Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20260929T062828Z-manuscript-relocation/diffusion/dataset/server_1/manuscript/scripts/cohort_calibration/package_analysis.py

@@ -0,0 +1,87 @@
+"""Bundle verified calibration figures, source tables and code into the paper."""
+from pathlib import Path
+from datetime import datetime, timezone
+import hashlib
+import json
+import shutil
+import pandas as pd
+
+ROOT=Path(__file__).resolve().parent
+PAPER=ROOT.parents[1]
+
+
+def sha(path):
+    h=hashlib.sha256()
+    with path.open('rb') as f:
+        for x in iter(lambda:f.read(8*1024**2),b''):h.update(x)
+    return h.hexdigest()
+
+
+def main():
+    assert json.loads((ROOT/'QA.json').read_text())['state']=='passed'
+    assert json.loads((ROOT/'PUBLICATION_QA.json').read_text())['state']=='PASS'
+    destination=PAPER/'source_data/calibration'
+    scripts=PAPER/'scripts/cohort_calibration'
+    destination.mkdir(parents=True,exist_ok=True);scripts.mkdir(parents=True,exist_ok=True)
+    copied=[]
+    def copy(source,target):
+        target.parent.mkdir(parents=True,exist_ok=True)
+        shutil.copy2(source,target)
+        copied.append(target)
+    for source in sorted((ROOT/'figures').iterdir()):
+        if source.suffix in ['.pdf','.svg','.png']:
+            copy(source,PAPER/'figures'/source.name)
+        elif source.suffix in ['.csv','.json']:
+            copy(source,destination/'plot_sources'/source.name)
+    for source in sorted((ROOT/'summary').iterdir()):
+        if source.suffix in ['.csv','.json']:
+            copy(source,destination/source.name)
+    for name in ['MODEL_ASSETS.json','INPUTS.json','RUN_SIGNATURE.json','FINAL_MANIFEST.json']:
+        copy(ROOT/'results'/name,destination/('RUN_'+name if name=='FINAL_MANIFEST.json' else name))
+    for name in ['QA.json','QA_base.json','PUBLICATION_QA.json','TEST_RESULTS.xml',
+                 'RUN_PROTOCOL.md','analysis-report.md','stats-appendix.md','figure-catalog.md']:
+        copy(ROOT/name,destination/name)
+    for source in sorted(ROOT.glob('*.py')):
+        copy(source,scripts/source.name)
+    # Per-sample primary 2K endpoints, rather than multi-million-row duplicates
+    # from all block/model sensitivities. Complete arrays remain in the datalake.
+    rows=[]
+    for source in sorted((ROOT/'results/partitions').glob('*/random_donor_n10_seed*/sample_metrics.csv.gz')):
+        frame=pd.read_csv(source)
+        rows.append(frame[(frame.model_id=='2k_e3')&frame.block.isin(['input','encoded'])])
+    combined=pd.concat(rows,ignore_index=True)
+    target=destination/'primary_sample_metrics.csv.gz'
+    combined.to_csv(target,index=False,compression='gzip');copied.append(target)
+    readme=destination/'README.md'
+    readme.write_text('''# Figure 6: within-cohort paired calibration
+
+Primary paper model: frozen PBISC-Diffusion2K E3 EMA, not B048. Exactly10 paired
+donors calibrate each of100 selections per study. This bundle includes all
+split assignments, numerical figure source tables, primary sample metrics,
+model/source provenance and independent QA. Complete600-partition coefficient
+arrays and sensitivity sample metrics remain in the analysis datalake path
+recorded in RUN_FINAL_MANIFEST.json.
+
+Depth is standardized at K1000 and a fixed training anchor; this is not native
+count-depth recovery or generated-cell validation. Original SC reference gene
+support is retained. The same-condition mean control is explicitly post-hoc.
+Selection ranges are not independent-cohort confidence intervals.
