (BC [0.124, 0.736]), vs 58-pair full FT +0.354 [0.015, 0.622] (significant). vs 58-pair LoRA +0.194 [-0.232, 0.422] and recipes +0.055 [-0.300, 0.314] stay n.s. Also dropping LDG (the 'two failures' of within-5) gives 5 pairs +0.574 [0.128, 0.810] and full FT +0.395 [0.047, 0.674]. A symmetric exclusion of every double-failure state brings all D back to about 0. Donor-equal weighting keeps all four n.s. (full FT +0.278 [-0.022, 0.601]). Supported: no significant advantage on the 13-state mean at 58 pairs and with recipes. Not supported: the general 'about as faithfully' at 5 pairs and vs full FT.", + "recomputed": "Mean |log2| (pc 0.5): scratch_n5 1.524, lora_n5 1.300; scratch_full 1.412, real_lora 1.393, real_full 1.275; scratch_recipe 1.306, lora_recipe 1.395; zs 1.763 (pc 0: 1.534/1.304/1.422/1.405/1.283/1.318/1.406/1.775). D [percentile; BC]: 5 pairs +0.224 [-0.078, 0.406; 0.010, 0.458]; 58 vs LoRA +0.019 [-0.350, 0.255; -0.238, 0.379]; 58 vs full +0.137 [-0.138, 0.406; -0.129, 0.413]; recipes -0.089 [-0.379, 0.178; -0.396, 0.164]. State-level counts, pretrained-closer/scratch-closer, percentile: 0/2, 1/0, 1/2, 1/3 (total 3 vs 7, 42 n.s.). Same counts under BC: 2/3, 3/1, 2/3, 2/4 (total 9 vs 11). Integrity check: per-donor sums reproduce summary gen_over_obs and spearman_donors to 1e-15. Outputs are in .../out/scratch_vs_lora_20261005/: verify_recompute_within_{main,label_marker,sensitivity,donorweight}.py with .csv/.log. No existing file was modified. No archive update was made by this verifier; that is left to the coordinator." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "Every quoted number reproduces exactly. D and its CIs match within_bootstrap.csv to within 5e-7, and the screen's per-donor counts replay exactly from the per-cell extracts (120 of 120 arm x state rows). The 13-state aggregate is not significant under every confound check: leave-one-donor-out (17 of 17 runs, all four comparisons), donor-mean weighting, common-depth thinning of generated and observed cells (V1), marker-depth matching (V2), monocyte/T-subtype composition standardisation, and swapping in the fold-0 LoRA models trained on 5/10/20/40 pairs (aggregate D -0.14 to +0.19). The 5-pair 'borderline' result is itself a depth/composition artifact. LoRA-5 is the deepest arm and the screen never thins generated cells. At common depth its D falls to +0.096 [-0.226, 0.345]; after composition standardisation it is -0.004 [-0.231, 0.170]. The evidence is still thinner than 'about as faithfully' implies, for three reasons. (1) The average hides a robust lineage split. Over the 5 T/NK states, pretrained is closer at 58 pairs vs LoRA in all 7 scoring variants (D +0.42 to +0.64, every CI excludes 0). Using only the 11 marker states, i.e. dropping the two scANVI label states that within-3 itself attributes to pretrained over-labelling, full FT is significantly closer: +0.354 [0.015, 0.622]. That holds under reseeding, trained-arm common depth and composition standardisation. (2) The 3-vs-7 state tally is unstable: donor-mean weighting gives 7 pretrained vs 5 scratch, V2 gives 6 vs 7, composition standardisation 6 vs 10. (3) Only one fold and one trained model per arm exist, so the donor bootstrap omits training variance. LoRA-5's fold-0 mean |log2| (1.26-1.30) is its best of five folds (1.53-2.94 elsewhere). Fold 0 also has the lowest observed cycling-T, proliferating-T-label and plasmablast levels of all folds.", + "recomputed": "Mean |log2| over 13 rare states: zs 1.763, lora_n5 1.300, scratch_n5 1.524, real_lora 1.393, real_full 1.275, scratch_full 1.412, lora_recipe 1.395, scratch_recipe 1.306. D: +0.224 [-0.078, 0.406] (BC [0.010, 0.458]); +0.019 [-0.350, 0.255]; +0.137 [-0.138, 0.406]; -0.089 [-0.379, 0.178]. Tally 3/7/42 (BC 9/11/32). V1 common depth (trained arms): +0.096/+0.020/+0.160/-0.092, all n.s. Composition-standardised: -0.004/-0.024/+0.133/-0.206, all n.s. Donor-mean: +0.237/+0.083/+0.278 [-0.022, 0.601]/-0.043. 11 marker states: +0.410 [-0.001, 0.649]/+0.194/+0.354 [0.015, 0.622]/+0.055. T/NK 5 states vs 58-pair LoRA: +0.481 [0.073, 0.880], significant in every scheme. Files: out/scratch_vs_lora_20261005/verify_confound_within_{repro,stateset,stateset_by_scheme,lodo,donormean,depth,composition,folds}.csv" + } + ] + }, + { + "id": "within-2", + "claim": "Pretrained models reproduce T-cell cycling and interferon marker states better than scratch in every matched comparison. Across the pretrained trained arms vs the scratch arms: Cycling T/NK is 0.93-1.34x vs 0.42-0.51x; ISG-high T is 0.37-0.53x vs 0.15-0.18x; Cycling activated T is 0.41-0.58x vs 0.28-0.40x. D > 0 in 12 of 12 cases (3 states x 4 comparisons). Three are significant: Cycling T/NK with 58-pair LoRA (D=+1.147 [0.200, 1.783]), Cycling activated T vs full FT (+0.870 [0.239, 1.395]), and ISG-high T with recipes (+1.524 [0.081, 2.490]).", + "evidence": "sepsis_fold0_summary.csv gen_over_obs. Cycling T/NK: lora_n5 1.154, real_lora 0.932, real_full 1.336, lora_recipe 0.987 vs scratch_n5 0.508, scratch_full 0.416, scratch_recipe 0.479. ISG-high T: 0.457/0.373/0.449/0.533 vs 0.146/0.170/0.184. Cycling activated T: 0.443/0.409/0.578/0.457 vs 0.404/0.315/0.279. D and CIs are in within_bootstrap.csv for these 3 states and the 4 main comparisons.", + "how_to_recompute": "Use the per-donor CSV. The ratio without pseudocount is (sum gen_state/sum gen_parent) / (sum obsthin_state/sum obs_parent); it equals the summary gen_over_obs. For states 'Cycling T/NK', 'ISG-high T' and 'Cycling activated T (CD38/HLA-DR)', compute D and the donor-bootstrap CI as in within-1 for each matched pair (scratch_n5/lora_n5, scratch_full/real_lora, scratch_full/real_full, scratch_recipe/lora_recipe).", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "confirmed", + "reason": "All ratios, ranges, the 12/12 positive D values and the three significant CIs reproduce exactly. Exactly 3 of 12 percentile CIs exclude 0. Under BC it is 4 of 12, adding Cycling T/NK with recipes, BC [0.483, 1.831]. A direct cell-level check supports the direction more strongly than the fidelity D does. Marker-defined cycling T (% of T cells) is lower in scratch in all 4 matched comparisons (Fisher p 0.0012 / 0.0088 / <1e-4 / 0.0055). Donor-bootstrap CIs of the percentage-point difference all exclude 0. Donor-equal weighting keeps 12/12. Caveats: (1) The 12 cases are not independent: Cycling T/NK and Cycling activated T share the 6 cycling genes and the T parent, and scratch_full is used twice. (2) Cycling activated T at 5 pairs is a near-tie: 0.404x vs 0.443x, 61 vs 61 generated cells, D=+0.135 [-0.371, 0.568]. (3) Pretrained arms also under-generate ISG-high T (0.37-0.53x) and Cycling activated T (0.41-0.58x), so 'better' means 'less under-generated'.", + "recomputed": "gen/obs without pseudocount. Cycling T/NK: pretrained 1.154 / 0.932 / 1.336 / 0.987 vs scratch 0.508 / 0.416 / 0.479. ISG-high T: 0.457 / 0.373 / 0.449 / 0.533 vs 0.146 / 0.170 / 0.184. Cycling activated T: 0.443 / 0.409 / 0.578 / 0.457 vs 0.404 / 0.315 / 0.279. D with percentile CI, order Cycling T/NK, Cycling activated T, ISG-high T. 5 pairs: +0.741 [-0.434, 1.701], +0.135 [-0.371, 0.568], +1.633 [-0.160, 3.041]. 58 pairs vs LoRA: +1.147 [0.200, 1.783], +0.372 [-0.370, 0.958], +1.124 [-0.231, 2.229]. 58 pairs vs full FT: +0.815 [-0.142, 1.869], +0.870 [0.239, 1.395], +1.393 [-0.223, 2.664]. Recipes: +1.031 [-0.008, 1.526], +0.705 [-0.075, 1.497], +1.524 [0.081, 2.490]. Cycling T as % of T: scratch 0.206 / 0.156 / 0.166 vs pretrained 0.436 / 0.311 / 0.454 / 0.339." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "The ratios, the 12-of-12 D>0 pattern and the three significant CIs reproduce exactly. D>0 for all 12 cases also survives leave-one-donor-out and every depth scheme. ISG-high T is not a composition artifact: standardised ratios are 0.11-0.18 for scratch vs 0.37-0.50 for pretrained. Three confounds undercut 'better in every matched comparison'. (1) Parent composition. Pretrained arms over-generate cells labelled 'Proliferating T cell' (1.49-2.49x), and 82-94% of their cycling-rule T cells sit inside that label (observed 72-74%). Re-weighting T/NK subtypes to the observed mix gives the following. Pretrained Cycling T/NK falls to 0.67-1.00x (scratch 0.43-0.57x). Pretrained Cycling activated T falls to 0.21-0.38x (scratch 0.28-0.32x). For Cycling activated T, scratch becomes significantly closer at 5 pairs (D_std -0.587 [-1.029, -0.166]); with recipes D_std is about 0 (-0.036); the vs-full-FT significance disappears (+0.357 [-0.137, 0.754]). Cycling T/NK with recipes drops to +0.234 (n.s.). (2) Each of the three significant calls depends on the scoring scheme. Cycling T/NK vs 58-pair LoRA is n.s. under zero-shot-depth and marker-depth matching. Cycling activated T vs full FT is n.s. under zero-shot depth and after composition standardisation. ISG-high T with recipes is n.s. under both common-depth schemes; it keeps significance in 12 of 17 leave-one-donor-out runs and after composition standardisation. (3) The near-1x 'reproduction' is specific to fold 0. Fold 0 has the lowest observed Cycling T/NK of all folds (0.40% vs 1.04-1.30% in folds 1-4), and the same 58-pair LoRA gives 0.34-0.66x in those folds. Within label-called proliferating T cells, every arm under-expresses the cycling program (pass rate 0.44-0.78x of observed). The direction (pretrained higher on Cycling T/NK and ISG-high T) looks robust; the magnitude and the 'every comparison' wording do not.", + "recomputed": "Cycling T/NK pretrained 1.154/0.932/1.336/0.987 vs scratch 0.508/0.416/0.479. ISG-high T 0.457/0.373/0.449/0.533 vs 0.146/0.170/0.184. Cycling activated T 0.443/0.409/0.578/0.457 vs 0.404/0.315/0.279. D: Cycling T/NK vs 58-pair LoRA +1.147 [0.200, 1.783]; Cycling activated T vs full FT +0.870 [0.239, 1.395]; ISG-high T with recipes +1.524 [0.081, 2.490]; 12 of 12 D>0. Composition-standardised D: Cycling activated T -0.587*/+0.119/+0.357/-0.036; Cycling T/NK +0.762/+1.004*/+1.192*/+0.234; ISG-high T +1.942*/+1.156/+1.412/+1.460* (* = CI excludes 0)." + } + ] + }, + { + "id": "within-3", + "claim": "Scratch is significantly closer mainly where the pretrained model over-generates scANVI-labelled states, and this does not mean scratch reproduces the biology better. Proliferating T (label) is 1.49-2.49x in pretrained arms vs 0.98-1.37x in scratch arms; scratch is significantly closer in 3 of 4 comparisons (5 pairs D=-0.863 [-1.273, -0.459]; full FT -0.709 [-1.031, -0.097]; recipes -0.791 [-1.078, -0.085]). Plasmablast (label) is 1.87-3.48x vs 1.13-1.24x, significant only vs full FT (D=-1.401 [-2.018, -0.134]). Among label-called proliferating T cells, only 16.4-23.5% pass the 2-of-6 cycling-marker rule in scratch, vs 22.8-27.8% in pretrained arms and 35.6-37.4% in observed cells thinned to arm depth. Marker-defined cycling T cells are 0.16-0.21% of generated T cells in scratch vs 0.31-0.45% in pretrained (observed 0.35-0.38%).", + "evidence": "within_bootstrap.csv rows for 