Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261013T072655Z-enriched-140-slide-master/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/build_enriched_master_20261013.py

@@ -0,0 +1,228 @@
+"""Extend the author-ordered master with source-bound Figure 1-3 material."""
+from pathlib import Path
+from datetime import datetime, timezone
+from collections import defaultdict
+import copy
+import csv
+import hashlib
+import json
+import os
+import sys
+if os.environ.get('CUDA_VISIBLE_DEVICES')!='':
+    raise SystemExit('CPU only: set CUDA_VISIBLE_DEVICES empty')
+D=Path(__file__).resolve().parents[1]
+A=D/'data/figure_enrichment_20261013'
+F=D/'figures/figure_enrichment_20261013'
+PARENT=D/'data/figure_order_fixed_20261013/spec.json'
+OUT='PBISC-Diffusion_expanded_master_20261013_enriched'
+sys.path.insert(0,str(D));sys.path.insert(0,str(D/'tools'))
+import generate_deck as G
+from reorder_master_figures_20261013 import draw_card,draw_note
+from pptx import Presentation
+from pptx.util import Inches
+
+
+def sha(path):
+    with Path(path).open('rb') as f:return hashlib.file_digest(f,'sha256').hexdigest()
+
+
+def read(name):return json.loads((A/name).read_text())
+
+
+def normalize(s,group,key=None):
+    s=copy.deepcopy(s)
+    if isinstance(s.get('sources'),list):
+        s['source_bindings']=s['sources']
+        s['sources']='; '.join(x['path'] if isinstance(x,dict) else x for x in s['sources'])
+    s.update(kind='content',deck_group=group,section=group,section_label=group,layout='full')
+    s.setdefault('keywords',[])
+    s['figures']=[str(Path(p).relative_to(D)) if Path(p).is_absolute() else p for p in s['figures']]
+    if key:s['addition_key']='enrichment:'+key
+    if 'notes' in s:s.setdefault('footnote_full',s['notes'])
+    s.setdefault('footnote_full',s.get('footnote',''))
+    return s
+
+
+def revise(s,new):
+    lineage={k:s[k] for k in ['source_page_131','parent_page','original_slide','original_title','original_footnote','addition_key'] if k in s}
+    # Old presenter notes can describe a different analysis even when the slide
+    # image is replaced. Keep historical wording in the preserved parent deck.
+    for field in ('say','text_source','numbers','refs','refs_full','keywords','layout','rows','figures_before_palette'):
+        s.pop(field,None)
+    s.update(new);s.update(lineage)
+    s['previous_figures']=copy.deepcopy(parent_by[s['parent_page']].get('figures',[]))
+    return s
+
+
+def realword(text):
+    return (text.replace('Real R2','Real-cell repeat').replace('real-R2','real-cell repeat')
+            .replace('real R2','real-cell repeat').replace('Real R1','Observed reference')
+            .replace('real R1','observed reference').replace('real-R1','observed-reference'))
+
+
+def build():
+    global parent_by
+    parent=json.loads(PARENT.read_text())['slides']
+    parent_by={i:copy.deepcopy(s) for i,s in enumerate(parent,1)}
+    by={s['source_page_131']:copy.deepcopy(s) for s in parent}
+    for i,s in enumerate(parent,1):by[s['source_page_131']]['parent_page']=i
+    arch=read('arch_result.json'); ar={r['role']:r for r in arch['slides']}
+    mismatch=read('mismatch_result.json')
+    labels=read('labels_result.json')
+    trajectory=read('trajectory_provenance.json')
+    trajectory_slide=normalize({'title':'From Gaussian noise to generated cell states',
+        'figures':[str(F/'trajectory_four_pca_frames.png')],
+        'footnote':'Recorded ctx2048 E12 EMA; same-donor pseudobulk illustration. Colour reveal shows step progress, not identity confidence.',
+        'footnote_full':trajectory['caption'],'sources':str(A/'trajectory_provenance.json')},
+        'Figure 1','trajectory_four_pca_frames')
+    real_control=normalize({'title':'Real-cell repeat: a sampling reference',
+        'figures':[labels['control_explanation']['png']],
+        'footnote':'Compare disjoint real-cell samples from the same donors. This measures sampling variability, not a theoretical ceiling.',
+        'footnote_full':'Previously called Real R2. The reference uses one 100-cell draw per donor; the control uses a disjoint 100-cell draw from the same 42 donors. Biological n remains 42.',
+        'sources':labels['control_explanation']['provenance']},'Figure 2','real_cell_repeat')
+    width=revise(by[12],normalize(ar['width_validation'],'Figure 2'))
+    revise(by[21],normalize(mismatch['slides'][0],'Figure 3'))
+    mismatch_extra=[normalize(s,'Figure 3',s['key']) for s in mismatch['slides'][1:]]
+    by[15]['previous_figures']=by[15]['figures'];by[15]['figures']=[str(Path(labels['ablation']['png']).relative_to(D))]
+    by[15]['footnote']='SLE pseudobulk; common 1,006-gene evaluation. Real-cell repeat compares disjoint observed-cell draws from the same donors.'
+    by[15]['sources']=labels['ablation']['provenance']
+    # Clarify remaining technical shorthand in supplementary figures as well.
+    replacements={41:labels['metrics_overview']['png'],31:str(F/'labels_within_type_centroids.png'),47:str(F/'labels_detection_and_library.png')}
+    for old,path in replacements.items():
+        by[old]['previous_figures']=by[old]['figures'];by[old]['figures']=[str(Path(path).relative_to(D))]
+    def add(role):return normalize(ar[role],'Figure 2',ar[role]['key'])
+    main2=[by[19],by[20],add('diagram'),add('hint_diagram'),width,by[13],add('architecture_arm_table'),
+           by[14],add('gene_validation'),add('gene_arm_table'),add('architecture_and_gene_cost'),real_control,by[15]]
+    slides=[]
+    for s in parent:
+        old=s['source_page_131']
+        if old==19:slides.extend(main2);continue
+        if old in (20,12,13,14,15):continue
+        slides.append(by[old])
+        if old==16:slides.append(trajectory_slide)
+        if old==21:slides.extend(mismatch_extra)
+    assert len(slides)==140,len(slides)
+    assert sorted(s['parent_page'] for s in slides if 'parent_page' in s)==list(range(1,132))
+    assert sorted(s['original_slide'] for s in slides if 'original_slide' in s)==list(range(1,77))
+    page_by_old={s['source_page_131']:i for i,s in enumerate(slides,1) if 'source_page_131' in s}
+    page_by_role={s['role']:i for i,s in enumerate(slides,1) if s.get('role')}
+    groups=defaultdict(list)
+    for i,s in enumerate(slides,1):groups[s['deck_group']].append(i)
+    def pg(*olds):return ', '.join(str(page_by_old[o]) for o in olds)
+    def rp(*roles):return ', '.join(str(page_by_role[r]) for r in roles)
+    def setcard(old,title,rows,foot):
+        s=by[old];s.update(title=title,rows=rows,footnote=foot,footnote_full=foot)
+    by[1]['subtitle']=f'{len(slides)} slides | Author-defined Figures 1-6 | Expanded model and expression views'
+    by[1]['title']='PBISC-Diffusion: manuscript evidence'
+    by[2]['title']='Figures 1-6: the expanded evidence map'
+    by[3].update(title='How this expanded deck is organized',rows=[
+        ('Figure 1: noise to cells','A recorded four-stage PCA view follows the experiment overview and corpus statistics.'),
+        ('Figure 2: model design and ablations','Context/trunk and hint diagrams, full validation curves, SLE benchmarks, exact parameters and observed training times.'),
+        ('Figure 3: expression mismatch to adaptation','Direct same-person expression comparisons lead into full FT, LoRA rank and paired-data efficiency.')],
+        footnote='Existing outputs are reused. Historical 131-page lineage is preserved; new views do not imply new training or inference.')
+    setcard(4,'Manuscript map: the six figures',[
+        ('Figure 1','Experimental overview, corpus and recorded denoising','Overview / illustration'),
+        ('Figure 2','Model components, validation, SLE ablations and cost','Pseudobulk development'),
+        ('Figure 3','Expression mismatch > full FT > rank > pair budget','Healthy measured pairs'),
+        ('Figure 4','Disease logFC, DEG and CAT benchmarks','Disease measured pairs'),
+        ('Figure 5','Rare-population identity in paired samples','Qualitative assembly'),
+        ('Figure 6','Published-finding reproduction from bulk-only samples','Planned')],
+        'Six author-defined scopes. Training validation and SLE biological benchmarks are labelled separately within Figure 2.')
+    traj_page=next(i for i,s in enumerate(slides,1) if s.get('addition_key')=='enrichment:trajectory_four_pca_frames')
+    setcard(7,'Figure 1: overview, corpus and denoising',[
+        ('1a','Pretraining > paired adaptation > bulk-to-cell inference','p. '+pg(9)),
+        ('1b-c','Corpus statistics and generation workflow','pp. '+pg(11,16)),
+        ('1d','Four recorded PCA stages: step 0 / 17 / 33 / 50','p. '+str(traj_page)),
+        ('Colour','Colour reveals endpoint cell types as denoising progresses','Recorded illustration')],
+        'Trajectory colours show progress, not calibrated confidence. Architecture and gene-context ablations remain in Figure 2.')