+
+Plots can be rebuilt from publication_split_summary.csv using
+scripts/cohort_calibration/make_figures.py. Full experiment reproduction also
+requires the processed expression sources and frozen checkpoints listed in
+INPUTS.json/MODEL_ASSETS.json; edit the recorded local asset paths for another
+host. Checkpoints and raw sequencing data are not bundled or claimed newly
+public. The primary papers and GEO/Zenodo accessions are cited in Methods.
+''')
+    copied.append(readme)
+    manifest=dict(state='completed',created_at=datetime.now(timezone.utc).isoformat(),
+                  experiment_root=str(ROOT),n_partitions=600,n_primary_sample_metric_rows=len(combined),
+                  qa='passed; see QA.json and PUBLICATION_QA.json',
+                  files=[dict(bundled_path=str(p.relative_to(PAPER)),bytes=p.stat().st_size,sha256=sha(p)) for p in sorted(set(copied))])
+    (destination/'MANIFEST.json').write_text(json.dumps(manifest,indent=2)+'\n')
+    print(json.dumps(dict(state='completed',n_files=len(manifest['files']),
+                          primary_sample_metric_rows=len(combined),destination=str(destination))))
+
+
+if __name__=='__main__':main()

paper-releases/pbisc-diffusion/updates/20260929T062828Z-manuscript-relocation/diffusion/dataset/server_1/model8/plans/threshold_rescore_20260922/tools/write_architecture_table.py

@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+"""Regenerate tables/results_architecture.tex from the scANVI snapshot.
+
+The four width arms are read out three ways on the same checkpoints: the inherited cut at p>0.5, the
+swept cut at tau=-2.0, and the drawn read-out that samples each gene from its predicted detection
+probability.  Writing the table from the snapshot rather than by hand keeps the manuscript numbers tied
+to the metric files, and makes it obvious when an arm has not been scored yet.
+
+The response column is the change in detected genes per cell per unit of threshold, measured at part 16
+between the two thresholded read-outs on the 2,980-gene panel; it is what shows that a wide trunk stops
+responding to the read-out at all.
+"""
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+M8 = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion")
+A = M8 / "analysis/scsemi_benchmark_diffusion_20260913"
+OUT = Path("/home/sj_server_1/PBISC/paper-release-20260928/manuscript"  # moved from bulk2sc-paper/nature-communications-overleaf on 2026-09-29
+           "/tables/results_architecture.tex")
+EPOCHS = (14, 15, 16)
+# label, trunk, conditioning width, parameters, validation loss at part 16,
+# (default tag, swept tag, drawn tag), detected-genes response per threshold unit
+ARMS = [
+    ("Base",           r"$1\times$", "512",   "54.2M",  0.1968, ("6k", "6kt20", "bern6k"), 77),
+    ("Wide condition", r"$1\times$", "2,048", "102.1M", 0.1969, ("ctx2048", "ctx2048t20", "bernctx2048"), 79),
+    ("Wide trunk",     r"$4\times$", "512",   "240.0M", 0.1355, ("4xctx512", "4xctx512t20", "bern4xctx512"), 10),
+    ("Both wide",      r"$4\times$", "2,048", "312.6M", 0.1356, ("4x", "4xt20", "bern4x"), 6),
+]
+
+CAPTION = (
+    "Capacity lowers the loss and does not move the biology; it also removes the read-out's control. "
+    "Four models differ only in denoiser trunk width and conditioning width, trained for 16 schedule "
+    "parts on the 6,000-gene panel with the same data, seed, batch size, learning rate and schedule. "
+    "Scores are means over parts 14--16 on the held-out lupus split under three read-outs of the same "
+    "checkpoints; $r$ and DEG higher-better, composition $L_1$ lower-better. Validation loss is the "
+    "denoising objective at part 16. Response is the change in detected genes per cell when the support "
+    "threshold is lowered by one unit, measured at part 16: the two wide-trunk models barely respond, so "
+    "their sparsity cannot be set by a threshold at all."