'Proliferating T (label)' and 'Plasmablast (label)'. within_label_marker_check.csv, column pct_of_pt_label_passing_cycling_rule: scratch_n5 18.52, scratch_full 16.36, scratch_recipe 23.47; lora_n5 22.77, real_lora 26.14, real_full 27.78, lora_recipe 24.29; observed thinned 35.59-37.39. Column cycling_rule_pct_of_T: scratch 0.21/0.16/0.17, pretrained 0.44/0.31/0.45/0.34, observed thinned 0.35-0.38.", + "how_to_recompute": "Ratios and D are computed as in within-1. The label/marker check needs the per-cell extracts, not the per-donor CSV. In work/sepsis/fold0/.npz, T cells are cells with aifi_l1 or aifi_l2 == 'T cell'; label-called cells have aifi_l2 == 'Proliferating T cell'; a cell is cycling if at least 2 of MKI67, TOP2A, TYMS, CENPF, UBE2C, BIRC5 have count > 0 (markers / marker_symbols). The observed side uses tools/sepsis_states.observed() for the same donors, binomially thinned with the arm's 'ISG-high T' thin (seed 20261005). Script: within_label_marker_check.py, run with PYTHONDONTWRITEBYTECODE=1.", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "weakened", + "reason": "Label-state ratios, D values and CIs reproduce exactly. The label/marker percentages also reproduce in my independent per-cell recompute, and per-donor T and PT-label counts and thin factors match the per-donor CSV and summary exactly. The PT-label pass-rate contrast is thin. Each arm has only 98-224 label cells (18-51 passing), and the ranges overlap: scratch_recipe 23.5% is above LoRA-5 22.8%. Matched differences are n.s. at 5 pairs (-4.2 pp, Fisher p=0.35, donor bootstrap [-14.7, +6.1]) and with recipes (-0.8 pp, p=1.0). Only the 58-pair scratch arm differs: vs LoRA -9.8 pp (p=0.07, bootstrap [-17.1, -1.6]); vs full FT -11.4 pp (p=0.03, bootstrap [-18.4, -3.7]). Every generated arm is far below observed (16-28% vs about 36-37%), so pretrained label calls are also marker-poor. 'Not better biology' is better supported by the marker-defined cycling-T deficit, which is significant in all 4 matched pairs. On 'mainly label states': of the 7 significant scratch-closer calls, 4 are label states and 2 are LDG. The seventh is unmentioned: DN2/ABC B with recipes (scratch 0.459x vs LoRA 0.283x, D=-0.696 [-1.206, -0.077]), a marker state that pretrained under-generates. Trivial rounding: PB label scratch_n5 is 1.1247 (stated 1.13) and real_lora is 3.4749 (stated 3.48).", + "recomputed": "Proliferating T (label) gen/obs: pretrained 2.489 / 1.490 / 1.846 / 1.760, scratch 1.366 / 1.127 / 0.977. D: 5 pairs -0.863 [-1.273, -0.459]; vs LoRA-58 -0.401 [-0.815, 0.204]; vs full FT -0.709 [-1.031, -0.097]; recipes -0.791 [-1.078, -0.085]. Plasmablast (label) gen/obs: pretrained 1.874 / 3.475 / 3.280 / 2.300, scratch 1.125 / 1.240 / 1.175. D: 5 pairs -0.735 [-1.113, 0.042]; vs LoRA -1.484 [-2.268, 0.344]; vs full FT -1.401 [-2.018, -0.134]; recipes -0.967 [-1.540, 0.358]. PT-label cells passing the 2-of-6 rule: scratch_n5 25/135 = 18.52%, scratch_full 18/110 = 16.36%, scratch_recipe 23/98 = 23.47%. Pretrained: lora_n5 51/224 = 22.77%, real_lora 40/153 = 26.14%, real_full 50/180 = 27.78%, lora_recipe 43/177 = 24.29%. Observed, thinned per arm (seed 20261005): 36.26-37.39%; 50-seed Monte Carlo range 35.1-37.4%. Cycling-rule % of T: scratch 0.206 / 0.156 / 0.166; pretrained 0.436 / 0.311 / 0.454 / 0.339; observed thinned 0.355-0.378." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "The ratios, D, CIs and label/marker percentages reproduce exactly. The proliferating-T-label calls are robust to leave-one-donor-out (17/17, 17/17, 14/17), and 'mainly' is 4 of the 7 scratch-closer calls (the others are LDG twice and DN2). Three problems remain. (1) The premise that the pretrained model over-generates scANVI-labelled states holds only in fold 0. Fold 0's donors have the lowest observed proliferating-T-label (0.73% vs 1.23-2.16% in folds 1-4) and plasmablast (0.54% vs 0.92-2.59%) levels of all folds. The same 58-pair LoRA gives 0.49-1.09x and 0.24-1.29x in folds 1-4, and 0.78x and 0.76x pooled over all 82 donors (README), so it under-generates there. Even within fold 0, LoRA models trained on different pair counts span 1.04-5.17x for plasmablasts. So the scratch-closer label calls may not generalise. (2) The within-label cycling-marker deficit of scratch is significant only at 58 pairs: pretrained minus scratch pass rate is +9.8 pp [1.6, 17.1] vs LoRA and +11.4 [3.7, 18.4] vs full FT. It is not significant at 5 pairs (+4.2 [-6.1, 14.7]) or with recipes (+0.8 [-8.0, 10.5]), and the quoted ranges overlap (scratch_recipe 23.5% vs lora_n5 22.8%). Every arm is far below observed (pass rate 0.44-0.78x). The marker-defined cycling-share gap (0.16-0.21% vs 0.31-0.45%) is significant in all four pairs, but it mostly reflects the label composition shown in within-2. (3) Under donor-mean weighting, proliferating-T-label significance is lost vs full FT and recipes, and plasmablast-label significance is lost vs full FT.", + "recomputed": "Proliferating T label: pretrained 2.489/1.490/1.846/1.760 vs scratch 1.366/1.127/0.977. D -0.863 [-1.273, -0.459]; -0.401 [-0.815, 0.204]; -0.709 [-1.031, -0.097]; -0.791 [-1.078, -0.085]. Plasmablast label: 1.874/3.475/3.280/2.300 vs 1.125/1.240/1.175; vs full FT -1.401 [-2.018, -0.134]. Pass rate within label (gen): scratch 18.5/16.4/23.5, pretrained 22.8/26.1/27.8/24.3, observed 35.8-37.4. Cycling share of T: scratch 0.206/0.156/0.166, pretrained 0.436/0.311/0.454/0.339. 58-pair LoRA proliferating-T label by fold: 1.49 (fold 0), 0.89, 0.49, 1.09, 0.51." + } + ] + }, + { + "id": "within-4", + "claim": "For the sepsis-induced monocyte states, neither initialisation wins consistently. C1Q-high: with 5 pairs both reach the observed frequency (scratch 1.005x, LoRA 1.075x; D=-0.096 [-0.403, 0.504]). With 58 pairs scratch drops to 0.467x while LoRA gives 1.053x and full FT 0.917x (D=+1.014 [-0.264, 1.343] and +0.970 [-0.099, 1.249]); this is not significant by the percentile CI but is significant in the bias-corrected sensitivity. MS1 (IL1R2+): both are about 0.1x at 5 pairs (0.096x vs 0.152x). With 58 pairs scratch is higher: 0.405x vs LoRA 0.206x (D=-0.972 [-1.442, 0.035]) and vs full FT 0.361x (D=-0.164 [-0.808, 0.857]); with recipes 0.462x vs 0.301x (D=-0.618 [-1.379, 0.399]). MS1 donor ranking is high in every trained arm (Spearman 0.78-0.91; scratch-pretrained differences at most 0.11).", + "evidence": "within_bootstrap.csv rows for 'C1Q-high monocyte (reference)' and 'MS1 monocyte (IL1R2+)': D, D_ci_*, D_bc_ci_*. The C1Q 58-pair BC intervals are [0.382, 1.467] for LoRA and [0.534, 1.484] for full FT; the MS1 58-pair LoRA BC interval is [-1.492, -0.098]. within_state_table.csv has the spearman__ columns.", + "how_to_recompute": "Compute r and D as in within-1 for the two states. Spearman is scipy.stats.spearmanr of per-donor gen_pct vs obsthin_pct in the per-donor CSV, using donors where both are defined. scratch_recipe has no generated monocytes for donor Abd-S47, so 16 donors are used there.", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "confirmed", + "reason": "Every number reproduces exactly, including the BC intervals and the Spearman range. 'Neither wins consistently' holds. C1Q favours pretrained at 58 pairs only under BC. MS1 favours scratch vs 58-pair LoRA only under BC. All other comparisons are n.s. Donor-equal weighting keeps the pattern: C1Q at 58 pairs is scratch 0.46x vs LoRA 0.75x and full FT 0.85x; MS1 at 58 pairs is scratch 0.36x vs LoRA 0.22x and full FT 0.37x. Note that MS1 donor ranking is just as high in zero-shot (rho 0.790), so neither initialisation gains it through training. C1Q with recipes (0.532x vs 0.626x, D=+0.231 [-0.644, 0.783]) is not mentioned but fits the claim.", + "recomputed": "C1Q-high gen/obs and D. 5 pairs: scratch 1.005 vs LoRA 1.075, D -0.096 [-0.403, 0.504]. 58 pairs: scratch_full 0.467 vs real_lora 1.053, D +1.014 [-0.264, 1.343], BC [0.382, 1.467]. vs real_full 0.917: D +0.970 [-0.099, 1.249], BC [0.534, 1.484]. MS1: 5 pairs 0.096 vs 0.152. 58 pairs: 0.405 vs real_lora 0.206, D -0.972 [-1.442, 0.035], BC [-1.492, -0.098]. vs real_full 0.361: D -0.164 [-0.808, 0.857]. Recipes: 0.462 vs 0.301, D -0.618 [-1.379, 0.399]. MS1 Spearman in trained arms: 0.782-0.914. |scratch - pretrained| = 0.056 / 0.103 / 0.004 / 0.022 (max 0.103). scratch_recipe uses 16 donors because Abd-S47 has 0 generated monocytes." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "All quoted numbers reproduce, including the BC intervals and MS1 Spearman 0.782-0.914 (largest scratch-pretrained gap 0.103; scratch_recipe uses 16 donors). Leave-one-donor-out overturns nothing. The individual sub-claims are confounded or specific to fold 0. (1) C1Q at 5 pairs, 'both reach the observed frequency'. scratch_n5 over-generates CD16 monocytes (26.9% vs 15.4% observed), which is where C1Q-high cells sit. Standardising monocyte subtypes gives scratch 0.61x vs LoRA 0.85x. LoRA-5 is the deepest, unthinned arm and drops to 0.81x at common depth (0.44x at zero-shot depth). (2) C1Q at 58 pairs. After composition standardisation the pretrained advantage is significant even by the percentile CI (+0.824 [0.243, 1.197] vs LoRA; +0.835 [0.347, 1.331] vs full FT), and it is BC-significant under every depth scheme. But fold 0 is this LoRA's best C1Q fold (1.04x vs 0.27-0.76x in folds 1-4; pooled 0.50x), and the fold-0 LoRA-40 model gives 0.44x (D vs scratch -0.08). (3) MS1 at 58 pairs. The scratch advantage is BC-significant in every depth scheme and significant under donor-mean weighting (-0.728 [-1.099, -0.163]), but it is measured against this LoRA's worst MS1 fold (0.20x vs 0.72-1.21x elsewhere; pooled 0.71x). (4) The high MS1 Spearman is partly the sepsis/control split. Within the 12 sepsis donors it ranges 0.55-0.90, and the scratch-pretrained gap reaches 0.28 at 5 pairs. 