+    setcard(20,'Figure 2: architecture and gene context',[
+        ('2a-b','Context / trunk widths and predicted / oracle hint input','pp. '+rp('diagram','hint_diagram')),
+        ('2c-d','Architecture validation progression and SLE benchmark','pp. '+pg(12,13)),
+        ('2e-f','Gene-context benchmark and validation progression','pp. '+pg(14)+', '+rp('gene_validation')),
+        ('2g','All arms: parameters and observed training time','p. '+rp('architecture_and_gene_cost')),
+        ('2h-i','Real-cell repeat explained; common-axis model comparison','pp. '+str(next(i for i,s in enumerate(slides,1) if s.get('addition_key')=='enrichment:real_cell_repeat'))+', '+pg(15)),
+        ('Arm details','Exact widths, genes, endpoints, run completion and timing','pp. '+rp('architecture_arm_table','gene_arm_table'))],
+        'Validation curves use corpus recipes; biological benchmarks use SLE pseudobulk. No measured bulk-sc paired performance in Figure 2.')
+    mp=[i for i,s in enumerate(slides,1) if s.get('role','').startswith('f3_expression')]
+    setcard(53,'Figure 3: the paired-adaptation story',[
+        ('3a','Same-person expression mismatch and donor-level error','pp. '+', '.join(map(str,mp))),
+        ('3b','Effect on generated cells; raw / calibrated / full FT / LoRA','pp. '+pg(23,5)),
+        ('3c','Full FT versus LoRA r4 / r16 / r64','p. '+pg(51)),
+        ('3d','Training-pair budget: historical n + 7 validation','p. '+pg(49)),
+        ('3e-f','Genuine five-pair performance and donor-level errors','pp. '+pg(54,55))],
+        'All performance panels use healthy measured bulk-sc pairs. Input calibration is a separately labelled ten-pair diagnostic.')
+    setcard(61,'Figure 4: disease signal preservation',[
+        ('4a','Disease signal with five training donors','p. '+pg(63)),
+        ('4b','Five-fold held-out disease comparisons','p. '+pg(67)),
+        ('4c','Agreement across disease-gene list lengths','p. '+pg(68)),
+        ('4d','Cell-type expression comparison with deconvolution','p. '+pg(71)),
+        ('Scope','Sepsis, Imperial, Giroux and Ji','Measured paired data')],
+        'Keep cohort-specific exceptions visible. Detailed quality and partial corrected DEG remain in Supplementary.')
+    def extent(a,b):return f'pp. {min(groups[a])}-{max(groups[b])}'
+    setcard(109,'Supplementary evidence map',[
+        ('S01-S02','Methods and additional SLE controls',extent('Supp S01','Supp S02')),
+        ('S03-S06','Input diagnostics, evaluation, healthy quality and budgets',extent('Supp S03','Supp S06')),
+        ('S07-S09','Cohorts, disease quality, corrected DEG and diagnostics',extent('Supp S07','Supp S09')),
+        ('S10-S11','Paired rare-state quantities and external screens',extent('Supp S10','Supp S11')),
+        ('S12-S14','Transfer, scratch, readout and bulk-only cohort shift',extent('Supp S12','Supp S14')),
+        ('Archive','Historical scoring and earlier bulk-only plans',extent('Historical archive','Historical archive'))],
+        'Width-validation curves moved from supplementary into Figure 2; the prior figure remains in the preserved source deck.')
+    labels_by_role={'diagram':'Fig. 2a','hint_diagram':'Fig. 2b','width_validation':'Fig. 2c',
+        'gene_validation':'Fig. 2f','architecture_and_gene_cost':'Fig. 2g',
+        'architecture_arm_table':'Fig. 2 details','gene_arm_table':'Fig. 2 details',
+        'f3_expression_mismatch':'Fig. 3a','f3_expression_discrepancy':'Fig. 3a detail'}
+    for s in slides:
+        if s.get('role') in labels_by_role:s['section_label']=labels_by_role[s['role']]
+        if s.get('source_page_131')==13:s['section_label']='Fig. 2d'
+        if s.get('source_page_131')==14:s['section_label']='Fig. 2e'
+        if s.get('source_page_131')==15:s['section_label']='Fig. 2i'
+        if s.get('source_page_131')==23:s['section_label']='Fig. 3b detail'
+        if s.get('addition_key')=='enrichment:real_cell_repeat':s['section_label']='Fig. 2h'
+        if s.get('addition_key')=='enrichment:trajectory_four_pca_frames':s['section_label']='Fig. 1d'
+        for field in ('footnote','footnote_full','say'):
+            if s.get(field):s[field]=realword(s[field])
+        if s.get('keywords'):s['keywords']=[realword(t) for t in s['keywords']]
+        if s.get('refs') and len(G.wrap(s.get('footnote',''),G.FOOT_PT,G.FOOT_W))>1:s.pop('refs')
+        # Short on-slide titles; full original titles remain in the source records.
+        short={'Architecture arms: dimensions, completion and endpoint rule':'Architecture arms: dimensions and endpoints',
+            'All ablation arms: parameters and observed training time':'All ablation arms: parameters and training time',
+            'Same-person RNA assays give different expression values':'Bulk and pooled cells: different expression values'}
+        s['title']=short.get(s['title'],s['title'])
+    spec={'deck':{'output_pptx':OUT+'.pptx','output_pdf':OUT+'.pdf'},'slides':slides}
+    prs=Presentation();prs.slide_width=Inches(G.SLIDE_W);prs.slide_height=Inches(G.SLIDE_H)
+    renderers={'title':G.build_title,'agenda':G.build_agenda,'divider':G.build_divider,
+        'content':G.build_content,'plan_card':draw_card,'evidence_note':draw_note}
+    for n,s in enumerate(slides,1):
+        fn=renderers[s['kind']]
+        fn(prs,s) if s['kind']=='title' else fn(prs,s,n)
+    prs.save(D/(OUT+'.pptx'))
+    (A/'spec.json').write_text(json.dumps(spec,indent=2)+'\n')
+    mapping=[{'current_page':i,'parent_reordered_page':s.get('parent_page',''),
+        'original_expanded_page':s.get('source_page_131',''),'original_76_page':s.get('original_slide',''),
+        'group':s['deck_group'],'panel':s.get('section_label',''),'title':s['title'],
+        'addition_key':s.get('addition_key','')} for i,s in enumerate(slides,1)]
+    with (A/'page_map.csv').open('w',encoding='utf-8-sig',newline='') as f:
+        w=csv.DictWriter(f,fieldnames=list(mapping[0]));w.writeheader();w.writerows(mapping)
+    sources={str(p):{'sha256':sha(p)} for p in [PARENT,Path(__file__),D/'generate_deck.py',
+        D/'tools/reorder_master_figures_20261013.py',A/'arch_result.json',A/'mismatch_result.json',
+        A/'labels_result.json',A/'trajectory_provenance.json']}
+    for record in arch['sources']+mismatch['sources']+trajectory['sources']:
+        p=Path(record['path']);actual=sha(p)
+        assert actual==record['sha256'],p
+        sources[str(p)]={'sha256':actual}
+    for s in slides:
+        for f in s.get('figures',[]):sources[str(D/f)]={'sha256':sha(D/f)}
+    protected={str(D/f):sha(D/f) for s in parent for f in s.get('figures',[])}
+    for suffix in ('pptx','pdf'):
+        p=D/('PBISC-Diffusion_expanded_master_20261013_reordered.'+suffix);protected[str(p)]=sha(p)
+    manifest={'schema':'pbisc.enriched-master.v1','created_at':datetime.now(timezone.utc).isoformat(),
+        'parent_spec':str(PARENT),'parent_spec_sha256':sha(PARENT),'slides':slides,'groups':dict(groups),
+        'parent_pages':131,'total_slides':len(slides),'new_slides':len(slides)-131,
+        'sources':sources,'protected_sources':protected,'new_training':False,'new_inference':False,
+        'new_projection_fit':False,'descriptive_analysis':'Saved same-donor expression values summarized on 5,508 shared genes; no calibration refitting.',
+        'output':{'pptx':{'path':str(D/(OUT+'.pptx')),'sha256':sha(D/(OUT+'.pptx'))}}}
+    (A/'manifest.json').write_text(json.dumps(manifest,indent=2)+'\n')
+    print(json.dumps({'slides':len(slides),'groups':dict(groups),'pptx':str(D/(OUT+'.pptx'))}))
+
+
+if __name__=='__main__':build()

paper-releases/pbisc-diffusion/updates/20261013T072655Z-enriched-140-slide-master/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/enrich_architecture_20261013.py

@@ -0,0 +1,430 @@
+#!/usr/bin/env python3
+"""Source-bound Figure 2 enrichment. CPU-only; no model or array imports.
+
+Reads saved JSON histories/configs, small result CSVs, and implementation text.
+Does not train, sample, rescore, fetch remote data, or overwrite old figures.