+)
+
+
+def snapshot() -> pd.DataFrame:
+    dirs = [A / "metrics_arch_epochs/scanvi_server2", A / "metrics_prop_epochs/scanvi_server2"]
+    cands = [p for d in dirs if d.exists() for p in d.glob("summary_tidy_*.csv")]
+    if not cands:
+        raise SystemExit("no scANVI snapshot found")
+    newest = max(cands, key=lambda p: p.stat().st_mtime)
+    t = pd.read_csv(newest)
+    t = t[(t["comparison"] == "all_diffusion") & (t["celltype_gate"] == "common_celltype_gate")]
+    return t.drop_duplicates("method").set_index("method")
+
+
+def mean_scores(t: pd.DataFrame, tag: str):
+    rows = [t.loc[f"diffusion_{tag}_e{e}"] for e in EPOCHS if f"diffusion_{tag}_e{e}" in t.index]
+    if len(rows) != len(EPOCHS):
+        return None
+    return (float(np.mean([r["mean_logFC_pearson"] for r in rows])),
+            float(np.mean([r["total_top50_overlap"] for r in rows])),
+            float(np.mean([r["composition_l1"] for r in rows])))
+
+
+def main() -> int:
+    t = snapshot()
+    scores = {}
+    missing = []
+    for label, *_ , tags, _resp in [(a[0], a[1], a[2], a[3], a[4], a[5], a[6]) for a in ARMS]:
+        for which, tag in zip(("default", "swept", "drawn"), tags):
+            m = mean_scores(t, tag)
+            scores[(label, which)] = m
+            if m is None:
+                missing.append(f"{label}/{which} ({tag})")
+    if missing:
+        print("not scored yet: " + ", ".join(missing), file=sys.stderr)
+        return 1
+
+    best = {}
+    for which in ("default", "swept", "drawn"):
+        for j, better in enumerate((max, max, min)):
+            best[(which, j)] = better(scores[(a[0], which)][j] for a in ARMS)
+    # bold the best value of each metric across the whole table, not within a read-out
+    overall = [max(scores[(a[0], w)][0] for a in ARMS for w in ("default", "swept", "drawn")),
+               max(scores[(a[0], w)][1] for a in ARMS for w in ("default", "swept", "drawn")),
+               min(scores[(a[0], w)][2] for a in ARMS for w in ("default", "swept", "drawn"))]
+    best_loss = min(a[4] for a in ARMS)
+
+    def fmt(value: float, j: int) -> str:
+        text = f"{value:.3f}" if j != 1 else f"{value:.1f}"
+        return rf"\textbf{{{text}}}" if abs(value - overall[j]) < 1e-9 else text
+
+    lines = [
+        r"\begin{table}[t]", r"\centering", r"\footnotesize",
+        r"\setlength{\tabcolsep}{2.2pt}",
+        rf"\caption{{{CAPTION}}}", r"\label{tab:architecture}",
+        r"\begin{tabular}{lrrrrrrrrrrrrrr}", r"\toprule",
+        r" & \multicolumn{2}{c}{Width} & & Val. & \multicolumn{3}{c}{Cut at $p>0.5$}"
+        r" & \multicolumn{3}{c}{Cut at $\tau=-2.0$} & \multicolumn{3}{c}{Drawn from $p$} & Response \\",
+        r"\cmidrule(lr){2-3}\cmidrule(lr){6-8}\cmidrule(lr){9-11}\cmidrule(lr){12-14}",
+        r"Model & trunk & cond. & Params & loss & $r$ & DEG & $L_1$ & $r$ & DEG & $L_1$"
+        r" & $r$ & DEG & $L_1$ & genes/unit \\",
+        r"\midrule",
+    ]
+    for label, trunk, cond, params, loss, _tags, response in ARMS:
+        loss_text = rf"\textbf{{{loss:.4f}}}" if abs(loss - best_loss) < 1e-9 else f"{loss:.4f}"
+        cells = []
+        for which in ("default", "swept", "drawn"):
+            s = scores[(label, which)]
+            cells += [fmt(s[0], 0), fmt(s[1], 1), fmt(s[2], 2)]
+        lines.append(f"{label:14s} & {trunk} & {cond} & {params} & {loss_text} & "
+                     + " & ".join(cells) + f" & {response} \\\\")
+    lines += [r"\bottomrule", r"\end{tabular}", r"\end{table}", ""]
+    OUT.write_text("\n".join(lines))
+    print(f"written: {OUT}")
+    for label, *_ in ARMS:
+        s = scores[(label, "drawn")]
+        print(f"  {label:15s} drawn  {s[0]:.3f} / {s[1]:.1f} / {s[2]:.3f}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

paper-releases/pbisc-diffusion/updates/20260929T062828Z-manuscript-relocation/diffusion/dataset/server_1/model8/plans/utility_studies_final6k_20260925/behcet/tools/make_manifest.py

@@ -0,0 +1,105 @@
+#!/usr/bin/env python
+"""Write behcet/MANIFEST.json: sha256 of every input, tool and output of this task dir."""