'Neither wins consistently' survives, but the per-state magnitudes depend on the fold and the specific LoRA model.", + "recomputed": "C1Q: 5 pairs 1.005/1.075, D -0.096 [-0.403, 0.504]; 58 pairs 0.467 vs 1.053/0.917, D +1.014 [-0.264, 1.343] (BC [0.382, 1.467]) and +0.970 [-0.099, 1.249] (BC [0.534, 1.484]). MS1: 0.096/0.152; 0.405 vs 0.206, D -0.972 [-1.442, 0.035] (BC [-1.492, -0.098]); vs full FT 0.361, D -0.164 [-0.808, 0.857]; recipes 0.462/0.301, D -0.618 [-1.379, 0.399]. Composition-standardised C1Q: scratch_n5 0.61, lora_n5 0.85, scratch_full 0.42, real_lora 0.74, real_full 0.74." + } + ] + }, + { + "id": "within-5", + "claim": "Both initialisations fail on the hardest states, so the significant 'scratch closer' calls there compare two failures. Granulocyte-like/LDG cells are 0.025-0.13x in every trained arm. At 5 pairs scratch is 0.129x vs LoRA 0.055x (D=-1.231 [-2.073, -0.327]); with recipes 0.059x vs 0.025x (D=-1.193 [-2.181, -0.356]). Cycling monocytes are 0.06-0.11x in every trained arm, with no significant difference in any comparison.", + "evidence": "sepsis_fold0_summary.csv gen_over_obs for 'Granulocyte-like (LDG)': 0.055/0.129/0.038/0.041/0.043/0.025/0.059. For 'Cycling monocyte (Ye 2026)': 0.112/0.061/0.090/0.114/0.087/0.096/0.087. within_bootstrap.csv verdicts.", + "how_to_recompute": "Use the pooled ratios from per-donor sums (within-1 formula, with or without +0.5). D and CIs come from the same donor bootstrap.", + "status": "confirmed", + "votes": [ + { + "lens": "recompute", + "verdict": "confirmed", + "reason": "The values reproduce exactly. Every trained arm under-generates LDG by about 8- to 40-fold and cycling monocytes by about 9- to 16-fold. So both significant LDG 'scratch closer' calls compare two failures. No cycling-monocyte comparison is significant by percentile or BC CI. Donor-equal weighting gives the same picture: LDG 0.019-0.098x, cycling monocytes 0.049-0.128x.", + "recomputed": "LDG gen/obs: lora_n5 0.0545, scratch_n5 0.1287, real_lora 0.0383, real_full 0.0413, scratch_full 0.0431, lora_recipe 0.0253, scratch_recipe 0.0586. D: 5 pairs -1.231 [-2.073, -0.327]; recipes -1.193 [-2.181, -0.356]; 58 pairs n.s. (vs LoRA -0.169 [-1.018, 0.507], vs full FT -0.059 [-0.630, 0.503]). Cycling monocytes: 0.112 / 0.061 / 0.090 / 0.114 / 0.087 / 0.096 / 0.087. D: 5 pairs +0.845 [-0.483, 2.226]; vs LoRA-58 +0.048 [-0.947, 1.070]; vs full FT +0.385 [-0.429, 1.285]; recipes +0.141 [-0.775, 0.976]. All BC intervals also include 0." + }, + { + "lens": "confound", + "verdict": "confirmed", + "reason": "The numbers reproduce: LDG is 0.025-0.129x in every trained arm, cycling monocytes 0.061-0.114x, the two D values and CIs match, and no cycling-monocyte comparison is significant. The 'two failures' reading is robust. Under every depth scheme and after composition standardisation, all trained arms stay at or below 0.13x for LDG and 0.17x for cycling monocytes, and cycling-monocyte differences stay n.s. throughout, including donor-mean weighting. The LDG scratch-closer calls hold in 17 of 17 leave-one-donor-out runs and after composition standardisation (-1.503 and -1.306). They lose significance only under marker-depth matching, which fits the claim's framing. No donor dominates: LDG cells come from all 17 donors, top-donor share 0.28. The pattern also generalises: the 58-pair LoRA's LDG is at most 0.16x in all five folds. The one high cycling-monocyte value elsewhere (LoRA-5 1.33x in fold 2) comes from a single donor (Res-S150, 131 of 186 generated cells).", + "recomputed": "LDG: 0.055/0.129/0.038/0.041/0.043/0.025/0.059. 5 pairs D -1.231 [-2.073, -0.327]; recipes -1.193 [-2.181, -0.356]. Cycling monocyte 0.112/0.061/0.090/0.114/0.087/0.096/0.087, all D n.s. (+0.845, +0.048, +0.385, +0.141). Under V1 common depth (trained arms), LDG for every arm is 0.02-0.12x." + } + ] + }, + { + "id": "within-6", + "claim": "Pretrained arms keep more disease directions, mostly because scratch generates no difference rather than the opposite one. The eligible set is 25 (state, contrast) pairs: informative rare states with |obs_diff| >= 0.5 pp in all trained arms. Directions kept: 5 pairs, LoRA 23 vs scratch 17; 58 pairs, LoRA 20 and full FT 20 vs scratch 19; recipes, LoRA 23 vs scratch 16. With 58 pairs every scratch loss is weak (|gen_diff| < 0.5 pp: 6 for scratch_full, 9 for scratch_recipe). With 5 pairs scratch has 3 strong reversals: DN2 adult (+4.74 vs observed -3.16 pp), Cytotoxic CD4 T adult (+1.70 vs -3.11) and all-sepsis (+1.00 vs -0.87). LoRA-5 has 2: C1Q-high adult (-0.72 vs +4.72) and Proliferating T label adult (+0.63 vs -0.51). The control groups are tiny: 2 adult HC, 1 paediatric HC and 2 ICU controls.", + "evidence": "within_contrast_counts.csv, set='rare, common eligible set': columns kept, reversed_strong, reversed_weak, flat and not_kept. Per-pair detail is in within_contrast_directions.csv (eligible_common, status, reversal_strength).", + "how_to_recompute": "Use out/sepsis_fold0_contrasts.csv. obs_diff and gen_diff are the mean per-donor obsthin_pct and gen_pct of group A minus group B; the same values can be rebuilt from the per-donor CSV 'group' column. Keep (state, contrast) pairs whose state is not a control and whose min over the 7 trained arms of |obs_diff| is >= 0.5 with one observed sign; this gives 25 pairs. For each arm, count pairs where sign(gen_diff) == sign(obs_diff). Call a loss strong if |gen_diff| >= 0.5.", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "weakened", + "reason": "All counts reproduce exactly from the per-donor CSV; group means of per-donor percentages match sepsis_fold0_contrasts.csv to 1e-14. The headline 'pretrained arms keep more disease directions' is thin. At 58 pairs the gap is 1 of 25: discordant pairs are 2 vs 1 (vs LoRA) and 3 vs 2 (vs full FT), exact sign p=1.0. At 5 pairs the discordant pairs are 8 vs 2 (p=0.11). Only recipes is clear: 7 vs 0 (p=0.016). These p-values are optimistic because the 25 pairs are not independent: all-sepsis overlaps the adult and paediatric contrasts, 8 of the 25 rest on the single paediatric HC donor and 9 on the 2 adult HC donors. 'Mostly no difference rather than the opposite' does hold. At 58 pairs every scratch loss is weak. At 5 pairs, 5 of 8 losses are weak or flat (4 weak reversals plus 1 exact zero) and 3 are strong reversals.", + "recomputed": "Eligible pairs: 25 (9 adult, 8 paediatric, 8 all-sepsis). Kept: zs 21, lora_n5 23, scratch_n5 17, real_lora 20, real_full 20, scratch_full 19, lora_recipe 23, scratch_recipe 16. Losses: scratch_n5 3 strong + 4 weak + 1 flat; scratch_full 6 weak; scratch_recipe 9 weak; lora_n5 2 strong; real_lora 5 weak; real_full 5 weak; lora_recipe 2 weak. Strong reversals: scratch_n5 DN2 adult gen +4.740 vs obs -3.164; Cytotoxic CD4 T adult +1.697 vs -3.111; Cytotoxic CD4 T all-sepsis +0.998 vs -0.875. lora_n5: C1Q-high adult -0.722 vs +4.716; PT label adult +0.631 vs -0.509. Donors: 9 adult sepsis, 3 paediatric sepsis, 2 adult HC, 1 paediatric HC, 1 adult ICU control, 1 paediatric ICU control." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "The counts recompute exactly: 25 eligible pairs; directions kept 23 vs 17 at 5 pairs, 20/20 vs 19 at 58 pairs, 23 vs 16 with recipes. The strong reversals are as stated, and every 58-pair scratch loss is below 0.5 pp. Four problems weaken the interpretation. (1) Most differences are not significant. Counting pairs where only pretrained vs only scratch keeps the direction gives 8 vs 2 at 5 pairs (exact sign test p=0.11), 2 vs 1 and 3 vs 2 at 58 pairs (p=1), and 7 vs 0 with recipes (p=0.016, uncorrected, pairs not independent). At 58 pairs there is effectively no difference. (2) 8 of the 25 pairs are the paediatric contrast, which has a single control donor (PHC03, 83-235 generated monocytes, 0-2 generated state cells), so one cell can decide a call. Dropping HC07 or Res-S139 flips an observed direction; the adult proliferating-T-label contrast (-0.51 pp) sits right at the eligibility threshold. (3) 'Weak' depends on an absolute 0.5 pp cut while generation is compressed. Measured against each arm's generated level, 3 of the 6 scratch_full losses are reversals of at least half that level (ISG-high T paediatric -0.31 pp = 1.0x; ISG-high T all -0.17 = 0.68x; LDG adult -0.13 = 0.77x). The same is true for 2 of the 9 scratch_recipe losses (cycling monocyte adult -0.43 = 0.92x; LDG adult -0.18 = 0.87x). (4) Some pretrained losses are reversals of the same kind.", + "recomputed": "Eligible pairs 25. Kept: zs 21, lora_n5 23, scratch_n5 17, real_lora 20, real_full 20, scratch_full 19, lora_recipe 23, scratch_recipe 16. Strong reversals (absolute cut): scratch_n5 3, lora_n5 2, all 58-pair arms 0. Strong reversals (relative cut): scratch_full 3, scratch_recipe 2, real_lora 1, real_full 2. Discordant counts and sign-test p: 8/2 p=0.109; 2/1 p=1; 3/2 p=1; 7/0 p=0.016. Leave-one-donor-out kept-difference ranges: 5 pairs +3..+7, 58 pairs vs LoRA -1..+2, vs full FT 0..+3, recipes +4..+8 (dropping PHC03 removes 8 pairs)." + } + ] + }, + { + "id": "within-7", + "claim": "Donor-level tracking of rare-state frequencies is similar for scratch and pretrained. Mean Spearman over the 13 rare states, pretrained vs scratch: 0.295 vs 0.187 at 5 pairs (difference +0.109 [-0.080, 0.294]); 0.293 vs 0.274 for 58-pair LoRA (+0.019 [-0.121, 0.186]); 0.428 vs 0.274 for full FT (+0.154 [-0.008, 0.313]); 0.398 vs 0.357 with recipes (+0.041 [-0.106, 0.217]). Only 2 of 52 state-level differences are significant, both favouring pretrained: Cytotoxic CD4 T at 5 pairs (0.743 vs -0.395; +1.137 [0.461, 1.603]) and Cycling T/NK vs full FT (0.466 vs -0.171; +0.637 [0.008, 1.111]).", + "evidence": "within_spearman.csv, kind=main: rows 'MEAN over 13 informative rare states' and state rows with verdict 'pretrained tracks donors better'.", + "how_to_recompute": "For each arm and state, compute Spearman over donors of gen_pct vs obsthin_pct in the per-donor CSV, dropping donors with an undefined pct; this reproduces summary spearman_donors exactly. The difference is pretrained minus scratch. For the CI, use the same idx matrix as within-1, compute Spearman on the resampled donor rows (ties averaged), average over the 13 states per replicate, and take the 2.5 and 97.5 percentiles.", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "weakened", + "reason": "All numbers reproduce exactly. Per-arm, per-state Spearman matches the summary to 1e-16. One bootstrap replicate in the 58-pair LoRA comparison is undefined (constant HSPC vector); mean and nanmean give the same CI. But 'similar' is inferred from non-significance despite a consistent tilt. All four mean differences favour pretrained (+0.019 to +0.154). vs full FT the bias-corrected interval excludes 0: [0.003, 0.333]. Both significant state-level calls favour pretrained and none favour scratch. A better phrasing: 'not significantly different, with a small consistent tilt towards pretrained (borderline vs full FT)'.", + "recomputed": "Mean rho over the 13 rare states, with percentile and BC CIs. 5 pairs: lora_n5 0.295 vs scratch_n5 0.187, +0.109 [-0.080, 0.294], BC [-0.079, 0.295]. 58 vs LoRA: real_lora 0.293 vs scratch_full 0.274, +0.019 [-0.121, 0.186], BC [-0.116, 0.200]. 58 vs full FT: real_full 0.428 vs scratch_full 0.274, +0.154 [-0.008, 0.313], BC [0.003, 0.333]. Recipes: lora_recipe 0.398 vs scratch_recipe 0.357, +0.041 [-0.106, 0.217], BC [-0.102, 0.232]. Significant state-level calls: Cytotoxic CD4 T at 5 pairs (0.743 vs -0.395, +1.137 [0.461, 1.603]) and Cycling T/NK vs full FT (0.466 vs -0.171, +0.637 [0.008, 1.111]). Per comparison 1/0/1/0 of 13 (2 of 52)." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "Reproduced exactly: mean Spearman 0.295 vs 0.187, 0.293 vs 0.274, 0.428 vs 0.274 and 0.398 vs 0.357, with matching CIs; 2 of 52 state-level differences are significant, both favouring pretrained. But 'similar' asserts equivalence, which the data cannot show. All four aggregate differences favour pretrained, as do 31 of the 52 state differences. Full FT is +0.154 with 95% CI [-0.008, 0.313] and 90% CI [0.014, 0.289]. The CI upper bounds (0.19-0.31) are as large as the mean Spearman values themselves. Low tracking is not just sampling noise: given the generated cell counts, a perfectly tracking generator would reach mean Spearman 0.78-0.85, and both arms reach only about a quarter to a half of that. Restricted to the 12 sepsis donors, the lean is mixed (0.205 vs 0.271, 0.271 vs 0.256, 0.271 vs 0.363, 0.406 vs 0.347), so part of the all-donor lean comes from the sepsis/control split. The fair reading is 'no detectable difference, both poor', not demonstrated similarity.", + "recomputed": "Mean Spearman differences (pretrained - scratch): +0.109 [-0.080, 0.294]; +0.019 [-0.121, 0.186]; +0.154 [-0.008, 0.313] (90% CI [0.014, 0.289]); +0.041 [-0.106, 0.217]. Significant states: Cytotoxic CD4 T at 5 pairs (0.743 vs -0.395, +1.137 [0.461, 1.603]) and Cycling T/NK vs full FT (0.466 vs -0.171, +0.637 [0.008, 1.111]). Noise ceilings: scratch 0.82/0.79/0.79/0.78, pretrained 0.85/0.80/0.82/0.80." + } + ] + }, + { + "id": "within-8", + "claim": "Within the cohort, the cohort's own training pairs, not pretraining, are what recover the sepsis monocyte states. Pretrained zero-shot (no pairs) is the least faithful arm (mean |log2| 1.763). Every scratch arm beats it on average: 58 pairs D=-0.350 [-0.660, -0.128]; recipes -0.457 [-0.729, -0.238]; 5 pairs -0.239 [-0.515, 0.005]. The gain comes from C1Q-high (zero-shot 0.025x vs scratch 0.467-1.005x), MS1 (0.084x vs 0.405-0.462x with 58 pairs) and Proliferating T label (3.878x vs 0.977-1.366x). Zero-shot stays closer for LDG (0.140x vs 0.043-0.059x) and Cycling activated T.", + "evidence": "within_bootstrap.csv, kind=reference: aggregate and state rows. within_arm_fidelity.csv aggregate row for zs__raw: abs_log2_pc 1.763.", + "how_to_recompute": "Same as within-1, with zs__raw as the 'pretrained' arm. Note that zs__raw is thinned to its own lower depth (thin 0.42-0.50 for marker states).", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "weakened", + "reason": "The core numbers reproduce exactly, and zero-shot is the least faithful arm. Recovery of C1Q-high by pairs is robust: all three scratch arms are significantly closer than zero-shot. Weak points: (i) 'Every scratch arm beats it' is n.s. at 5 pairs. (ii) The 58-pair aggregate advantage is fragile. It loses significance without the two label states (-0.227 [-0.614, 0.060]) and under donor-equal weighting (-0.226 [-0.666, 0.126]). Only recipes stays significant both ways. (iii) MS1 recovery is partial (0.40-0.46x) and absent at 5 pairs (0.096x vs 0.084x). (iv) The LDG range leaves out scratch_n5 at 0.129x, which is about equal to zero-shot's 0.140x (D=+0.121, n.s.). (v) Unmentioned: zero-shot is also significantly closer than scratch_n5 for Plasmablast (markers): 0.782x vs 0.350x, D=+1.159 [0.089, 1.551]. (vi) The note 'thin 0.42-0.50 for marker states' is wrong. Zero-shot marker-state thins span 0.416-0.699 (DN2/ABC B 0.681, Cytotoxic CD4 T 0.699). Finally, 'not pretraining' overreaches for C1Q: at 58 pairs, pretrained plus pairs keeps C1Q at 0.92-1.05x vs scratch 0.47x (BC-significant).", + "recomputed": "zs mean |log2| 1.763, the highest of the 8 arms. D vs zs: 5 pairs -0.239 [-0.515, 0.005] (BC [-0.507, 0.008]); 58 pairs -0.350 [-0.660, -0.128] (BC [-0.619, -0.101]); recipes -0.457 [-0.729, -0.238] (BC [-0.748, -0.251]). zs gen/obs vs scratch arms: C1Q 0.025 vs 0.467-1.005; MS1 0.084 vs 0.096 / 0.405 / 0.462; PT label 3.878 vs 0.977-1.366; LDG 0.140 vs 0.129 / 0.043 / 0.059. Cycling activated T: zs 0.646 vs 0.279-0.404, and zs is significantly closer in all 3 comparisons (+0.687 [0.026, 1.339], +1.041 [0.062, 1.920], +1.214 [0.207, 2.213]). Significant scratch-closer vs zs: C1Q 3/3, PT label 3/3, MS1 at 58 pairs and with recipes. Over the 11 marker states: 5 pairs -0.108 [-0.439, 0.168], 58 pairs -0.227 [-0.614, 0.060], recipes -0.332 [-0.682, -0.091]. Donor-weighted 13-state D: -0.157 [-0.536, 0.122], -0.226 [-0.666, 0.126], -0.354 [-0.757, -0.007]." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "The numbers reproduce, and the core conclusion is robust. Zero-shot is the least faithful of the 8 arms under all 7 scoring variants (mean |log2| 1.76-1.97). Its C1Q (0.02-0.03x vs scratch 0.35-1.06x) and MS1 (0.04-0.08x vs 58-pair scratch 0.30-0.51x) failures hold under every depth scheme and in 17 of 17 leave-one-donor-out runs. The 58-pair and recipe scratch aggregates beat zero-shot under every scheme; the one borderline case is the 58-pair arm at zero-shot depth (-0.293 [-0.645, -0.003], BC n.s.). Several stated details do not hold up. (1) 'Zero-shot stays closer for Cycling activated T' is a depth and composition artifact. Zero-shot T cells carry about 2.4x more marker-gene depth than their library predicts, so thinning observed cells to 0.46 by library over-thins them. With marker-depth matching, zero-shot's ratio falls from 0.65x to 0.35x (D vs scratch -0.22/+0.06/+0.32, all n.s.). After T-subtype standardisation (zero-shot makes 3.9x proliferating-labelled T cells) it is 0.23x, below scratch's 0.28-0.32x. The LDG exception does hold. (2) The 5-pair 'beats it on average' is not significant under the published scoring, under common depth, or in 13 of 17 leave-one-donor-out runs; it becomes significant only under marker-depth matching or composition standardisation. (3) The zero-shot thinning constants for marker states are 0.416-0.699 (DN2 0.681, cytotoxic CD4 0.699), not 0.42-0.50. (4) Pairs do not recover MS1 at 5 pairs (scratch 0.096x vs zero-shot 0.084x), and no arm recovers cycling monocytes.", + "recomputed": "Zero-shot mean |log2| 1.763 (worst; 1.761-1.966 across schemes). Scratch minus zero-shot D: 5 pairs -0.239 [-0.515, 0.005]; 58 pairs -0.350 [-0.660, -0.128]; recipes -0.457 [-0.729, -0.238]. At zero-shot depth (V1 all arms): -0.110/-0.293/-0.428. Marker depth (V2 all arms): -0.383*/-0.578*/-0.695*. Composition-standardised: -0.321*/-0.440*/-0.575*. Zero-shot Cycling activated T: 0.646x published, 0.346x under V2." + } + ] + }, + { + "id": "cross-1", + "claim": "With no sepsis pairs used for training, neither the from-scratch model nor the pretrained model fine-tuned on other cohorts (LoRA r16/r64 or full FT) generates the sepsis-defining rare monocyte states: C1Q-high monocytes come out at 0.03-0.09x and MS1 monocytes at 0.10-0.27x. Pretrained LoRA trained on the sepsis cohort's own pairs (cross-fitted, held-out donors) gets 0.50-0.53x and 0.64-0.71x. What recovers these states is target-cohort pairs on the pretrained model, not pretraining alone. Scratch vs pretrained cross-cohort LoRA is either a tie or a choice between two failures.", + "evidence": "C1Q-high (gen state/parent vs obs thinned state/parent): scratch s162000 0.068 [0.050,0.097] (86/20,808 vs 6,788/111,269); r16 0.035 [0.022,0.056] (41/22,524 vs 5,862/111,269); r64 0.056 [0.038,0.088]; full 108.5k 0.093 [0.063,0.141]; xcal 0.074 [0.051,0.111]; own-cohort LoRA-5 0.531 [0.317,0.822]; LoRA-58 0.502 [0.344,0.730]; zero-shot 0.026. MS1: scratch 0.102 [0.059,0.159]; r16 0.121 [0.079,0.184]; r64 0.113; full 0.265 [0.157,0.469]; xcal 0.724 [0.546,1.045]; own5 0.644 [0.214,1.263]; own58 0.707 [0.437,1.139]. D (scratch - r16): C1Q -0.96 [-1.55,-0.43] (scratch closer); MS1 +0.24 [-0.63,+1.12] (n.s.). MS1 adult sepsis minus adult HC, generated vs observed difference in pp: scratch +1.18 vs +18.97 (6%; 9.5% at r16 depth); r16 +1.23 vs +12.51 (10%); full +3.03 vs +11.89 (25%); xcal +8.85 vs +17.53 (50%); own58 +10.88 vs +17.50 (62%). C1Q contrasts: scratch reproduces -1% to +4% of the observed difference (paediatric reversed); r16 reverses all three; own58 reproduces 30-34%.", + "how_to_recompute": "out/sepsis_loco_per_donor.csv, filtered by arm and state: ratio = (sum gen_state / sum gen_parent) / (sum obsthin_state / sum obs_parent) over the 82 donors. CI: sort the donors; idx = np.random.default_rng(20261005).integers(0,82,(2000,82)); per row of idx, pooled sums with +0.5 added to both state sums; 2.5/97.5 percentiles of the ratio. These are cross_arm_ci.csv rows with variant=own_ref. D rows are in cross_bootstrap.csv (kind=main). Contrast: mean gen_pct over adult_sepsis donors minus mean over adult_HC donors, and the same for obsthin_pct (cross_contrasts.csv, columns gen_diff, obs_diff, gen_over_obs_diff).", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "weakened", + "reason": "All numbers reproduce to the stated precision: ratios, CIs, D values and contrasts. The conclusion 'what recovers these states is target-cohort pairs on the pretrained model' fails for MS1. The xcal arm (pretrained zero-shot plus cross-cohort calibration, no sepsis pairs, no training) gets MS1 to 0.724x [0.546,1.045], the same level as own-cohort LoRA (0.644x and 0.707x). xcal also reproduces 50% of the adult MS1 contrast (own58: 62%). It tracks MS1 across donors better than any of the 12 arms (rho 0.647 vs own58 0.571). Only for C1Q-high does the attribution hold: every arm without target-cohort pairs is at or below 0.093x (xcal 0.074, zs 0.026), and only own-cohort LoRA reaches about 0.5x. 'Tie or a choice between two failures' holds: C1Q D is significant but both arms are below 0.1x, and the MS1 difference is n.s. Minor: the MS1 range 0.10-0.27x leaves out full FT at 20k steps (0.292x).", + "recomputed": "C1Q ratio [bootstrap CI, +0.5]: s162k 0.068 [0.050,0.097] (86/20,808 vs 6,788/111,269); r16 0.035 [0.022,0.056] (41/22,524 vs 5,862/111,269); r64 0.056 [0.038,0.088]; full108.5k 0.093 [0.063,0.141]; full20k 0.084 [0.056,0.127]; xcal 0.074 [0.051,0.111]; zs 0.026 [0.017,0.041]; own5 0.531 [0.317,0.822]; own58 0.502 [0.344,0.730]. MS1: s162k 0.102 [0.059,0.159]; r16 0.121 [0.079,0.184]; r64 0.113 [0.080,0.166]; full108.5k 0.265 [0.157,0.469]; full20k 0.292 [0.188,0.475]; xcal 0.724 [0.546,1.045]; zs 0.080 [0.056,0.117]; own5 0.644 [0.214,1.263]; own58 0.707 [0.437,1.139]. D(s162k-r16): C1Q -0.962 [-1.549,-0.429]; MS1 +0.242 [-0.631,+1.123]. At matched depth: C1Q -1.174 [-1.749,-0.651]; MS1 -0.404 [-1.294,+0.471]. Adult MS1 contrast, gen vs obs pp (fraction): s162k +1.18 vs +18.97 (0.062; 0.095 at r16 depth); r16 +1.23 vs +12.51 (0.099); r64 +0.99 vs +11.87 (0.084); full108.5k +3.03 vs +11.89 (0.255); xcal +8.85 vs +17.53 (0.505); own58 +10.88 vs +17.50 (0.621). C1Q contrast fraction, adult/paediatric/all: s162k +0.041/-0.014/+0.012; r16 -0.042/-0.006/-0.013; own58 0.312/0.344/0.302. MS1 donor Spearman: xcal 0.647 (highest of 12 arms), own58 0.571, s162k 0.523. Files (all new, under out/scratch_vs_lora_20261005/): verify_recompute_cross_script.py and its outputs (_report.txt, _ratios.csv, _D_per_state.csv, _aggregate.csv, _spearman_arms.csv, _spearman_diff.csv, _contrasts.csv, _contrast_counts.csv, _depth.csv), plus verify_recompute_cross_extra_* and verify_recompute_cross_obsdepth_*. CPU only. No existing file was modified. No archive delta from this subagent." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "The C1Q half holds; the MS1 attribution fails. Every quoted ratio, CI, D and contrast reproduces exactly from sepsis_loco_per_donor.csv. C1Q is robust to every confound tried. Cross-cohort arms stay at 0.03-0.09x pooled, 0.03-0.09x generated-weighted or donor-mean, and 0.04-0.11x at detection-matched depth. The scratch-closer D vs r16 survives all 82 leave-one-donor-out re-bootstraps. Own-cohort LoRA-58 beats every cross-cohort arm in all 5 within-cohort folds (0.27-1.04 vs at most 0.18). The causal sentence fails for MS1. xcal (pretrained zero-shot + cross-cohort calibration; zero sepsis pairs, no training) reaches 0.72x pooled, 0.90x detection-matched and 0.61-1.37x per fold. It is at or above own-cohort LoRA-58 in folds 0, 1 and 3. So MS1 comes back with input calibration alone, without target-cohort pairs. The MS1 range is also understated: full FT 20k gives 0.29x (0.44x detection-matched). The LoRA-5 reference depends on folds and donors: per-fold C1Q 0.01-1.10 and MS1 0.01-2.87. Its donor-mean ratios are 0.23 (C1Q) and 0.31 (MS1), and one donor supplies 36% of its generated MS1 cells. Within-cohort fold 0 (17 donors), scratch trained on own pairs also reaches C1Q 0.47-1.0x and MS1 0.40-0.46x with 58 pairs. So the claim does not show that 'on the pretrained model' is necessary. 'Scratch vs cross-cohort LoRA is a tie or two failures' holds.", + "recomputed": "Outputs in /home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/rare_states_screen_20261004/out/scratch_vs_lora_20261005/verify_confound_cross_*.csv (script verify_confound_cross_analysis.py; this subagent published no archive delta). C1Q pooled: s162k 0.068 [0.050,0.097], r16 0.035 [0.022,0.056], r64 0.057, full108k 0.094, full20k 0.085, xcal 0.075, own5 0.531, own58 0.502. Generated-weighted: s162k 0.056, r16 0.033, own5 0.49, own58 0.45. Donor-mean: own5 0.23, own58 0.38. Detection-matched: s162k 0.094, r16 0.044, r64 0.070, full108k 0.105, own58 0.553. MS1 pooled: s162k 0.102, r16 0.121, r64 0.113, full108k 0.265, full20k 0.292, xcal 0.724 [0.546,1.045], own5 0.644, own58 0.707. MS1 detection-matched: xcal 0.897, own5 0.930, own58 1.009, full20k 0.440, s162k 0.172. D C1Q -0.962 [-1.549,-0.429] (82/82 LODO); D MS1 +0.242 n.s. MS1 per fold, xcal vs own58: 0.64/0.20, 0.88/0.71, 0.61/1.09, 1.37/1.21, 0.61/0.84. Contrasts: MS1 adult s162k +1.18 vs +18.97 pp; own58 +10.88 vs +17.50 pp. C1Q: s162k 1-4% (paediatric reversed), r16 reverses all three, own58 30-34%." + } + ] + }, + { + "id": "cross-2", + "claim": "Proliferating/cycling states are where pretraining clearly helps across cohorts. The final scratch model almost never generates cycling cells: cycling T/NK at 0.049x, against 0.15-0.46x for every pretrained cross-cohort arm. This holds in all 4 comparisons and survives the depth correction. Cycling activated T and cycling monocytes show the same direction, and it stays significant at matched depth vs r64 (activated T only), full FT and xcal.", + "evidence": "Cycling T/NK: scratch 35/72,018 cells = 0.049x [0.027,0.085], vs 1.00% observed (3,461/345,949). r16 0.151x [0.092,0.261]; r64 0.172x; full 0.356x; xcal 0.458x; own58 0.536x. D (scratch - pretrained), own reference: +1.62 [+0.99,+2.31] r16, +1.81 [+1.28,+2.43] r64, +2.86 [+2.27,+3.60] full, +3.22 [+2.72,+3.86] xcal. At r16 depth the scratch ratio is 0.064 and D vs r16 is +1.23 [+0.61,+1.91]. Cycling monocyte: scratch 21/20,808 = 0.024x [0.013,0.042] vs r16 0.062, r64 0.070, full 0.114, xcal 0.252, own-cohort 0.23-0.30; D vs r16 +1.34 [+0.48,+2.25], n.s. at matched depth (+0.48 [-0.39,+1.36]), still significant vs full (+1.27 [+0.41,+2.20]). Cycling activated T: 0.051x vs 0.098-0.245x. Proliferating T (label, no depth issue): scratch 0.532x vs r16 0.789x, D +0.57 [+0.09,+0.80]; n.s. vs r64 (1.04x) and full (1.13x).", + "how_to_recompute": "Ratios: as in cross-1 with state='Cycling T/NK' (parent T cell + NK cell), 'Cycling monocyte (Ye 2026)' or 'Cycling activated T (CD38/HLA-DR)'. D = |log2 r_scratch| - |log2 r_pre| using the shared idx matrix. Matched-depth variant: replace the scratch arm's per-donor obsthin_state with the comparator's obsthin_state for the same donor and state (obs_parent is identical across arms). Rows are in cross_bootstrap.csv with variant own_ref or pretrained_depth_ref.", + "status": "confirmed", + "votes": [ + { + "lens": "recompute", + "verdict": "confirmed", + "reason": "Every number reproduces. Cycling T/NK is significant against all 4 pretrained arms, both at the screen's own reference and at matched depth. Cycling activated T stays significant at matched depth against r64, full FT and xcal (n.s. against r16). Cycling monocytes stay significant against full FT and xcal (n.s. against r16 and r64). That is exactly what the claim says. Proliferating T (label) is significant only against r16. Caveats: (1) 'Clearly helps' rests on the marker-rule states. The depth-independent label state supports it in only 1 of 4 comparisons. By label, scratch makes proliferating T cells at 0.53x, so 'almost never generates cycling cells' holds only under the marker rule. (2) The range '0.15-0.46x for every pretrained cross-cohort arm' leaves out the zero-shot arm at 0.83x, which only strengthens the claim. (3) The result depends on the checkpoint. At s50k, cycling monocytes at matched depth beat r16 (D -0.778 [-1.586,-0.075]), while the cycling T/NK deficit remains (+0.911 [+0.324,+1.579]).", + "recomputed": "Cycling T/NK: s162k 35/72,018 = 0.049 [0.027,0.085]; observed 3,461/345,949 (1.00%). r16 0.151 [0.092,0.261]; r64 0.172 [0.107,0.281]; full20k 0.348 [0.226,0.563]; full108.5k 0.356 [0.236,0.563]; xcal 0.458 [0.298,0.736]; zs 0.830 [0.565,1.294]; own5 0.288; own58 0.536. D at own reference: r16 +1.623 [+0.989,+2.314]; r64 +1.812 [+1.279,+2.428]; full108.5k +2.857 [+2.271,+3.595]; xcal +3.219 [+2.718,+3.860]. Matched depth: s162k ratio 0.064 at r16 depth (0.069 at r64, 0.066 at full108.5k). D r16 +1.226 [+0.613,+1.907]; r64 +1.303 [+0.797,+1.897]; full108.5k +2.410 [+1.829,+3.139]; xcal (r16 reference) +2.822 [+2.318,+3.458]. Cycling monocyte: s162k 21/20,808 = 0.024 [0.013,0.042]; r16 0.062; r64 0.070; full108.5k 0.114; full20k 0.150; xcal 0.252; own5 0.304; own58 0.228. D vs r16 +1.344 [+0.478,+2.253]. Matched depth: r16 +0.482 [-0.386,+1.362]; r64 +0.604 [-0.152,+1.446]; full108.5k +1.268 [+0.413,+2.199]; xcal +2.477 [+1.652,+3.334]. Cycling activated T: s162k 0.051 [0.032,0.077] vs r16 0.098, r64 0.140, full20k 0.207, full108.5k 0.227, xcal 0.245 (zs 0.361). Matched depth: r16 +0.399 [-0.077,+0.966]; r64 +0.794 [+0.358,+1.278]; full108.5k +1.561 [+1.059,+2.187]; xcal +1.716 [+1.223,+2.271]. Proliferating T (label): s162k 0.532 [0.377,0.807]. vs r16 0.789, D +0.568 [+0.093,+0.795]; r64 1.037, D +0.856 [-0.296,+1.132]; full108.5k 1.127, D +0.737 [-0.410,+1.219]; xcal 0.642, D +0.272 [-0.044,+0.519]." + }, + { + "lens": "confound", + "verdict": "confirmed", + "reason": "The core claim holds under every confound tried, and all quoted values reproduce. For cycling T/NK, the scratch-worse verdict survives all 82 leave-one-donor-out re-bootstraps in all 4 comparisons. It also holds under generated-weighted and donor-mean pooling. Independent detection-matched depth gives the same answer: scratch 0.054x vs r16 0.143, r64 0.141, full 0.279 and xcal 0.404, with the comparator closer in all 4. At that depth, cycling activated T is also significant vs r16. Scratch is below r16, full FT and xcal in every one of the 5 donor folds. The direction replicates within-cohort (fold 0: scratch arms 0.42-0.51x vs pretrained 0.93-1.34x). Scratch also makes fewer cells with >=2 cycling genes across all labels (0.17% vs 0.21-0.56%), so this is not a parent-label artefact. Caveat 1: drop 'Proliferating T (label, no depth issue)' as cycling evidence. Only 4.7% of scratch's and 5.7% of r16's labelled cells carry >=2 cycling genes, against 37% of observed cells at r16 depth. That label result is significant only vs r16 (76/82 LODO), and scANVI labels on 20k-UMI / 236-gene cells are not obviously depth-neutral. Caveat 2: 'across cohorts' rests on one held-out cohort and one scratch initialisation seed.", + "recomputed": "Cycling T/NK: s162k 0.049 [0.027,0.085] (35/72,018 vs 3,461/345,949); r16 0.152, r64 0.173, full108k 0.357, xcal 0.459, own58 0.536. D own reference: r16 +1.62 [0.99,2.31], r64 +1.81, full +2.86, xcal +3.22. At r16 depth: +1.23 [0.61,1.91]. Detection-matched ratios: s162k 0.054, r16 0.143, r64 0.141, full108k 0.279, xcal 0.404. Cycling activated T, detection-matched: s162k 0.057 vs r16 0.080 (significant). Cycling monocyte: s162k 0.024; at matched depth n.s. vs r16 (+0.48), significant vs full (+1.27). Proliferating T label: 0.532 vs 0.789, D +0.568 [0.093,0.795]. Labelled proliferating T cells with >=2 cycling genes: s162k 4.7%, r16 5.7%, observed at r16 depth 36.8%. Files: verify_confound_cross_{comparisons,detmatched_*,per_fold,label_validity}.csv." + } + ] + }, + { + "id": "cross-3", + "claim": "Pooled over the 13 informative non-control states, the final scratch model is further from the observed frequencies than every pretrained cross-cohort arm under the screen's definition. The gap to LoRA, however, is entirely depth-driven. Against r16 and r64 the scratch deficit is 0.30-0.41 log2 units, about 1.2-1.3-fold, and it vanishes at matched depth. Against full FT (+0.25) and especially the zero-pair calibration control xcal (+0.68), a deficit remains.", + "evidence": "Mean |log2 ratio| [95% CI], own reference: scratch 2.615 [2.428,2.865]; r16 2.315 [2.122,2.537]; r64 2.210 [2.061,2.442]; full 108.5k 1.923 [1.742,2.140]; full 20k 1.749; xcal 1.533 [1.335,1.757]; zero-shot 1.824; own-cohort LoRA-5 1.535 [1.261,1.923]; LoRA-58 1.287 [1.049,1.610]. Mean paired D (scratch - pretrained): r16 +0.300 [+0.093,+0.520]; r64 +0.405 [+0.171,+0.618]; full +0.692 [+0.417,+1.000]; xcal +1.082 [+0.920,+1.289]. At matched depth: -0.106 [-0.243,+0.112], -0.044 [-0.208,+0.131], +0.245 [+0.030,+0.539], +0.676 [+0.551,+0.891]. Per-state significant counts, own reference (pretrained closer / scratch closer / n.s.): r16 5/1/7, r64 4/1/8, full 6/2/5, xcal 6/1/6. At matched depth: 3/3/7, 2/4/7, 3/3/7, 5/2/6. States within 2-fold of observed: scratch 5/13 (6 at matched depth), r16 6, r64 5, full 3, xcal 5, own58 7.", + "how_to_recompute": "cross_bootstrap_aggregate.csv: kind=arm gives mean_abs_log2; kind=comparison_main gives mean_D. Mean over the 13 non-control states of |log2(((sum gen_state + 0.5)/sum gen_parent) / ((sum obsthin_state + 0.5)/sum obs_parent))|; CI from the same 2000 donor resamples (seed 20261005), averaging the 13 states within each replicate. Per-state counts: cross_bootstrap.csv, informative & ~is_control, grouped by comparison and variant, verdict counts.", + "status": "confirmed", + "votes": [ + { + "lens": "recompute", + "verdict": "confirmed", + "reason": "Every number reproduces exactly, including all mean |log2| values and CIs, the 8 mean-D CIs, the 8 per-state verdict tallies and the within-2-fold counts. An independent observed-side test backs the matched-depth premise. Thinning observed cells until their