+"""
+from __future__ import annotations
+
+import csv
+import hashlib
+import json
+import math
+from pathlib import Path
+
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from matplotlib.patches import FancyBboxPatch
+from matplotlib.ticker import MaxNLocator
+
+D = Path(__file__).resolve().parents[1]
+M = D.parents[1]
+OUT = D / "figures/figure_enrichment_20261013"
+DATA = D / "data/figure_enrichment_20261013"
+FULL = (12.3, 5.6)
+INK, GREY, GOLD, DARK, PALE = "#1A1A1A", "#6E6E6E", "#C5AA4B", "#806A25", "#F7F2DF"
+COLORS = ["#C5AA4B", "#8F772B", "#B59A50", "#66531E", "#D8C57E", "#9D8240"]
+STYLES = ["-", "--", "-.", ":", "--", "-."]
+MARKERS = ["o", "s", "^", "D", "P", "v"]
+PREFIX = "s0_6k_allK_topup16_seed20260627_"
+RUNS = [
+    ("base", "1× baseline", PREFIX + "bundle_v2", "server_1", "architecture"),
+    ("context", "Context ×4", PREFIX + "archctx2048_v1", "server_2", "architecture"),
+    ("trunk", "Trunk ×4", PREFIX + "arch4x_ctx512_v1", "server_1", "architecture"),
+    ("both", "Trunk + context ×4", PREFIX + "arch4x_v1", "server_2", "architecture"),
+    ("sixteen", "Trunk + context ×16", PREFIX + "arch16x_v1", "server_2", "architecture"),
+    ("hint", "1× + 9-fraction hint", PREFIX + "prop_v1", "server_1", "architecture"),
+    ("2k", "2k genes", "f01_2k_1x_seed20260627", "server_1", "gene_panel"),
+    ("3k", "3k genes", "f01_3k_1x_seed20260627", "server_2", "gene_panel"),
+    ("6k", "6k genes", "f01_6k_1x_seed20260627", "server_1", "gene_panel"),
+    ("15k", "15,114 genes", "f01_p15k_1x_seed20260627", "server_2", "gene_panel"),
+]
+SOURCES: dict[str, dict] = {}
+ASSETS: list[str] = []
+
+
+def sha(path: Path) -> str:
+    h = hashlib.sha256()
+    with path.open("rb") as f:
+        for b in iter(lambda: f.read(1 << 20), b""):
+            h.update(b)
+    return h.hexdigest()
+
+
+def source(path: Path, role: str, host: str = "server_1", origin: str | None = None) -> Path:
+    if not path.is_file():
+        raise FileNotFoundError(path)
+    SOURCES[str(path)] = {"path": str(path), "sha256": sha(path), "role": role,
+                          "host": "server_1", "original_host": host,
+                          "original_path": origin or str(path)}
+    return path
+
+
+def read_json(path: Path, role: str, host: str = "server_1", origin: str | None = None) -> dict:
+    return json.loads(source(path, role, host, origin).read_text())
+
+
+def write_csv(name: str, rows: list[dict]) -> None:
+    p = DATA / ("arch_" + name + ".csv")
+    fields = list(dict.fromkeys(k for row in rows for k in row))
+    with p.open("w", newline="") as f:
+        w = csv.DictWriter(f, fieldnames=fields)
+        w.writeheader(); w.writerows(rows)
+    ASSETS.append(str(p))
+
+
+def count_parameters(g: int, c: int, widths: list[int], t: int, hint: int) -> int:
+    """Exact Linear/LayerNorm parameter arithmetic from model.py; no torch/model allocation."""
+    linear = lambda a, b: a * b + b
+    n = linear(t, c) + linear(c, c)
+    n += linear(3*g+6+hint, c) + 2*c + linear(c, c)
+    n += 2*c + linear(c, c)
+    edges = [(4*g, widths[0])] + list(zip(widths[:-1], widths[1:]))
+    edges += [(widths[i+1], widths[i]) for i in reversed(range(len(widths)-1))]
+    n += sum(linear(a, b) + linear(c, b) + 2*b for a, b in edges)
+    n += linear(widths[0], widths[0]) + 2*widths[0] + 2*linear(widths[0], g)
+    return n
+
+
+def sat_epoch(hist: list[dict], tolerance: float) -> int | None:
+    if len(hist) != 16 or hist[-1]["epoch"] != 16:
+        return None
+    end = hist[-1]["validation"]["loss"]
+    return next(h["epoch"] for i, h in enumerate(hist)
+                if all(x["validation"]["loss"] <= end*(1+tolerance) for x in hist[i:]))
+
+
+def load_runs() -> tuple[list[dict], list[dict]]:
+    rows, epochs = [], []
+    for key, label, name, host, family in RUNS:
+        orig = M / "runs" / name
+        local = orig if host == "server_1" else D / "data/sec03_server2_snapshots" / name
+        cfg = read_json(local/"effective_config.json", "saved training configuration", host, str(orig/"effective_config.json"))
+        status = read_json(local/"RUN_STATUS.json", "saved run completion/failure metadata", host, str(orig/"RUN_STATUS.json"))
+        hist = read_json(local/"history.json", "complete saved epoch history", host, str(orig/"history.json"))["epochs"]
+        hist = sorted(hist, key=lambda x: x["epoch"])
+        assert [h["epoch"] for h in hist] == list(range(1, len(hist)+1))
+        g, c, w = cfg["data"]["n_genes"], cfg["model"]["context_dim"], cfg["model"]["hidden_dims"]
+        hint = 9 if key == "hint" else 0
+        params = count_parameters(g, c, w, cfg["model"]["time_embedding_dim"], hint)
+        if "parameters" in status:
+            assert params == status["parameters"], (key, params, status["parameters"])
+        else:
+            manifest = read_json(M/"plans/arch_ablation_6k_20260913/MANIFEST.json", "16x declared parameter count")
+            assert params == manifest["variants"]["16x"]["params"]
+        ts = sum(h["train_seconds"] for h in hist)
+        steps = sum(h["train"]["optimizer_steps"] for h in hist)
+        final_steps = hist[-1]["train"]["optimizer_steps"]
+        row = dict(key=key, label=label, family=family, run=name, original_host=host,
+                   original_path=str(orig), local_metadata=str(local), status=status["status"],
+                   n_genes=g, detection_floor="1%" if family == "architecture" else "0.1%",
+                   hidden_dims=" / ".join(map(str,w)), context_dim=c,
+                   condition_dim=3*g+6+hint, hint_dim=hint, parameters=params,
+                   completed_parts=len(hist), optimizer_steps=steps,
+                   train_seconds_completed_parts=ts,
+                   validation_seconds_completed_parts=sum(h["validation_seconds"] for h in hist),
+                   train_seconds_per_step_weighted=ts/steps,
+                   last_part_seconds_per_step=hist[-1]["train_seconds"]/final_steps,
+                   final_val_loss=hist[-1]["validation"]["loss"],
+                   saturation_part_0_5pct=sat_epoch(hist,.005),
+                   saturation_part_1pct=sat_epoch(hist,.01),
+                   saturation_part_2pct=sat_epoch(hist,.02),
+                   late_loss_drop_pct_per_part=((hist[11]["validation"]["loss"]-hist[15]["validation"]["loss"])/4/hist[15]["validation"]["loss"]*100) if len(hist)==16 else None,
+                   training_seed=cfg["training"]["seed"], batch_size=cfg["training"]["batch_size"],
+                   error=status.get("error", ""),
+                   runtime_note="Observed train-loop elapsed time including loading; excludes validation, checkpoint writes, restart gaps and failed partial parts. Loader/host changes prevent controlled speed claims.")
+        rows.append({**row, "history":hist, "config":cfg})
+        elapsed=0.0
+        for h in hist:
+            elapsed += h["train_seconds"]
+            epochs.append(dict(key=key,family=family,run=name,part=h["epoch"],
+                               global_step=h["global_step"],train_loss=h["train"]["loss"],
+                               validation_loss=h["validation"]["loss"],
+                               relative_to_own_E16=h["validation"]["loss"]/hist[-1]["validation"]["loss"] if len(hist)==16 else None,
+                               train_seconds=h["train_seconds"],validation_seconds=h["validation_seconds"],
+                               optimizer_steps=h["train"]["optimizer_steps"],
+                               train_seconds_per_optimizer_step=h["train_seconds"]/h["train"]["optimizer_steps"],
+                               cumulative_train_seconds=elapsed,
+                               source_path=str(local/"history.json"),original_host=host))
+    # Same architecture validation elements; different gene panels are not a common absolute-loss axis.