+
+from __future__ import annotations
+
+import datetime as _dt
+import hashlib
+import json
+from pathlib import Path
+
+M8 = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion")
+T = M8 / "plans/utility_studies_final6k_20260925/behcet"
+R = M8 / "plans/behcet_c1q_repro_20260921"
+S = M8 / "plans/aifi_scanvi_rebuild_20260916"
+DL = Path("/home/sj_server_1/gblee/datalake/paired-pbmc-bulk-sc/datasets")
+PAPER = Path("/home/sj_server_1/PBISC/paper-release-20260928/manuscript")  # moved from bulk2sc-paper/nature-communications-overleaf on 2026-09-29
+SEEDS = M8 / "plans/utility_studies_final6k_20260925/behcet_seeds"
+
+
+def sha(path: Path) -> str:
+    h = hashlib.sha256()
+    with path.open("rb") as fh:
+        for chunk in iter(lambda: fh.read(1 << 24), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+INPUTS = {
+    "final_checkpoint": M8 / "runs/f01_6k_1x_seed20260627/checkpoint_epoch_016.pt",
+    "behcet_condition_npz": R / "conditions/behcet_donor_conditions.npz",
+    "observed_h5ad_GSE198616_pbmc_all8": R / "work/gse198616_pbmc_all8.h5ad",
+    "observed_build_manifest": R / "work/BUILD_MANIFEST.json",
+    "observed_labels_6k_v2_old": R / "work/scanvi_labels_real/labels.parquet",
+    "observed_labels_f01common_new": T / "work/scanvi_labels_real_f01common/labels.parquet",
+    "observed_labels_f01common_new_manifest": T / "work/scanvi_labels_real_f01common/labels.manifest.json",
+    "generated_npz_seed20260913": M8 / "analysis/scsemi_benchmark_diffusion_20260913/generated_behcetf016kbern.npz",
+    "generated_h5ad_seed20260913": R / "work_gen/generated_behcetf016kbern_poisson.h5ad",
+    "generated_labels_seed20260913": R / "work_gen/scanvi_labels_f016kbern_f01common/labels.parquet",
+    "generated_labels_seed20260913_manifest": R / "work_gen/scanvi_labels_f016kbern_f01common/labels.manifest.json",
+    "generated_npz_s1": SEEDS / "generated_behcetf016kbern_s1.npz",
+    "generated_npz_s1_receipt": SEEDS / "generated_behcetf016kbern_s1.log",
+    "generated_npz_s2": SEEDS / "generated_behcetf016kbern_s2.npz",
+    "generated_npz_s2_receipt": SEEDS / "generated_behcetf016kbern_s2.log",
+    "generated_h5ad_s1": T / "work/generated_behcetf016kbern_s1_poisson.h5ad",
+    "generated_labels_s1": T / "work/scanvi_labels_f016kbern_s1_f01common/labels.parquet",
+    "generated_h5ad_s2": T / "work/generated_behcetf016kbern_s2_poisson.h5ad",
+    "generated_labels_s2": T / "work/scanvi_labels_f016kbern_s2_f01common/labels.parquet",
+    "shared_gene_space_genf016k": R / "work/shared_gene_space_genf016k.csv",
+    "shared_gene_space_gen6k": R / "work/shared_gene_space_gen6k.csv",
+    "modules_v2": R / "modules_c1q_behcet_v2.json",