marker detection matches scratch takes about 0.25 of observed depth for monocytes and 0.70 for T cells, inside the pretrained arms' ranges (0.21-0.29 and 0.62-0.88) rather than at 1.0. Caveats: (1) The result depends on the checkpoint. At matched depth the s50k checkpoint is significantly closer than r16 and r64 and ties full FT, so 'a deficit remains vs full FT' holds only for s162k. (2) The full-FT deficit at matched depth is marginal (CI lower bound +0.030). (3) The matched-depth tie with LoRA is an aggregate. Scratch's significant wins are on states where every arm fails (C1Q, DN2, LDG), and its losses are on cycling, proliferating and plasmablast-marker states. (4) The xcal matched-depth figure uses the r16 reference for scratch. With xcal's own depth the deficit is larger (+0.951).", + "recomputed": "Mean |log2| over the 13 states (+0.5) [CI]: s162k 2.615 [2.428,2.865]; r16 2.315 [2.122,2.537]; r64 2.210 [2.061,2.442]; full108.5k 1.923 [1.742,2.140]; full20k 1.749 [1.602,1.952]; xcal 1.533 [1.335,1.757]; zs 1.824 [1.663,2.013]; own5 1.535 [1.261,1.923]; own58 1.287 [1.049,1.610]. Mean D at own reference: r16 +0.300 [+0.093,+0.520]; r64 +0.405 [+0.171,+0.618]; full108.5k +0.692 [+0.417,+1.000]; xcal +1.082 [+0.920,+1.289]; full20k +0.866 [+0.616,+1.104]; zs +0.791 [+0.564,+1.049]. Matched depth: r16 -0.106 [-0.243,+0.112]; r64 -0.044 [-0.208,+0.131]; full108.5k +0.245 [+0.030,+0.539]; xcal +0.676 [+0.551,+0.891] (with xcal's depth: +0.951 [+0.799,+1.160]); full20k +0.413 [+0.227,+0.650]. Per-state tallies (pretrained closer / scratch closer / n.s.), own reference: r16 5/1/7, r64 4/1/8, full 6/2/5, xcal 6/1/6. Matched depth: 3/3/7, 2/4/7, 3/3/7, 5/2/6. States within 2-fold: s162k 5 (6 at r16 depth), r16 6, r64 5, full108.5k 3, full20k 6, xcal 5, zs 5, own5 5, own58 7. Log2 deficits 0.300 and 0.405 = 1.23x and 1.32x. s50k at matched depth: vs r16 -0.386 [-0.540,-0.173]; vs r64 -0.328 [-0.491,-0.168]; vs full108.5k -0.035 [-0.240,+0.243]." + }, + { + "lens": "confound", + "verdict": "confirmed", + "reason": "The numbers reproduce exactly and the central reading is robust. Matched to the LoRA depth, the scratch-vs-r16/r64 gap stays null under four treatments. Those are pooled (-0.106 / -0.044), generated-weighted (+0.004 / +0.008), donor-mean (+0.049 / +0.089) and my detection-matched depth (-0.008 [-0.20,+0.20] / +0.12 n.s.). It is also stable to dropping any one donor. The deficit vs xcal is robust everywhere (+0.68 matched, +0.68 detection-matched). Caveats: (i) The own-reference deficit vs r16 (+0.30) loses significance if any one of cycling T/NK, plasmablast markers or cycling monocyte is dropped. (ii) The full-FT matched deficit (+0.245 [0.030,0.539]) loses significance if any of cycling T/NK, cycling activated T, MS1 or cycling monocyte is dropped. It stays significant under generated-weighted (+0.39), donor-mean (+0.39) and detection-matched (+0.28 [0.02,0.51]) estimators, so the full-FT deficit is the cycling and MS1 states. (iii) The ranking is specific to the final checkpoint. s50k ties r16 at own reference (+0.01) and beats it at matched depth (-0.39).", + "recomputed": "Mean |log2|: s162k 2.615 [2.428,2.865], r16 2.315, r64 2.210, full108k 1.923, full20k 1.749, xcal 1.533, zs 1.824, own5 1.535, own58 1.287. Mean D, own reference: r16 +0.300 [0.093,0.520], r64 +0.405, full +0.692, xcal +1.082. Matched depth: -0.106 [-0.243,0.112], -0.044, +0.245 [0.030,0.539], +0.676. Detection-matched: -0.008, +0.121, +0.275 [0.016,0.514], +0.680. LODO verdict stability 1.0 in every case. States within 2-fold: scratch 5 (6 at r16 depth), r16 6, r64 5, full108k 3, xcal 5, own58 7. Per-state verdict counts reproduce: own reference 5/1/7, 4/1/8, 6/2/5, 6/1/6; matched depth 3/3/7, 2/4/7, 3/3/7, 5/2/6. Files: verify_confound_cross_aggregate.csv and verify_confound_cross_detmatched_aggregate.csv." + } + ] + }, + { + "id": "cross-4", + "claim": "The screen's library-based depth matching does not hold for the cross-cohort scratch model. Its cells carry a 20,606-UMI median library, 18x the pretrained arms' and 10x the observed 1,967, but only 236 detected genes. Their marker-gene detection matches the ~1,000-UMI pretrained arms. The screen therefore left the observed cells unthinned for scratch (thin = 1.0, vs 0.34-0.62 for the pretrained arms). This biases the scratch ratios for marker-rule states downward by 1.1-2.0x. Corrected, the scratch ISG states reach about 1.0x, while the cycling, C1Q-high and MS1 deficits remain.", + "evidence": "Median library on measured genes: scratch s50k 13,891, s100k 19,138, s150k 20,524, s162k 20,606; r16 1,135; r64 1,013; full 20k 1,070; full 108.5k 1,055. Benchmark observed median 1,967 UMIs and 411 detected genes; generated detected genes: scratch 236, r16 243, r64 232, full 224/232. Monocyte marker panel: mean detected markers per cell (of 73) scratch 5.94 vs 5.29-6.53; UMIs per detected marker 2.79 vs 2.54-2.70; HLA-DRA detected in 65.1% of scratch monocytes vs 64.0/60.4/64.5% (r16/r64/full). T cells: NKG7 50.0% vs 50.8/49.6/55.0%; GZMA 43.1% vs 42.9/42.3/43.9%. Screen thin constant: 1.0 for all scratch arms; monocytes 0.398/0.376/0.338/0.370 and T cells 0.601/0.539/0.619/0.578 for r16/r64/F20k/F108k. Scratch ratio, own reference > at r16 depth: ISG-high T 0.599 > 1.001 (0.98-1.14 across the 4 pretrained references); ISG-high monocyte (control) 0.473 > 0.952; MS1 0.102 > 0.160; Plasmablast markers 0.338 > 0.530; DN2 0.163 > 0.244; C1Q 0.068 > 0.078; cycling T/NK 0.049 > 0.064; cycling monocyte 0.024 > 0.044. Label states are unchanged.", + "how_to_recompute": "cross_depth_check.csv is computed from work/sepsis/loco/.npz (lib_measured, markers, aifi labels; parents via tools/screen.py parent_mask) and plans/paired_crosscohort_20261002/scores/ye2026_sepsis//summary.json (median_umi_per_cell_scoring_genes, precision_coverage_depth_matched.detected_median_before). Matched-depth ratios: per-donor CSV with the scratch arm's gen counts over the comparator's obsthin_state/obs_parent; these are cross_arm_ci.csv rows with variant=pretrained_depth_ref.", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "confirmed", + "reason": "Every library, detected-gene, marker-detection and thin-constant value reproduces from the npz files and summary.json. So do the shifts from own reference to matched depth. The thin constants are internally consistent: the implied observed median library is 2,819-2,822 for monocytes and 1,980-1,983 for T cells in every arm, matching the observed cells directly (2,821 and 1,981). An independent observed-side test supports the core inference. Thinning observed cells until their marker detection matches the scratch arm takes about 0.25 of observed depth for monocytes (0.56 by HLA-DRA) and about 0.70 for T cells. The pretrained arms need 0.21-0.29, 0.44-0.54 and 0.62-0.88, and none of these is near the screen's thin of 1.0 for scratch. If anything, the correction is conservative for monocyte states. Minor imprecisions: the library is 18-20x the pretrained arms', not 18x. Pretrained thin constants span 0.34-0.78 across all marker parents; the quoted 0.34-0.62 covers monocyte and T-cell parents only. The correction factor for rare states is 1.12-1.82 at the r16 reference and up to 2.09 across references. Generated T-cell NKG7 and GZMA detection (42-55%) exceeds observed native detection (36% and 32%), so those two genes show scratch \u2248 pretrained but are not depth read-outs in themselves.", + "recomputed": "Median library on measured genes (npz, equal to the benchmark): s50k 13,891; s100k 19,138; s150k 20,524; s162k 20,606; r16 1,135; r64 1,013; full20k 1,070; full108.5k 1,055. s162k is 18.2-20.3x the pretrained arms and 10.5x the observed 1,967. Detected genes: s162k 236, r16 243, r64 232, full 224/232, observed 411. Monocytes, mean detected markers (of 73): s162k 5.942 vs 5.291-6.525. UMIs per detected marker: 2.789 vs 2.536-2.700. HLA-DRA: 65.1% vs 64.0 (r16) / 60.4 (r64) / 64.5 (full108.5k). T cells: NKG7 50.0% vs 50.8/49.6/55.0; GZMA 43.1% vs 42.9/42.3/43.9. Thin: 1.0 for all scratch arms. Monocytes 0.398/0.376/0.338/0.370 and T cells 0.601/0.539/0.619/0.578 for r16/r64/full20k/full108.5k; all marker parents 0.338-0.782. s162k own ratio, then at r16 depth: ISG-high T 0.599 to 1.001 (0.978-1.141 across the 4 references); ISG-high monocyte 0.473 to 0.952; MS1 0.102 to 0.160; Plasmablast markers 0.338 to 0.530; DN2 0.163 to 0.244; C1Q 0.068 to 0.078; cycling T/NK 0.049 to 0.064; cycling monocyte 0.024 to 0.044; also HSPC 1.313 to 1.472, Cytotoxic CD4 T 0.940 to 1.073, LDG 0.012 to 0.021, activated T 0.051 to 0.074. Correction factor, rare marker states: 1.12-1.82 at r16, 1.09-2.09 across the 4 references; controls 2.01-2.46. Label states identical in every arm (4,793 and 8,423 cells). Observed side (native C1Q 6,788 and MS1 18,712 reproduced): native monocytes 12.0 detected markers and 74.4% HLA-DRA. Detection-matched thinning: s162k 0.25 (markers) / 0.56 (HLA-DRA); r16 0.22/0.53; r64 0.21/0.44; full20k 0.24/0.45; full108.5k 0.29/0.54. T-cell markers: s162k 0.70 vs 0.62-0.88. Observed native T-cell NKG7 36.3%, GZMA 32.3%." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "The core is confirmed; the size of the correction is not robust. The library and detection facts reproduce from the per-cell extracts and the benchmark summary.json files. My independent detection-matched thinning confirms that scratch's effective marker depth is close to the pretrained arms', not to its 20k library. The scratch/r16 depth ratio is 0.75-1.4 across parents (monocytes 0.188 vs 0.154; all cells 0.564 vs 0.563). But the same check shows the library-based thinning is also off for the pretrained arms. For monocytes, detection-matched thinning is 0.15-0.22 against library 0.34-0.40; for T cells it is 0.66-1.0 against 0.54-0.62. So every 'corrected' ratio carries an unknown bias shared by all arms. The scratch bias factor depends on the reference. It is 1.1-1.8x at r16 library depth (controls 2.0-2.1) but 1.1-3.2x at scratch's own detection-matched depth (cycling monocyte 3.2, DN2 2.9, controls 3.7-4.7). 