+    assert len({h["validation"]["support_elements"] for r in rows[:6] for h in r["history"]}) == 1
+    return rows, epochs
+
+
+def setup() -> None:
+    OUT.mkdir(parents=True, exist_ok=True); DATA.mkdir(parents=True, exist_ok=True)
+    plt.rcParams.update({"font.family":"sans-serif", "font.sans-serif":["Arial","Liberation Sans"],
+                         "font.size":14,"axes.labelsize":14,"axes.titlesize":14,
+                         "xtick.labelsize":12,"ytick.labelsize":12,"legend.fontsize":12,
+                         "axes.spines.top":False,"axes.spines.right":False,
+                         "figure.facecolor":"white","axes.facecolor":"white",
+                         "pdf.fonttype":42,"svg.fonttype":"none","text.color":INK,"legend.frameon":False,
+                         "axes.labelcolor":INK,"savefig.facecolor":"white"})
+
+
+def save(fig: plt.Figure, name: str) -> str:
+    for ext in ["png","pdf","svg"]:
+        p=OUT/("arch_"+name+"."+ext)
+        fig.savefig(p,dpi=220,bbox_inches=None)
+        ASSETS.append(str(p))
+    plt.close(fig)
+    return str((OUT/("arch_"+name+".png")).relative_to(D))
+
+
+def card(ax, x: float,y: float,w: float,h: float,title: str,body: str="", fill: str=PALE,fs: int=14) -> None:
+    ax.add_patch(FancyBboxPatch((x,y),w,h,boxstyle="round,pad=0.012,rounding_size=0.045",fc=fill,ec="#DDD5BC",lw=1))
+    ax.text(x+w/2,y+h-.16,title,ha="center",va="top",fontsize=fs,fontweight="bold",color=DARK)
+    if body: ax.text(x+w/2,y+.12,body,ha="center",va="bottom",fontsize=fs-1,linespacing=1.45)
+
+
+def arrow(ax,x0,y0,x1,y1,style="-",color=GREY) -> None:
+    ax.annotate("",(x1,y1),(x0,y0),arrowprops=dict(arrowstyle="-|>",lw=1.25,linestyle=style,color=color,shrinkA=2,shrinkB=2))
+
+
+def width_diagram(rows: list[dict]) -> str:
+    fig,ax=plt.subplots(figsize=FULL);ax.set(xlim=(0,12.3),ylim=(0,5.6));ax.axis("off")
+    card(ax,.1,4.08,3.25,1.22,"Bulk condition + measurements","G+6 bulk/depth features\n+ 2G measurement features")
+    card(ax,3.8,4.08,3.65,1.22,"Context encoder · width C","Linear(3G+6 → C), then C → C")
+    card(ax,8.05,4.08,4.05,1.22,"Time + condition → shared context","time encoding → C; add, then mix")
+    arrow(ax,3.35,4.67,3.8,4.67);arrow(ax,7.45,4.67,8.05,4.67)
+    card(ax,.1,1.58,2.35,1.88,"Gene-space input","noisy value + support: 2G\nself-conditioning: 2G\nconcatenate → 4G",fs=13)
+    card(ax,2.9,1.58,6.62,1.88,"Denoiser trunk · seven conditioned blocks","",fill="#FCFAF1")
+    widths=[1024,1024,512,256,512,1024,1024]
+    for i,v in enumerate(widths):
+        x=3.10+i*.86; height=.36+.43*v/1024
+        ax.add_patch(FancyBboxPatch((x,2.08-height/2),.62,height,boxstyle="round,pad=0.01",fc=COLORS[i%3],ec=DARK,lw=.6))
+        ax.text(x+.31,2.08,str(v),ha="center",va="center",fontsize=11)
+        arrow(ax,x+.31,2.81,x+.31,2.08+height/2,color=DARK)
+    ax.text(6.14,3.70,"same C-dimensional context → a learned bias in every block",ha="center",fontsize=12,color=DARK)
+    ax.text(6.15,1.30,"3 matching-width skip additions",ha="center",fontsize=11.5,color=GREY)
+    arrow(ax,10.07,4.08,10.07,3.50)
+    ax.plot([10.07,9.72,9.72,3.41],[3.50,3.50,2.81,2.81],color=DARK,lw=1)
+    card(ax,9.92,1.58,2.18,1.88,"Output trunk","value head: G\nsupport logits: G\n→ joint clean estimate",fs=13)
+    arrow(ax,2.45,2.5,2.9,2.5);arrow(ax,9.52,2.5,9.92,2.5)
+    ax.text(.18,.97,"Context ×4: C 512 → 2,048; trunk widths stay fixed.",fontsize=14,fontweight="bold")
+    ax.text(.18,.56,"Trunk ×4: [1,024, 1,024, 512, 256] → [4,096, 4,096, 2,048, 1,024]; C stays 512.",fontsize=13)
+    ax.text(.18,.18,"Both ×4 / ×16 change both axes. Multipliers describe widths, not parameter counts. G = 6,000 for architecture arms.",fontsize=11.5,color=GREY)
+    fig.subplots_adjust(left=0,right=1,bottom=0,top=1)
+    return save(fig,"context_vs_trunk")
+
+
+def hint_diagram(report: dict) -> str:
+    fig,ax=plt.subplots(figsize=FULL);ax.set(xlim=(0,12.3),ylim=(0,5.6));ax.axis("off")
+    card(ax,.12,3.83,3.77,1.5,"Predicted hint · bulk-only input","6k pseudobulk → log1p(CPM)\nstandardize → MLP 6,000→512→256→9\nsoftmax → nine fractions",fs=13)
+    card(ax,.12,1.90,3.77,1.5,"Oracle hint · diagnostic only","Target donor's scANVI L1 labels\ncount and normalize nine cell types\nuses observed single-cell truth",fs=13)
+    card(ax,4.4,2.66,3.60,1.39,"Replace only the 9-value hint","[base G+6 | hint 9 | measurements 2G]\nG = 6,000 → 18,015 inputs",fs=13)
+    card(ax,8.53,2.66,3.63,1.39,"Same hint-trained denoiser","condition_encoder: 18,015 → 512\n54,189,792 denoiser parameters",fs=13)
+    arrow(ax,3.89,4.54,4.4,3.50);arrow(ax,3.89,2.60,4.4,3.10);arrow(ax,8,3.35,8.53,3.35)
+    ax.text(6.2,4.76,"Training: observed sample-level fractions",ha="center",fontsize=13,fontweight="bold")
+    ax.text(6.2,4.35,"Recipe block_id → sample_key → 9-fraction table",ha="center",fontsize=12,color=GREY)
+    arrow(ax,6.2,4.15,6.2,4.05)
+    ax.text(8.21,2.10,"After encoding: add time context, mix, inject into all 7 trunk blocks.",ha="center",fontsize=12)
+    ax.text(.18,1.22,"Pred and oracle are two inference conditions, not two trained diffusion models.",fontsize=15,fontweight="bold")
+    ax.text(.18,.79,"Predictor selected by corpus-validation L1: MLP 0.075 vs ridge 0.090; 106 MLP epochs.",fontsize=13)
+    ax.text(.18,.38,f"42 SLE donors: predicted vs oracle fraction L1 = {report['benchmark42']['predicted_vs_oracle_l1_mean']:.3f}. Auxiliary predictor adds 3,206,153 parameters; its runtime is not recorded.",fontsize=12,color=GREY)
+    fig.subplots_adjust(left=0,right=1,bottom=0,top=1)
+    return save(fig,"hint_injection")
+
+
+def loss_plot(rows: list[dict], family: str) -> str:
+    selected=[r for r in rows if r["family"]==family]
+    fig,axes=plt.subplots(1,2,figsize=FULL)
+    panel=family=="gene_panel"
+    for i,r in enumerate(selected):
+        h=r["history"]; x=[z["epoch"] for z in h]
+        final=h[-1]["validation"]["loss"]
+        yy=[z["validation"]["loss"]/(final if panel else 1) for z in h]
+        lab=r["label"]+(" · stopped after E2" if len(h)<16 else "")
+        axes[0].plot(x,yy,color=COLORS[i],ls=STYLES[i],marker=MARKERS[i],ms=5,lw=1.9,label=lab,
+                     markerfacecolor="white" if i%2 else COLORS[i],markeredgewidth=1.0)
+        late_yy=yy if panel else [(z["validation"]["loss"]/final-1)*100 for z in h]
+        if len(h)==16:
+            axes[1].plot(x,late_yy,color=COLORS[i],ls=STYLES[i],marker=MARKERS[i],ms=5,lw=1.9,
+                         markerfacecolor="white" if i%2 else COLORS[i],markeredgewidth=1.0)
+        se=r["saturation_part_1pct"]
+        if se is not None:
+            axes[1].scatter([se],[late_yy[se-1]],s=115,facecolors="none",edgecolors=INK,lw=1.1,zorder=6)
+    axes[0].set_xlim(.6,16.4);axes[1].set_xlim(6.5,16.5)
+    axes[0].set_xticks([1,4,8,12,16]);axes[1].set_xticks([7,10,12,14,16])
+    axes[0].set_title("All saved parts",loc="left",fontweight="bold")
+    axes[1].set_title("Late progression · hollow ring = 1% endpoint rule",loc="left",fontsize=13)
+    for ax in axes:
+        ax.set_xlabel("Schedule part (historical epoch)")
+        ax.set_ylabel("Corpus validation loss / own E16" if panel else "Corpus validation loss")
+        ax.grid(axis="y",color="#EAEAEA",lw=.7);ax.yaxis.set_major_locator(MaxNLocator(6))
+    if panel:
+        axes[1].set_ylim(.998,1.042);axes[1].axhline(1.01,ls=":",color=GREY,lw=1)
+    else:
+        axes[1].set_ylabel("Loss above own E16 (%)")
+        axes[1].set_ylim(-.1,5.5);axes[1].axhline(1.0,ls=":",color=GREY,lw=1)
+    handles,labels=axes[0].get_legend_handles_labels()
+    fig.legend(handles,labels,loc="upper center",bbox_to_anchor=(.51,.99),ncol=3 if not panel else 4,
+               fontsize=11,handlelength=2.4,columnspacing=1.8)
+    footer=("Panels predict different genes: absolute losses are not a fair cross-panel ranking. Every curve has one training seed."
+            if panel else "Same 6k panel / validation support. 16× has two parts only: no E16 or saturation claim. Overlapping curves are genuine.")
+    fig.text(.07,.025,footer,fontsize=11,color=GREY)
+    fig.subplots_adjust(left=.075,right=.985,bottom=.17,top=.78,wspace=.29)
+    return save(fig,"panel_loss_progression" if panel else "width_loss_progression")
+
+
+def compute_plot(rows: list[dict]) -> str:
+    fig,axes=plt.subplots(1,3,figsize=FULL,gridspec_kw={"width_ratios":[1.05,1,1]})
+    ys=list(range(len(rows)))
+    labels=[r["label"]+(" †" if r["key"]=="sixteen" else "") for r in rows]
+    for ax in axes:
+        ax.set_yticks(ys);ax.invert_yaxis();ax.set_ylim(9.65,-.65)
+        ax.axhline(5.5,color="#D6D6D6",lw=1);ax.grid(axis="x",color="#EAEAEA",zorder=0)
+        ax.spines['left'].set_visible(False);ax.tick_params(axis="y",length=0)
+    values=[[r["parameters"]/1e6 for r in rows],
+            [r["train_seconds_per_step_weighted"] for r in rows],
+            [r["train_seconds_completed_parts"]/3600 for r in rows]]
+    for j,(ax,v) in enumerate(zip(axes,values)):
+        ax.barh(ys,v,color=[COLORS[i%6] for i in ys],height=.61,zorder=3,edgecolor=DARK,lw=.4)
+        for i,val in enumerate(v):
+            x=val*1.10 if j==0 else val+max(v)*.035
+            ax.text(x,i,f"{val:,.1f}" if j!=1 else f"{val:.3f}",va="center",fontsize=10.5)
+        if j==0:
+            ax.set_xscale("log");ax.set_xlim(12,7200);ax.set_yticklabels(labels,fontsize=11.5)
+        else:
+            ax.set_xlim(0,max(v)*1.33);ax.set_yticklabels([])
+    axes[0].set_xlabel("Denoiser parameters (million; log)",fontsize=12)
+    axes[1].set_xlabel("Observed train seconds / update",fontsize=12)
+    axes[2].set_xlabel("Sum of logged training hours",fontsize=12)
+    axes[0].set_title("Capacity",loc="left",fontweight="bold")
+    axes[1].set_title("Weighted across completed parts",loc="left",fontsize=12)
+    axes[2].set_title("Completed parts only",loc="left",fontsize=12)
+    fig.text(.025,.062,"† 16×: E1–E2 only; others E1–E16. Hint pred/oracle share one training run. Auxiliary MLP: +3.206M parameters, timing NA.",fontsize=11)
+    fig.text(.025,.023,"Historical observations include data loading and changing loader/host conditions; they are not controlled hardware-speed comparisons.",fontsize=11,color=GREY)
+    fig.subplots_adjust(left=.207,right=.99,bottom=.19,top=.92,wspace=.28)
+    return save(fig,"all_arm_compute")
+
+
+def audit_table(rows: list[dict], family: str) -> str:
+    selected=[r for r in rows if r["family"]==family]
+    fig,ax=plt.subplots(figsize=FULL);ax.axis("off")
+    col=["Arm / inference use","G / condition","Trunk down widths","Context","Params","Parts","1% / 0.5% / 2%"]
+    body=[]
+    for r in selected:
+        dims=r['hidden_dims'].replace(' / ',' /\n',1) if r['key']=='sixteen' else r['hidden_dims']
+        sats=" / ".join("NA" if r[f"saturation_part_{k}"] is None else str(r[f"saturation_part_{k}"]) for k in ["1pct","0_5pct","2pct"])
+        body.append([r["label"]+("\npred + oracle: same weights" if r['key']=='hint' else ""),
+                     f"{r['n_genes']:,} /\n{r['condition_dim']:,}",dims,str(r['context_dim']),
+                     f"{r['parameters']/1e6:.3f}M",str(r['completed_parts']),sats])
+    table=ax.table(cellText=body,colLabels=col,cellLoc="left",colLoc="left",loc="upper left",
+                   bbox=[.005,.26,.99,.64],colWidths=[.205,.105,.26,.075,.10,.06,.195])
+    table.auto_set_font_size(False);table.set_fontsize(11.3)
+    for (i,j),cell in table.get_celld().items():
+        cell.set_edgecolor("#E4DFD2");cell.set_linewidth(.6);cell.PAD=.055
+        if i==0: cell.set_facecolor(PALE);cell.get_text().set_weight("bold");cell.get_text().set_color(DARK)
+        else: cell.set_facecolor("white" if i%2 else "#FCFBF6")
+    ax.text(.008,.955,"Architecture family · 1% detection-floor genes" if family=='architecture' else "Gene-panel family · 0.1% detection-floor genes",fontsize=15,fontweight="bold",transform=ax.transAxes)
+    ax.text(.008,.19,"Last column: first part after which every remaining validation loss is within the stated tolerance of that run's E16 loss.",fontsize=11.6,transform=ax.transAxes)
+    ax.text(.008,.125,"This is a retrospective endpoint rule, not proof of convergence; sensitivity to 0.5% / 1% / 2% is retained.",fontsize=11.6,transform=ax.transAxes)
+    if family=='architecture':
+        ax.text(.008,.068,"Pred also needs a separately trained 3.206M-parameter MLP; predictor training/inference timing is NA and not included.",fontsize=11.4,color=GREY,transform=ax.transAxes)
+    note=("16×: FAILED_PRESERVED after E2; memory floor during resumed part 3. 2× was retired before the reported suite; it has no curve here."
+          if family=='architecture' else "Each panel is one trained checkpoint. Five 6k sampling seeds reuse one checkpoint; 2k/3k/15k each have one sampling seed.")
+    ax.text(.008,.016 if family=='architecture' else .043,note,fontsize=11.3,color=GREY,transform=ax.transAxes)
+    fig.subplots_adjust(left=.025,right=.98,bottom=.02,top=.99)
+    return save(fig,"width_arm_register" if family=='architecture' else "panel_arm_register")
+
+
+def build() -> dict:
+    SOURCES.clear(); ASSETS.clear(); setup()
+    paths=[M/"src/pbisc_sdd/model.py",M/"src/pbisc_sdd/experiment.py",M/"src/pbisc_sdd/data.py",
+           M/"src/pbisc_sdd/ladder_experiment.py",M/"src/pbisc_sdd/proportion_hint.py",
+           M/"plans/prop_condition_6k_20260916/train_proportion_predictor.py",
+           M/"plans/prop_condition_6k_20260916/README.md",M/"plans/arch_ablation_6k_20260913/README.md",
+           D/"tools/sec03_common.py",D/"tools/sec03_panel.py",D/"tools/sec03_width_bench.py",
+           M/"scripts/sample_sdd_standalone.py"]
+    for p in paths: source(p,"implementation and historical protocol semantics")
+    assert "time_context + condition_context" in paths[0].read_text()
+    assert "base_condition, hint_values" in paths[1].read_text()
+    assert "condition, bulk_analog, output_analog" in paths[2].read_text()
+    report=read_json(M/"plans/prop_condition_6k_20260916/predictor_v1/predictor_report.json","saved separate hint predictor metrics")
+    assert report['winner_by_val']=='mlp' and len(report['labels'])==9
+    assert 6000*512+512+512*256+256+256*9+9 == 3206153
+    rows,epochs=load_runs()
+    public=[{k:v for k,v in r.items() if k not in ['history','config']} for r in rows]
+    write_csv("training_arms",public);write_csv("loss_and_timing_by_part",epochs)
+    variants=[]
+    for r in public:
+        for mode in (["pred","oracle"] if r['key']=='hint' else ["standard"]):
+            variants.append({**r,"inference_variant":mode,"independent_training_key":r['key'],
+                             "external_predictor_parameters":3206153 if mode=='pred' else 0,
+                             "external_predictor_training_seconds":None,
+                             "oracle_target_truth_access":mode=='oracle'})
+    write_csv("inference_variant_register",variants)
+    # Verify saved figure tables without regenerating scores.