'ISG states about 1.0x' is specific to the reference: detection-matched ISG-high T is 0.67x and the ISG-high monocyte control is 2.2x. The cycling, C1Q and MS1 deficits remain under every reference (0.05-0.17x). Minor: the gap is 18-20x, not 18x. The thinning range 0.34-0.62 covers only monocyte and T parents; B and CD4 parents reach 0.78, and xcal 0.77-0.99.", + "recomputed": "Median library: s50k 13,891, s100k 19,138, s150k 20,524, s162k 20,606; r16 1,135, r64 1,013, full20k 1,070, full108k 1,055; observed 1,967. Detected genes: scratch 236 vs 243/232/224/232; observed 411. Monocytes, mean detected markers of 73: s162k 5.94 vs r16 5.50, r64 5.29, full20k 5.77, full108k 6.53; observed native 12.0. HLA-DRA 65.1% vs 64.0/60.4/64.5%. T cells: NKG7 50.0% vs 50.8/49.6/55.0%; GZMA 43.1% vs 42.9/42.3/43.9%. Detection-matched thinning, monocytes: s162k 0.188, r16 0.154, r64 0.153, full108k 0.217 (library thinning 1.0/0.398/0.376/0.370). Scratch ratio, own reference > r16 depth > detection-matched: ISG-high T 0.599 > 1.002 > 0.669; ISG-high monocyte 0.473 > 0.952 > 2.230; MS1 0.102 > 0.160 > 0.172; C1Q 0.068 > 0.079 > 0.094; cycling T/NK 0.049 > 0.065 > 0.054; cycling monocyte 0.024 > 0.045 > 0.080. Files: verify_confound_cross_{depth_detection,depth_readouts,detmatched_ratios}.csv." + } + ] + }, + { + "id": "cross-5", + "claim": "Where the pretrained arms over-generate, the scratch model is closer to the observed frequency: HSPC (vs r64 and full FT), Cytotoxic CD4 T (vs xcal; also vs r64 and full at matched depth) and Plasmablast label (vs full FT). These are the only states where scratch beats pretrained significantly apart from the uniformly failed C1Q-high, DN2 and LDG.", + "evidence": "HSPC (observed 629-741 cells): scratch 1.313x [1.034,1.729] vs r64 2.957x [2.062,4.233] and full 2.966x [2.093,4.220]. D -1.17 [-1.49,-0.81] and -1.17 [-1.53,-0.79]; at matched depth -0.95 [-1.26,-0.60] and -0.98 [-1.34,-0.60]. vs r16 1.617x: n.s. Cytotoxic CD4 T: scratch 0.940x [0.710,1.263] vs xcal 2.024x, D -0.93 [-1.28,-0.26]; r64 1.701x and full 1.609x are significant only at matched depth (-0.57 [-0.87,-0.12], -0.58 [-0.88,-0.06]). Plasmablast label: scratch 0.875x [0.669,1.213] vs full 2.214x [1.474,3.605], D -0.95 [-1.65,-0.03]; vs r16 1.377x and r64 1.010x: n.s. ISG-high T: the pretrained arms are at 1.35-2.05x vs scratch 0.60x (1.00x at matched depth), all n.s. because the CIs are wide (e.g. r16 1.662 [1.007,3.119]).", + "how_to_recompute": "cross_bootstrap.csv rows kind=main with state in {HSPC, Cytotoxic CD4 T, Plasmablast (label), ISG-high T}. Ratios and CIs are in cross_arm_ci.csv. Recompute from the per-donor CSV as in cross-1.", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "weakened", + "reason": "The HSPC, Cytotoxic CD4 T, Plasmablast-label and ISG-high T numbers all reproduce. The exclusivity sentence fails under the matched-depth convention the claim set uses for xcal: scratch at r16 depth, the same variant that gives cross-3's xcal tally of 5/2/6. In that variant Plasmablast (markers) is a further significant scratch-closer state. Scratch is at 0.530x vs xcal 0.367x, D -0.530 [-0.866,-0.208]. Both arms under-generate here, so it is not a case 'where the pretrained arms over-generate', and the state is not C1Q, DN2 or LDG. It is n.s. at the own reference and with xcal's depth as the scratch reference. 'Uniformly failed' also overstates DN2: xcal 0.30x, own5 0.42x, and zero-shot 1.49x.", + "recomputed": "HSPC (observed thinned 629-741 cells): s162k 1.313 [1.034,1.729]; r16 1.617 [1.149,2.310]; r64 2.957 [2.062,4.233]; full108.5k 2.966 [2.093,4.220]; full20k 1.403; xcal 1.681. D vs r64 -1.169 [-1.494,-0.810] and vs full108.5k -1.174 [-1.526,-0.794]. At matched depth: -0.951 [-1.262,-0.595] and -0.980 [-1.342,-0.602]. vs r16 -0.300 [-0.635,+0.086], n.s. Cytotoxic CD4 T: s162k 0.940 [0.710,1.263]. xcal 2.024, D -0.932 [-1.280,-0.261]. r64 1.701: own -0.681 [-1.043,+0.078] n.s., matched -0.573 [-0.872,-0.120]. full108.5k 1.609: own -0.601 [-0.978,+0.111] n.s., matched -0.580 [-0.876,-0.058]. Plasmablast (label): s162k 0.875 [0.669,1.213]; full108.5k 2.214 [1.474,3.605], D -0.954 [-1.652,-0.029]; r16 1.377, D -0.269 n.s.; r64 1.010, D +0.178 n.s. ISG-high T: pretrained arms 1.352-2.051 (r16 1.662 [1.007,3.119]); s162k 0.599 (1.001 at r16 depth); all n.s. at own reference and matched depth. Every significant scratch-closer state (13 rare states x {r16, r64, full108.5k, xcal} x {own, matched}): r16 own C1Q; r16 matched C1Q, DN2, LDG; r64 own HSPC; r64 matched C1Q, DN2, HSPC, Cytotoxic CD4 T; full own Plasmablast label, HSPC; full matched Plasmablast label, HSPC, Cytotoxic CD4 T; xcal own Cytotoxic CD4 T; xcal matched (r16 reference) Plasmablast (markers) [0.530 vs 0.367, D -0.530 [-0.866,-0.208]] and Cytotoxic CD4 T. Plasmablast (markers) vs xcal is +0.117 [-0.232,+0.473] at own reference and -0.109 [-0.447,+0.235] with xcal's depth." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "HSPC and Cytotoxic CD4 T reproduce and are robust. They survive all 82 leave-one-donor-out re-bootstraps, all three pooling estimators and detection-matched depth. Under donor-balanced pooling, Cytotoxic CD4 T is even significant at own reference vs r64 and full FT. But the exhaustive 'only states' sentence is false. In the same cross_bootstrap.csv, Plasmablast (markers) vs xcal at matched depth is scratch-closer and significant: D -0.53 [-0.87,-0.21], 82/82 LODO, and also at detection-matched depth (0.54 vs 0.31). That is an under-generation case (xcal 0.37x), outside the 'pretrained over-generates' framing. Plasmablast label vs full FT is fragile: the upper CI is -0.03, and it is significant in only 61/82 leave-one-donor-out re-bootstraps. Scratch's HSPC advantage depends on the checkpoint: s100k sits at 2.58x, as high as r64 and full. How close Cytotoxic CD4 T is to 1 depends on weighting (pooled 0.94x, donor-mean 1.56x). Scratch stays closer than every comparator either way.", + "recomputed": "HSPC: s162k 1.315 [1.034,1.729], r64 2.958, full108k 2.967, r16 1.619. D -1.169 and -1.174; at matched depth -0.951 and -0.980; 82/82 LODO. Cytotoxic CD4 T: s162k 0.942, xcal 2.026, D -0.932 [-1.280,-0.261]. At matched depth: vs r64 -0.573 [-0.872,-0.120] (82/82), vs full -0.580 [-0.876,-0.058] (71/82). Plasmablast label vs full: 0.875 vs 2.214, D -0.954 [-1.652,-0.029] (61/82 LODO). Omitted case: Plasmablast markers vs xcal at matched depth, 0.530 vs 0.367, D -0.530 [-0.866,-0.208]. ISG-high T: all n.s. (r16 1.662 [1.007,3.119]). Files: verify_confound_cross_comparisons.csv and verify_confound_cross_ratios.csv." + } + ] + }, + { + "id": "cross-6", + "claim": "Donor-level tracking of rare-state frequency (Spearman over 82 donors of generated % vs observed-thinned %) is statistically indistinguishable between scratch and pretrained LoRA: no state differs vs r16 or r64. Full FT and xcal track 2 states better each. Every cross-cohort arm tracks donors only modestly (rho mostly 0.0-0.55). The own-cohort LoRA tracks best.", + "evidence": "Scratch minus r16: 0/13 significant, range -0.150 (ISG-high T 0.383 vs 0.533) to +0.246 (cycling activated T 0.163 vs -0.082; LDG 0.183 vs -0.063), mean +0.057. vs r64: 0/13, mean -0.002. vs full: Cytotoxic CD4 T 0.404 vs 0.616, -0.211 [-0.422,-0.001]; DN2 0.105 vs 0.390, -0.286 [-0.550,-0.013]. vs xcal: cycling monocyte 0.072 vs 0.417, -0.345 [-0.615,-0.061]; ISG-high T 0.383 vs 0.552, -0.169 [-0.348,-0.012]. Scratch rho on the key states: MS1 0.523, C1Q 0.100, cycling T/NK -0.054. Own-cohort LoRA-58: ISG-high T 0.699, Cytotoxic CD4 T 0.692, MS1 0.571, DN2 0.423, proliferating T 0.386, cycling T/NK 0.332.", + "how_to_recompute": "Per arm and state: spearmanr(gen_pct, obsthin_pct) over donors where both are non-NaN; this reproduces sepsis_loco_summary.csv spearman_donors. Difference CI: the same idx matrix (seed 20261005, 2000x82); per replicate, Spearman (average ranks) of each arm on the resampled donors, then the difference and 2.5/97.5 percentiles. Rows are in cross_spearman.csv (kind=main, variant=own_ref).", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "confirmed", + "reason": "All per-arm Spearman values (they match sepsis_loco_summary.csv), the difference CIs and the significance counts reproduce. Caveats: (1) 'Statistically indistinguishable' means no significant difference, not equivalence. The difference CIs are about \u00b10.24 wide (full widths 0.33-0.65). The four 'better' states against full FT and xcal only just clear significance (upper bounds -0.001 to -0.061), out of 52 tests where about 2.6 false positives are expected. (2) 'The own-cohort LoRA tracks best' holds only for LoRA-58: mean rho 0.360, highest of all arms, but top in only 5 of 13 states. LoRA-5 (mean 0.244) is no better than the cross-cohort arms, and xcal tracks MS1 better than LoRA-58 (0.647 vs 0.571).", + "recomputed": "s162k minus r16: 0/13 significant. Range -0.150 (ISG-high T 0.383 vs 0.533) to +0.246 (LDG 0.183 vs -0.063; cycling activated T 0.163 vs -0.082 = +0.245); mean +0.057. s162k minus r64: 0/13, range -0.210 to +0.179, mean -0.002. vs full108.5k: Cytotoxic CD4 T 0.404 vs 0.616, -0.211 [-0.422,-0.001]; DN2 0.105 vs 0.390, -0.286 [-0.550,-0.013]. vs xcal: cycling monocyte 0.072 vs 0.417, -0.345 [-0.615,-0.061]; ISG-high T 0.383 vs 0.552, -0.169 [-0.348,-0.012]. vs full20k: 0/13. vs own58: 5/13 (own58 higher). Scratch rho: MS1 0.523, C1Q 0.100, cycling T/NK -0.054. own58 rho: ISG-high T 0.699, Cytotoxic CD4 T 0.692, MS1 0.571, DN2 0.423, proliferating T 0.386, cycling T/NK 0.332. Mean rho over the 13 states: own58 0.360, xcal 0.298, s100k 0.261, own5 0.244, full20k 0.243, full108.5k 0.242, r64 0.238, s162k 0.236, zs 0.233, s50k 0.231, s150k 0.226, r16 0.179. Cross-cohort arms (130 values): 114 in [0, 0.55], 11 below 0, 5 above 0.55 (max xcal MS1 0.647)." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "All rho values and significance verdicts reproduce. 'Indistinguishable' is a low-power null; it does not show parity. The median 95% CI width of the scratch-minus-comparator rho is 0.45. For C1Q, MS1, the three cycling states and LDG, 55-79% of donors have zero generated scratch state cells (41-83% for r16), so rank ties dominate. At matched depth one state does differ vs r64: HSPC, with scratch higher (0.295 vs 0.096, CI [0.013,0.392]). 'Own-cohort LoRA tracks best' holds only for LoRA-58 on average (mean rho 0.360 vs xcal 0.298, at most 0.26 for the others). xcal beats LoRA-58 on MS1 (0.647 vs 0.571, a state the claim lists), cycling monocyte (0.417 vs 0.128), plasmablast label (0.413 vs 0.195) and plasmablast markers. LoRA-5 (mean 0.244) is no better than the cross-cohort arms. 