+    for name in ['sec03_width_loss.csv','sec03_panel_loss.csv','sec03_width_bench.csv','sec03_panel_bench.csv','sec03_panel_seeds.csv']:
+        source(D/'data'/name,"existing figure source table; no rescoring")
+    aliases={'base':'1x','context':'ctx4x','trunk':'trunk4x','both':'4x','sixteen':'16x','hint':'prop'}
+    with (D/'data/sec03_width_loss.csv').open() as f: old=list(csv.DictReader(f))
+    for r in rows[:6]:
+        saved={int(float(z['epoch'])):z for z in old if z['variant']==aliases[r['key']]}
+        for h in r['history']:
+            assert math.isclose(float(saved[h['epoch']]['val_loss']),h['validation']['loss'],rel_tol=1e-13)
+    common=("Source-only enrichment: saved configs/history + implementation; no new model training, inference or statistical scoring. "
+            "Validation curves describe internal pretraining-corpus recipes (192 × 64), not SLE held-out donor scores. "
+            "Figure 2 benchmark scores remain SLE pseudobulk, not measured bulk–sc paired performance. "
+            "Architecture and gene-panel families have different detection-floor gene selections and must be compared within family. ")
+    timing=("History train_seconds times train_one_epoch including data loading and heartbeat overhead. Divide its sum by recorded optimizer_steps for weighted seconds/update. "
+            "Sum only completed parts; validation, checkpoint writes, resume gaps and failed partial parts excluded. RUN_STATUS.wall_seconds resets on resumed invocations and is not a total. "
+            "Hosts, loader settings and competing load changed; do not interpret timing ratios as a controlled benchmark. Server_2 histories are existing local snapshots, not a live-status claim.")
+    specs=[
+        ("context_vs_trunk","Where context width and trunk width change",width_diagram(rows),11,
+         "Code-derived diagram. Width multipliers are not parameter multipliers; all architecture arms use 6,000 genes.",
+         "Seven conditioned blocks: down widths [1024,1024,512,256], up widths [512,1024,1024], three additive skips. Context = mixer(time encoder + condition encoder); each block adds its own projected context. No latent-space diffusion. Visual organization follows the public website's condition strip / input / denoiser / output cards; fresh vector implementation, not copied recordings."),
+        ("hint_injection","Predicted and oracle hints: one trained denoiser",hint_diagram(report),11,
+         "Pred/oracle share hint-trained weights. Oracle reads target scANVI fractions; the separate predictor's training time is unavailable.",
+         "Training sample fractions are joined by recipe block_id→sample_key. Pred is corpus-trained MLP (train5232/val648/test627 samples), chosen by validation L1 against ridge; log1p(CPM), train-standardized, two hidden GELU/dropout layers and softmax. Oracle aggregates target donor observed AIFI-L1 labels over nine categories. Hint is appended after base G+6, before 2G measurement context; first condition layer 18015→512. Baseline denoiser and hint denoiser are separately trained; pred and oracle share hint denoiser. Predictor params3206153 calculated from recorded architecture, excluded from denoiser count."),
+        ("width_loss_progression","Architecture: full and late validation progression",loss_plot(rows,"architecture"),12,
+         "Corpus validation; same 6k genes. Hollow ring: within 1% of own E16 thereafter. 16× stops at E2; saturation remains NA.",
+         "Full saved E1–E16 curves with a late-region view. No smoothing, curve fitting or extrapolation. One training seed. Endpoint proximity is a retrospective summary, not established convergence; tolerance sensitivity is in arm register."),
+        ("panel_loss_progression","Gene panels: progression relative to each endpoint",loss_plot(rows,"gene_panel"),13,
+         "Corpus validation, normalized to each panel's E16. Different gene sets prevent comparing absolute losses across panels.",
+         "All four panels have 16 saved parts. Curves use each panel's own denominator; this compares progression shape, not identical-objective performance. One training seed for every panel; five 6k generation seeds are not five training repetitions."),
+        ("all_arm_compute","All ablation arms: parameters and observed training time",compute_plot(rows),13,
+         "Observed train-loop time includes loading; excludes validation/restarts/checkpoint writes. 16× has E1–E2 only; no controlled speed claim.",timing),
+        ("width_arm_register","Architecture arms: dimensions, completion and endpoint rule",audit_table(rows,"architecture"),12,
+         "Six independent training runs; pred/oracle reuse the hint run. 16× incomplete; retired 2× is not a completed result.",
+         "Counts derived arithmetically from exact Linear/LayerNorm definitions and checked against all available RUN_STATUS parameters; 16× checked against recorded manifest. Saturation columns use 1%,0.5%,2% relative to own E16. 16× memory-floor failure preserved; discarded2× design not counted as trained evidence."),
+        ("panel_arm_register","Gene-panel arms: exact sizes and training provenance",audit_table(rows,"gene_panel"),13,
+         "Four separately trained panels; 6k sampling seeds reuse one checkpoint. Saturation is endpoint-relative, not proof of convergence.",
+         "Nominal15k is15114 genes. Fixed trunk/context widths; parameters change because input/output and condition projections depend on G. Architecture6k 1% and final6k0.1% are distinct panels/checkpoints despite equal parameter counts.")]
+    roles={"context_vs_trunk":"diagram", "hint_injection":"hint_diagram", "width_loss_progression":"width_validation",
+           "panel_loss_progression":"gene_validation", "all_arm_compute":"architecture_and_gene_cost",
+           "width_arm_register":"architecture_arm_table", "panel_arm_register":"gene_arm_table"}
+    source(Path(__file__).resolve(),"reproducible CPU-only enrichment renderer")
+    slides=[]
+    for key,title,png,after,foot,full in specs:
+        assert len(foot)<=180,(key,len(foot))
+        slides.append(dict(key="arch_"+key,kind="content",type="content",title=title,
+                           section="Figure 2",section_label="Figure 2 support",figures=[png],keywords=[],
+                           footnote=foot,footnote_full=common+full,
+                           notes=common+full,after_current=after,after_current_page=after,
+                           palette="native",role=roles[key],sources=list(SOURCES.values()),
+                           replace_current_page=43 if key=='width_loss_progression' else None,
+                           replacement_previous_131_page=12 if key=='width_loss_progression' else None))
+    payload=dict(schema="pbisc.figure-enrichment.architecture.v1",slides=slides,assets=ASSETS.copy(),
+                 sources=list(SOURCES.values()),training_runs=10,inference_variants=11,
+                 scope="Figure 2: all architecture/hint and final gene-panel arms in sec03; excludes retired2x and unrelated pilot/loss-policy studies.",
+                 style_reference="https://www.virtualhumanlab.com/research/pbisc-diffusion/; code-native redraw; no website data copied",
+                 timing_definition=timing,
+                 limitations=["No training seed replicates; five6k seeds are generation repeats.","16x onlyE1–E2; saturationNA; failedpartial elapsed excluded.","Separate hint predictor runtimeNA.","SLEscores and corpusloss are distinct evidence.","Historical loader/host changes preclude controlledspeedclaims."],
+                 source_validation={"width_csv_losses_match_history":True,"all_recorded_parameter_counts_match_arithmetic":True})
+    hashes={"schema":"pbisc.architecture-enrichment.hashes.v1", "sources":list(SOURCES.values()),
+            "assets":[{"path":p,"sha256":sha(Path(p))} for p in ASSETS],
+            "self_and_result_json_excluded_to_avoid_recursive_hashes":True}
+    p=DATA/"arch_hashes.json";p.write_text(json.dumps(hashes,indent=2)+"\n")
+    payload["assets"].append(str(p))
+    payload["assets"].extend([str(DATA/"arch_additions.json"),str(DATA/"arch_result.json")])
+    p=DATA/"arch_additions.json";p.write_text(json.dumps(payload,indent=2)+"\n")
+    p=DATA/"arch_result.json";p.write_text(json.dumps(payload,indent=2)+"\n")
+    return payload
+
+
+if __name__=="__main__":
+    result=build()
+    print(json.dumps({"slides":len(result['slides']),"assets":len(result['assets']),"sources":len(result['sources'])}))

paper-releases/pbisc-diffusion/updates/20261013T072655Z-enriched-140-slide-master/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/enrich_expression_mismatch_20261013.py

@@ -0,0 +1,264 @@
+"""Direct paired ImMicro expression mismatch, using the verified 5508-gene axis.
+
+Reads saved assay-level matrices and saved calibration conditions only. It never
+loads generated expression, fits calibration/PCA, initializes CUDA, or trains a
+model. The other-cohort compact reduction supplies only the immutable gene IDs
+of the identical final-model checkpoint, never any expression values.
+"""
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+from pathlib import Path
+
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+
+import numpy as np
+import pandas as pd
+import style as ST
+import matplotlib.pyplot as plt
+from matplotlib.colors import LogNorm
+
+D = Path(__file__).resolve().parents[1]
+M = D.parents[1]
+U = M / "plans/utility_studies_final6k_20260925"
+S = U / "umap/sample_level"
+C = M / "plans/realbulk_generation_20260922/conditions"
+FEATURES = Path("/home/sj_server_1/gblee/datalake/paired-pbmc-bulk-sc/analyses/offset_sample_size_20260914/features.npz")
+PANEL = S / "reduced/reduced_Arunachalam_paired12.npz"
+FIG = D / "figures/figure_enrichment_20261013"
+DATA = D / "data/figure_enrichment_20261013"
+RAW_COLOR, CAL_COLOR = "#777777", "#C5AA4B"
+
+
+def sha(path: Path) -> str:
+    h = hashlib.sha256()
+    with path.open("rb") as f:
+        for block in iter(lambda: f.read(1 << 20), b""):
+            h.update(block)
+    return h.hexdigest()
+
+
+def log_cpm(values: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+    totals = values.sum(axis=1)
+    if not np.isfinite(values).all() or (values < 0).any() or (totals <= 0).any():
+        raise ValueError("Common-axis expression contains missing/negative values or an empty library")
+    return np.log1p(values / totals[:, None] * 1e6), totals
+
+
+def build() -> dict:
+    sources: dict[str, dict] = {}
+    assets: list[str] = []
+    def source(path: Path, role: str) -> Path:
+        path = path.resolve()
+        sources[str(path)] = {"path": str(path), "sha256": sha(path), "bytes": path.stat().st_size, "role": role}
+        return path
+    def document(path: Path, role: str) -> dict:
+        return json.loads(source(path, role).read_text())
+    def save_table(frame: pd.DataFrame, name: str) -> Path:
+        path = DATA / f"mismatch_{name}.csv"
+        frame.to_csv(path, index=False); assets.append(str(path))
+        return path
+    def save(fig, name: str) -> Path:
+        for ext in ("png", "pdf", "svg"):
+            path = FIG / f"mismatch_{name}.{ext}"
+            fig.savefig(path, dpi=ST.DPI, bbox_inches="tight", pad_inches=0.07, facecolor="white")
+            assets.append(str(path))
+        plt.close(fig)
+        return FIG / f"mismatch_{name}.png"
+
+    FIG.mkdir(parents=True, exist_ok=True); DATA.mkdir(parents=True, exist_ok=True)
+    condition_manifest = document(C / "MANIFEST_ImMicro.json", "saved group-held-out ten-pair calibration recipe and condition hashes")
+    source(FEATURES, "actual paired ImMicro bulk counts and observed single-cell summed UMI, gene availability and donor IDs")
+    if sources[str(FEATURES.resolve())]["sha256"] != condition_manifest["source_sha256"]:
+        raise ValueError("Assay matrix differs from the saved ImMicro condition source")
+    panel_meta = document(PANEL.with_suffix(".json"), "receipt of compact final-model reduction used only to recover panel IDs")
+    immicro_meta = document(S / "reduced/reduced_ImMicro_native.json", "matching final-model ImMicro checkpoint lineage")
+    source(PANEL, "only panel_ids array read; no other-cohort expression is used")
+    if sha(PANEL) != panel_meta["out_sha256"]:
+        raise ValueError("Gene-axis carrier fails its saved reduction hash")
+    checkpoints = {v["checkpoint_sha256"] for meta in (panel_meta, immicro_meta) for v in meta["files"].values()}
+    if len(checkpoints) != 1:
+        raise ValueError("Panel carrier and ImMicro were generated from different checkpoints")
+    with np.load(PANEL, allow_pickle=False) as z:
+        panel_ids = z["panel_ids"].astype(str)
+    with np.load(FEATURES, allow_pickle=False) as z:
+        gene_ids = z["gene_ids"].astype(str)
+        donors = z["donor_ids"].astype(str)
+        pools = z["pools"].astype(str)
+        bvalid = z["bulk_valid"].astype(bool).all(axis=0)
+        svalid = z["sc_valid"].astype(bool).all(axis=0)
+        shared = bvalid & svalid
+        if not np.array_equal(shared, z["common_all_donors"].astype(bool)):
+            raise ValueError("Saved shared measurement support disagrees with actual assay masks")
+        mapping = {gene: i for i, gene in enumerate(gene_ids)}
+        positions = np.array([mapping[g] for g in panel_ids], dtype=int)
+        kept_panel_positions = np.flatnonzero(shared[positions])
+        cols = positions[kept_panel_positions]
+        if len(donors) != 107 or len(set(donors)) != 107 or len(panel_ids) != 6000 or len(cols) != 5508:
+            raise ValueError("Expected 107 unique donors and exactly 5508 shared genes of the final 6000-gene model panel")
+        bulk = np.asarray(z["bulk_raw"][:, cols], dtype=np.float64)
+        pseudo = np.asarray(z["pseudo_raw_sum"][:, cols], dtype=np.float64)
+    raw, raw_total = log_cpm(bulk)
+    observed, pseudo_total = log_cpm(pseudo)
+
+    # Use the stored calibrated inputs, not a newly estimated multiplier.
+    cal = None
+    for arm in ("raw", "cal", "pseudo"):
+        path = source(C / f"conditions_ImMicro_{arm}.npz", f"saved {arm} model-condition profile; only bulk_counts and identity metadata read")
+        if sha(path) != condition_manifest["arms"][arm]["sha256"]:
+            raise ValueError(f"Saved {arm} condition no longer matches its receipt")
+        with np.load(path, allow_pickle=False) as z:
+            if not np.array_equal(z["gene_ids"].astype(str), gene_ids) or not np.array_equal(z["sample_ids"].astype(str), donors):
+                raise ValueError("Condition donors/gene order do not match actual paired assays")
+            expression, _ = log_cpm(np.asarray(z["bulk_counts"][:, cols], dtype=float))
+        if arm == "cal": cal = expression
+        else:
+            expected = raw if arm == "raw" else observed
+            if not np.allclose(expression, expected, rtol=0, atol=1e-10):
+                raise ValueError(f"Common-axis {arm} condition does not reproduce assay expression")
+    if cal is None: raise ValueError("No saved paired-calibration values")
+    for pool, info in condition_manifest["calibration_per_group"].items():
+        ids = np.asarray(info["calibration_rows"], dtype=int)
+        if len(ids) != 10 or (pools[ids] == pool).any() or donors[ids].tolist() != info["calibration_samples"]:
+            raise ValueError("Calibration includes evaluated pool/donor or differs from saved recipe")
+
+    cached = document(S / "ImMicro/metrics_ImMicro.json", "old common-axis normalization and expression-space reference scale")
+    if cached["n_genes"] != len(cols) or cached["n_samples"] != len(donors):
+        raise ValueError("Old expression diagnostic uses a different donor/gene set")
+    old_distance_path = source(S / "ImMicro/per_sample_distance_ImMicro_native.csv", "existing full-expression distances used solely as a reproduction check")
+    old_distance = pd.read_csv(old_distance_path)
+    unit = cached["settings"]["native"]["sum"]["full"]["unit"]["D"]
+    max_distance_gap = 0.0
+    for label, matrix in (("raw", raw), ("cal", cal)):
+        old = old_distance[(old_distance.space == "full") & (old_distance.group == label)].set_index("donor_id")
+        if len(old) != 107: raise ValueError("Old full-gene distance coverage incomplete")
+        actual = np.linalg.norm(matrix - observed, axis=1) / unit
+        gap = float(np.max(np.abs(actual - old.loc[donors, "distance_to_own_observed_D"].to_numpy())))
+        max_distance_gap = max(max_distance_gap, gap)
+        if gap > 1e-5: raise ValueError(f"Recovered gene axis/normalization differs from saved {label} expression analysis: {gap}")
+
+    gene_table = pd.DataFrame({"gene_id": panel_ids[kept_panel_positions], "canonical_position": cols,
+                               "final_panel_position": kept_panel_positions, "measured_in_all_107_both_assays": True})
+    gene_path = save_table(gene_table, "gene_axis")
+    raw_error = np.abs(raw - observed).mean(axis=1)
+    cal_error = np.abs(cal - observed).mean(axis=1)
+    per_donor = pd.DataFrame({"donor_id": donors, "pool": pools, "genes": len(cols),
+        "bulk_raw_common_gene_total": raw_total, "sc_sum_common_gene_total": pseudo_total,
+        "raw_mae_log1p_cpm": raw_error, "saved_calibration_mae_log1p_cpm": cal_error,
+        "raw_minus_calibration_mae": raw_error - cal_error,
+        "raw_rmse_log1p_cpm": np.sqrt(((raw-observed)**2).mean(axis=1)),
+        "saved_calibration_rmse_log1p_cpm": np.sqrt(((cal-observed)**2).mean(axis=1))})
+    donor_path = save_table(per_donor, "per_donor_errors")
+    # Every donor × shared-gene value contributes; no marker or donor selection.