'rho mostly 0.0-0.55' holds; the maxima are xcal MS1 0.647 and full FT Cytotoxic CD4 T 0.616.", + "recomputed": "s162k minus r16: 0/13 significant, mean +0.057, range -0.150 to +0.246. s162k minus r64: 0/13 at own reference, mean -0.002; HSPC significant at matched depth. vs full: DN2 -0.286 [-0.550,-0.013], Cytotoxic CD4 T -0.211 [-0.422,-0.001]. vs xcal: cycling monocyte -0.345 [-0.615,-0.061], ISG-high T -0.169 [-0.348,-0.012]. vs own58: 5 states significant. Mean rho over 13 states: own58 0.360, xcal 0.298, s100k 0.261, own5 0.244, full20k 0.243, full108k 0.242, r64 0.238, s162k 0.236, zs 0.233, s50k 0.231, s150k 0.226, r16 0.179. Files: verify_confound_cross_spearman.csv and verify_confound_cross_spearman_diff.csv." + } + ] + }, + { + "id": "cross-7", + "claim": "The scratch model keeps the direction of sepsis-vs-control rare-state differences about as often as pretrained cross-cohort FT: 14/21 pairs vs 10-16/21. The kept effects, however, are 3-4x more compressed, at a median of 4% of the observed difference vs 13-15% for LoRA/full FT, 23% for xcal and 34% for the own-cohort LoRA-58.", + "evidence": "Common set: 21 (state, contrast) pairs over the 13 rare states with |obs_diff| >= 0.5 pp and the same sign in all 12 arms. Kept: scratch s162000 14 (s50k 15, s100k 15, s150k 16); r16 11; r64 10; full 20k 15; full 108.5k 16; xcal 19; zero-shot 15; own5 17; own58 19. Median gen_diff/obs_diff of the kept pairs: scratch 0.042 (0.067 at r16 depth); r16 0.137; r64 0.131; full 108.5k 0.150; full 20k 0.290; xcal 0.229; own5 0.245; own58 0.344. Every arm loses 2 pairs: Proliferating T label all-vs-all (obs -0.62 pp) and Plasmablast markers paediatric (obs -1.18 pp). Scratch loses but at least one pretrained arm keeps: C1Q paediatric (full, xcal keep); LDG paediatric (r16, full, xcal keep); ISG-high T all-vs-all (full, xcal keep). Cytotoxic CD4 T adult and all-vs-all are reversed by scratch, r16, r64 and full 108.5k; xcal and own-cohort keep them. Scratch keeps where LoRA reverses: C1Q adult (r16 and full reverse), C1Q all-vs-all (r16, r64), LDG adult and all-vs-all (r16, r64), cycling monocyte paediatric (r64). These scratch keeps are tiny effects (C1Q 1-4% of observed).", + "how_to_recompute": "From the per-donor CSV: per arm, state and contrast, obs_diff = mean obsthin_pct(group A) - mean obsthin_pct(group B) and gen_diff likewise. Group A/B: adult_sepsis vs adult_HC; ped_sepsis vs ped_HC; {adult_sepsis, ped_sepsis} vs {adult_HC, adult_ICU_ctrl, ped_HC, ped_ICU_ctrl}. same_direction = sign equality (this reproduces sepsis_loco_contrasts.csv). Common set = |obs_diff| >= 0.5 in all 12 arms with a consistent sign. Results are in cross_contrasts.csv (eligible_common) and cross_contrast_counts.csv.", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "weakened", + "reason": "The common set (21 pairs), kept counts, medians and pair-level keep/lose lists reproduce. 'Keeps the direction about as often' holds. 'Kept effects 3-4x more compressed vs LoRA' holds only under the screen's definition with medians taken over different kept sets, and does not survive adversarial checks. On the pairs kept by both arms, the gap to r16 shrinks to 2.3x (CI 1.08-5.39) and to r64 to 2.2x (CI 0.87-4.90, n.s.). At the matched depth that cross-3 and cross-4 argue is correct for scratch, the gap to LoRA is about 2x and n.s. (r16 CI 0.76-7.11; r64 0.58-6.22), and on shared pairs it is 1.2-1.3x, n.s. Over all 21 common pairs with reversals counted as negative, scratch's median fraction (0.024) is above r16 (0.005) and r64 (0.000). Compression relative to full FT, xcal and own-cohort LoRA-58 is robust across all variants. The '13-15% for LoRA/full FT' range also leaves out full FT at 20k steps (29%).", + "recomputed": "Common set: 21 pairs. Kept: s162k 14 (s50k 15, s100k 15, s150k 16); r16 11; r64 10; full20k 15; full108.5k 16; xcal 19; zs 15; own5 17; own58 19. Median gen_diff/obs_diff over kept pairs: s162k 0.042 (0.067 at r16 depth); r16 0.137; r64 0.131; full108.5k 0.150; full20k 0.290; xcal 0.229; zs 0.092; own5 0.245; own58 0.344. All arms lose Proliferating T label all-vs-all (obs -0.623) and Plasmablast markers paediatric (obs -1.18 to -1.67; -1.67 for scratch). Other keep/lose patterns as stated; C1Q all-vs-all is also reversed by full20k. Robustness (comparator median / scratch median; donor bootstrap, 2000 reps, seed 20261005). Each arm's own kept set: r16 3.29 [1.22,11.76]; r64 3.15 [1.00,9.66]; full108.5k 3.59 [1.88,9.43]; full20k 6.97 [2.52,13.28]; xcal 5.50; own58 8.27. Pairs kept by both: r16 2.32 [1.08,5.39]; r64 2.20 [0.87,4.90]; full108.5k 6.10 [2.15,8.90]. Scratch at r16 depth: r16 2.04 [0.76,7.11]; r64 1.95 [0.58,6.22]; full108.5k 2.23 [1.15,5.88]. r16 depth and shared pairs: r16 1.29 [0.66,3.23]; r64 1.22 [0.52,3.04]; full108.5k 3.55 [1.18,5.19]; xcal 5.98; own58 4.58. Signed median over all 21 common pairs: s162k 0.024; r16 0.005; r64 -0.000; full108.5k 0.106; full20k 0.170; xcal 0.147; own58 0.312." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "The counts and medians reproduce, as do every listed kept, lost and reversed pair. Direction parity is real but uninformative. On 21 pairs, scratch 14 (p=0.095), r16 11 (p=0.50) and r64 10 (p=0.67) are not above coin-flipping. Bootstrap-robust kept pairs are scratch 10, r16 9, r64 8, full 10 and xcal 13. The '3-4x more compressed' vs LoRA does not survive confound control. It mixes a depth artefact (scratch's obs_diff uses unthinned observed cells) with a selection artefact. Scratch's kept set includes tiny C1Q and LDG effects (1-4%) that LoRA reverses, and the medians are taken over different pair sets. On the pairs both arms keep, at the comparator's own depth, scratch is 0.107 vs r16 0.138 (1.3x) and 0.105 vs r64 0.131 (1.25x). Compression vs full FT 108.5k survives (0.066 vs 0.255, 3.8x), as it does vs xcal (6x) and own-cohort LoRA-58 (4.6x). The paediatric contrasts rest on 4 healthy-control donors and 0-2 generated control-group state cells.", + "recomputed": "Kept out of 21: zs 15, xcal 19, r16 11, r64 10, full20k 15, full108k 16, s50k 15, s100k 15, s150k 16, s162k 14, own5 17, own58 19. Median kept ratio: s162k 0.042 (0.067 at r16 depth, 0.062 at r64, 0.063 at full), r16 0.137, r64 0.131, full108k 0.150, full20k 0.290, xcal 0.229, own5 0.245, own58 0.344. Paired, pairs kept by both, scratch at comparator depth: 0.107 vs r16 0.138 (n=10); 0.105 vs r64 0.131 (n=10); 0.066 vs full108k 0.255 (n=13). Files: verify_confound_cross_contrast_counts.csv and verify_confound_cross_contrast_paired_compression.csv." + } + ] + }, + { + "id": "cross-8", + "claim": "Training the scratch model longer moves the rare monocyte states and LDG further away; only plasmablasts improve. The selected final checkpoint (s162000) is overall further from the observed frequencies than the s50000 checkpoint. This tracks the scratch model's collapsing coverage and inflating library.", + "evidence": "s50k > s100k > s150k > s162k ratios: C1Q 0.092 > 0.082 > 0.070 > 0.068; MS1 0.158 > 0.116 > 0.105 > 0.102; cycling monocyte 0.059 > 0.022 > 0.020 > 0.024; LDG 0.035 > 0.015 > 0.011 > 0.012; cycling T/NK 0.061 > 0.124 > 0.068 > 0.049; ISG-high T 0.697 > 0.567 > 0.602 > 0.599; Plasmablast label 0.486 > 0.675 > 0.794 > 0.875; proliferating T 0.444 > 0.521 > 0.564 > 0.532; HSPC 0.957 > 2.581 > 1.300 > 1.313. Mean |log2| 2.33 > 2.50 > 2.56 > 2.62. Paired mean D (s162000 - s50000) +0.288 [+0.153,+0.425]; per state, s50000 is closer in 5 (C1Q, cycling activated T, MS1, cycling monocyte, LDG) and s162000 in 3 (proliferating T, both plasmablast states). Library 13,891 > 19,138 > 20,524 > 20,606; benchmark coverage 0.063 > 0.037 > 0.034 > 0.032 (precision 0.999-1.000). At r16 depth, s50000 is overall closer than LoRA r16 (mean D -0.386 [-0.540,-0.173]), while s162000 ties r16 (-0.106 [-0.243,+0.112]).", + "how_to_recompute": "Ratios: sepsis_loco_summary.csv gen_over_obs for locoL_scratchsat_s{50000,100000,150000,162000}_K1000. Paired D: cross_bootstrap.csv kind=trajectory. Aggregate: cross_bootstrap_aggregate.csv (kind=comparison_trajectory, and the comparison_extra row for s50000 vs r16). Library: cross_depth_check.csv. Coverage: paired_crosscohort_20261002/scores/ye2026_sepsis//summary.json, per_arm..precision_coverage.", + "status": "weakened", + "votes": [ + { + "lens": "recompute", + "verdict": "weakened", + "reason": "Every ratio, mean |log2|, D, library and coverage value reproduces. The overall worsening from s50k to s162k (+0.288) is not a thinning artifact: it is unchanged with both checkpoints at r16 depth. But the headline 'only plasmablasts improve' contradicts the claim's own per-state result. Proliferating T (label) also improves significantly from s50k to s162k: 0.444 to 0.532, D -0.259 [-0.489,-0.035]. 'Tracks collapsing coverage and inflating library' is a descriptive co-trend over 4 checkpoints, not a tested link. Most of the change happens between s50k and s100k; s162k vs s100k and vs s150k are n.s. Cycling T/NK and HSPC are non-monotone, so the '>' signs in the evidence are separators, not inequalities.", + "recomputed": "Ratios s50k / s100k / s150k / s162k: C1Q 0.092 / 0.082 / 0.070 / 0.068; MS1 0.158 / 0.116 / 0.105 / 0.102; cycling monocyte 0.059 / 0.022 / 0.020 / 0.024; LDG 0.035 / 0.015 / 0.011 / 0.012; cycling T/NK 0.061 / 0.124 / 0.068 / 0.049; cycling activated T 0.079 / 0.108 / 0.073 / 0.051; ISG-high T 0.697 / 0.567 / 0.602 / 0.599; Plasmablast label 0.486 / 0.675 / 0.794 / 0.875; Plasmablast markers 0.262 / 0.307 / 0.330 / 0.338; proliferating T 0.444 / 0.521 / 0.564 / 0.532; HSPC 0.957 / 2.581 / 1.300 / 1.313. Mean |log2|: 2.327 / 2.495 / 2.560 / 2.615. Mean D: s162k-s50k +0.288 [+0.153,+0.425] (both at r16 depth +0.280 [+0.156,+0.423]); s162k-s100k +0.120 [-0.011,+0.242]; s162k-s150k +0.055 [-0.061,+0.172]. States where s50k is closer: C1Q +0.440 [+0.017,+0.891]; cycling activated T +0.622 [+0.083,+1.206]; MS1 +0.628 [+0.396,+0.950]; cycling monocyte +1.260 [+0.543,+2.129]; LDG +1.585 [+1.083,+2.159]. States where s162k is closer: proliferating T -0.259 [-0.489,-0.035]; Plasmablast label -0.846 [-1.168,-0.429]; Plasmablast markers -0.372 [-0.493,-0.233]. Library 13,891 / 19,138 / 20,524 / 20,606. Benchmark coverage 0.0626 / 0.0371 / 0.0340 / 0.0318; precision 0.9994-0.9998. At r16 depth: s50k vs r16 -0.386 [-0.540,-0.173]; s162k vs r16 -0.106 [-0.243,+0.112]." + }, + { + "lens": "confound", + "verdict": "weakened", + "reason": "The trajectory numbers reproduce. The core holds: s50k is closer overall than s162k, robust to leave-one-donor-out, leave-one-state-out, both re-weightings and detection-matched depth. s50k also beats r16 at matched depth, while s162k ties it. Two wording problems. First, 'only plasmablasts improve' contradicts the claim's own table: proliferating T (label) also improves significantly (0.444 to 0.532; D -0.26 [-0.49,-0.04], 82/82 LODO). Second, the C1Q step is fragile: D +0.44 [0.017,0.891], significant in only 57/82 leave-one-donor-out re-bootstraps and not significant at detection-matched depth. 'Tracks collapsing coverage and inflating library' is a 4-point correlation within one run. Monocyte effective marker depth also drops (detection-matched thinning 0.231 to 0.188; detected markers 6.33 to 5.94), which accounts for part of the monocyte decline. HSPC is non-monotone (0.96, 2.58, 1.30, 1.31). s162000 holds the step-152000 weights chosen by inner-validation loss (0.10115 vs 0.10493 at s50k), so checkpoint selection is blind to rare-state fidelity.", + "recomputed": "Mean |log2|: s50k 2.327, s100k 2.495, s150k 2.560, s162k 2.615. D (s162k minus s50k): +0.288 [0.153,0.425] pooled, +0.297 generated-weighted, +0.274 donor-mean, +0.242 [0.096,0.359] detection-matched; LODO 82/82. Per state at own reference: s50k closer in C1Q (57/82 LODO), cycling activated T, MS1, cycling monocyte and LDG. s162k closer in proliferating T label, plasmablast label and plasmablast markers. s50k vs r16 at r16 depth: -0.386 [-0.540,-0.173] (detection-matched -0.251 [-0.426,-0.026]). s162k vs r16: -0.106, n.s. Coverage 0.063/0.037/0.034/0.032; precision 0.999-1.000. Files: verify_confound_cross_aggregate.csv and verify_confound_cross_comparisons.csv." + } + ] + } + ] +} \ No newline at end of file