+    high = np.ceil(max(raw.max(), observed.max(), cal.max()) * 2) / 2
+    edges = np.linspace(0, high, 121)
+    matrices, density_rows = [], []
+    for arm, values in (("raw", raw), ("saved_calibration", cal)):
+        counts = np.zeros((120, 120), dtype=np.int64)
+        for start in range(0, len(donors), 8):
+            block = np.histogram2d(observed[start:start+8].ravel(), values[start:start+8].ravel(), bins=(edges, edges))[0]
+            counts += block.astype(np.int64)
+        if int(counts.sum()) != len(donors) * len(cols): raise ValueError("Density plot dropped paired gene values")
+        matrices.append(counts)
+        for i in range(120):
+            for j in range(120):
+                density_rows.append({"arm": arm, "sc_log1p_cpm_left": edges[i], "sc_log1p_cpm_right": edges[i+1],
+                    "bulk_log1p_cpm_left": edges[j], "bulk_log1p_cpm_right": edges[j+1], "donor_gene_pairs": int(counts[i,j])})
+    bins_path = save_table(pd.DataFrame(density_rows), "density_bins")
+    display_style = {"font.family": "sans-serif", "font.sans-serif": ["Arial", "Liberation Sans", "DejaVu Sans"],
+        "font.size": 14, "axes.labelsize": 14, "axes.titlesize": 14, "xtick.labelsize": 12, "ytick.labelsize": 12,
+        "legend.fontsize": 12, "legend.frameon": False, "axes.spines.top": False, "axes.spines.right": False,
+        "figure.facecolor": "white", "axes.facecolor": "white", "pdf.fonttype": 42, "svg.fonttype": "none"}
+    with plt.rc_context(display_style):
+        fig, axes = plt.subplots(1, 2, figsize=ST.FULL, sharex=True, sharey=True)
+        fig.subplots_adjust(left=0.08, right=0.88, bottom=0.18, top=0.86, wspace=0.26)
+        vmax = max(int(x.max()) for x in matrices)
+        for ax, counts, title in zip(axes, matrices, ["Measured bulk", "Saved paired calibration · diagnostic"]):
+            im = ax.imshow(np.ma.masked_equal(counts.T, 0), origin="lower", extent=(0, high, 0, high),
+                           aspect="equal", cmap="Blues", norm=LogNorm(vmin=1, vmax=vmax), interpolation="nearest")
+            ax.plot([0, high], [0, high], "--", color="#252525", linewidth=1)
+            ax.set_title(title, loc="left", pad=10)
+            ax.set_xlabel("Same-donor scRNA-seq sum\n(pseudobulk; log1p CPM)")
+            ax.set_ylabel("Bulk expression (log1p CPM)")
+            ax.set_xlim(0, high); ax.set_ylim(0, high)
+        cax = fig.add_axes([0.91, 0.23, 0.018, 0.56])
+        cb = fig.colorbar(im, cax=cax); cb.set_label("Donor–gene pairs per bin", fontsize=12)
+        fig.text(0.08, 0.98, f"107 paired healthy donors × {len(cols):,} shared final-model genes", fontsize=15, ha="left", va="top")
+        fig.text(0.08, 0.025, "Dashed line: equal expression · all measured zeros retained · same gene axis and normalization", fontsize=11, color=ST.SUBINK)
+        expression_figure = save(fig, "expression_values")
+
+        fig, axes = plt.subplots(1, 2, figsize=ST.FULL)
+        fig.subplots_adjust(left=0.085, right=0.98, bottom=0.21, top=0.86, wspace=0.32)
+        jitter = np.linspace(-0.10, 0.10, len(donors))
+        # Fixed source donor order; all 107 are shown, no donor ranking/selection.
+        for i in range(len(donors)):
+            axes[0].plot([jitter[i], 1+jitter[i]], [raw_error[i], cal_error[i]], color="#B9B9B9", linewidth=0.6, alpha=0.45)
+        axes[0].scatter(jitter, raw_error, s=19, color=RAW_COLOR, zorder=3)
+        axes[0].scatter(1+jitter, cal_error, s=19, color=CAL_COLOR, zorder=3)
+        for x, values in ((0, raw_error), (1, cal_error)):
+            median = float(np.median(values))
+            axes[0].plot([x-0.15, x+0.15], [median, median], color=ST.INK, linewidth=2.0, zorder=4)
+            axes[0].annotate(f"Median {median:.3f}", (x, median), xytext=(0, 15), textcoords="offset points", ha="center", fontsize=12,
+                             bbox={"facecolor": "white", "edgecolor": "none", "pad": 1.5})
+        axes[0].set_xticks([0, 1], ["Measured bulk", "Paired-calibrated bulk\n(diagnostic)"])
+        axes[0].set_xlim(-0.3, 1.3); axes[0].set_ylim(bottom=0)
+        axes[0].set_ylabel("Same-donor mean absolute error\n(log1p-CPM units)")
+        axes[0].set_title("Each line = one donor", loc="left")
+        for values, label, color in ((raw_error, "Measured bulk", RAW_COLOR), (cal_error, "Saved paired calibration", CAL_COLOR)):
+            xs = np.sort(values); ys = 100 * np.arange(1, len(xs)+1) / len(xs)
+            axes[1].step(np.r_[0, xs], np.r_[0, ys], where="post", label=label, color=color, linewidth=2)
+        axes[1].set_xlabel("Same-donor mean absolute error\n(log1p-CPM units)")
+        axes[1].set_ylabel("Donors at or below this error (%)")
+        axes[1].set_ylim(0, 102); axes[1].set_xlim(left=0); axes[1].legend(loc="lower right")
+        axes[1].set_title("All 107 donors · lower error is better", loc="left")
+        for ax in axes: ax.grid(axis="y", color="#EAEAEA", linewidth=0.7); ax.set_axisbelow(True)
+        fig.text(0.085, 0.035, "Calibration used 10 pairs from other sc pools; the evaluated donor and its entire pool were excluded.", fontsize=11, color=ST.SUBINK)
+        errors_figure = save(fig, "same_donor_errors")
+
+    source(Path(__file__), "figure builder; no model generation or calibration fit")
+    source(D / "tools/style.py", "original 76-slide figure typography")
+    for path in (gene_path, donor_path, bins_path): source(path, "exact display source table")
+    notes = (f"ImMicro, 107 unique paired healthy donors. All {len(cols)} final-model panel genes measured by both assays in every donor are included; "
+        "the other 492 model genes are excluded by availability, never filled as measured zeros. Each assay row is separately normalized to 1e6 "
+        "over the same 5508 genes and transformed by natural log(1+CPM). Axes are gene-expression profiles, not cell-type counts, learned embeddings or PCA. "
+        "Observed pseudobulk is the sum of the donor's measured single-cell UMI counts. The saved calibrated condition uses the historical gene-wise "
+        "multiplier from exactly 10 other-pool paired donors, with the evaluated donor's whole sc pool excluded. It is a paired-information diagnostic, "
+        "not zero-shot deployment, LoRA output or the current strict-five budget. Source condition hashes and raw/pseudobulk normalization are verified. "
+        "The full-expression distances reproduce the old saved analysis within its rounded CSV precision. Assay differences can include cell capture, "
+        "pool/batch, depth and assay chemistry; this plot does not attribute them to one mechanism or treat scRNA-seq as noise-free truth. "
+        "No individual genes or donors were chosen for a favorable example. The plots describe 107 biological donors; donor–gene points are not independent replicates.")
+    footer1 = "107 paired donors; all 5,508 shared final-model genes. log1p CPM on the same gene axis. Saved calibration uses 10 other-pool pairs."
+    footer2 = "Each line = one donor. Error averages all 5,508 genes; no nearest-neighbor score. Ten-pair, other-pool calibration is diagnostic."
+    slides = []
+    for key, title, figure, footer, role in [
+        ("mismatch_expression_values", "Same-person RNA assays give different expression values", expression_figure, footer1, "f3_expression_mismatch"),
+        ("mismatch_same_donor_errors", "Paired calibration reduces same-donor expression error", errors_figure, footer2, "f3_expression_discrepancy")]:
+        slides.append({"kind": "content", "layout": "full", "key": key, "role": role, "title": title,
+            "section_label": "Fig. 3 · Input expression", "figures": [str(figure.relative_to(D))], "keywords": [],
+            "footnote": footer, "footnote_full": notes, "notes": notes,
+            "sources": "; ".join(sources), "palette": "biological", "after_original": 17})
+    summary = {"n_donors": 107, "n_shared_genes": len(cols), "donor_gene_pairs_per_density_panel": 107*len(cols),
+        "raw_median_mae_log1p_cpm": float(np.median(raw_error)), "calibrated_median_mae_log1p_cpm": float(np.median(cal_error)),
+        "raw_mean_mae_log1p_cpm": float(raw_error.mean()), "calibrated_mean_mae_log1p_cpm": float(cal_error.mean()),
+        "donors_with_lower_error_after_saved_calibration": int((cal_error < raw_error).sum()),
+        "reproduction_max_abs_gap_in_old_rounded_distance_D": max_distance_gap,
+        "checkpoint_sha256": next(iter(checkpoints)), "panel_carrier_expression_used": False,
+        "mask": "final 6000-gene panel intersect all-donor bulk_valid and sc_valid; 5508 genes",
+        "normalization": "ln(1 + 1e6 * count_g / sum_counts_over_same_5508_genes)",
+        "new_calibration_fit": False, "new_generation": False, "new_training": False,
+        "new_work": "paired expression binning and donor-wise descriptive errors from existing assays/calibrated conditions"}
+    analysis_path = DATA / "mismatch_analysis.json"
+    analysis_path.write_text(json.dumps(summary, indent=2) + "\n"); assets.append(str(analysis_path)); source(analysis_path, "descriptive expression comparison and validation")
+    result = {"schema": "pbisc.figure-enrichment.expression-mismatch.v1", "slides": slides,
+              "sources": list(sources.values()), "assets": assets, "summary": summary,
+              "recommended_footer": footer1, "no_new_generation": True}
+    (DATA / "mismatch_result.json").write_text(json.dumps(result, indent=2) + "\n")
+    return result
+
+
+if __name__ == "__main__":
+    result = build()
+    print(json.dumps({"slides": len(result["slides"]), "summary": result["summary"],
+                      "result": str(DATA / "mismatch_result.json")}, indent=2))