+    "analyzer_analyze_c1q_subset": R / "analyze_c1q_subset.py",
+    "marker_list_source_gate_diagnostic": R / "tools/gate_diagnostic.py",
+    "generated_identity_check_reference": R / "tools/generated_identity_check.py",
+    "scanvi_model_f01common": S / "runs/aifi_scanvi_f01common_v1/scanvi_model/model.pt",
+    "scanvi_model_f01common_manifest": S / "runs/aifi_scanvi_f01common_v1/MODEL_MANIFEST.json",
+    "predict_scanvi": S / "predict_scanvi.py",
+    "scanvi_common": S / "scanvi_common.py",
+    "poisson_readout": S / "poisson_readout.py",
+    "GSE198616_sample_info": DL / "GSE198616/raw/GSE198616_sample_info.csv.gz",
+    "GSE198533_series_matrix": DL / "GSE198533/metadata/GSE198533_series_matrix.txt.gz",
+    "existing_out_v2_real_gen6k_sample": R / "out/v2_real_gen6k/c1q_high_vs_other_monocyte_sample_summary.csv",
+    "existing_out_v2_real_gen6k_manifest": R / "out/v2_real_gen6k/ANALYSIS_MANIFEST.json",
+    "existing_out_v2_real_pbmc_all8_sample": R / "out/v2_real_pbmc_all8/c1q_high_vs_other_monocyte_sample_summary.csv",
+    "existing_out_v2_generated_thr00_sample": R / "out/v2_generated_thr00/c1q_high_vs_other_monocyte_sample_summary.csv",
+    "existing_out_v2_generated_thr00_manifest": R / "out/v2_generated_thr00/ANALYSIS_MANIFEST.json",
+    "existing_out_v2_generated_thr20_sample": R / "out/v2_generated_thr20/c1q_high_vs_other_monocyte_sample_summary.csv",
+    "existing_out_v2_generated_thr20_manifest": R / "out/v2_generated_thr20/ANALYSIS_MANIFEST.json",
+    "existing_out_v2_generated_f016kbern_sample": R / "out/v2_generated_f016kbern/c1q_high_vs_other_monocyte_sample_summary.csv",
+    "existing_out_v2_generated_f016kbern_group": R / "out/v2_generated_f016kbern/c1q_high_vs_other_monocyte_within_sample_summary.csv",
+    "existing_out_v2_generated_f016kbern_manifest": R / "out/v2_generated_f016kbern/ANALYSIS_MANIFEST.json",
+    "existing_out_three_arm_primary": R / "out/three_arm_primary.csv",
+    "existing_out_GATE_DIAGNOSTIC": R / "out/GATE_DIAGNOSTIC.json",
+    "existing_out_GENERATED_IDENTITY": R / "out/GENERATED_IDENTITY.json",
+    "manuscript_table_results_behcet_readonly": PAPER / "tables/results_behcet.tex",
+    "manuscript_results_tex_readonly": PAPER / "sections/results.tex",
+}
+
+
+def main() -> int:
+    inputs = {}
+    for key, path in INPUTS.items():
+        inputs[key] = {"path": str(path), "sha256": sha(path), "bytes": path.stat().st_size} if path.exists() else {"path": str(path), "missing": True}
+    tools = {str(p.relative_to(T)): sha(p) for p in sorted((T / "tools").glob("*")) if p.is_file()}
+    outputs = {}
+    for p in sorted(T.rglob("*")):
+        if p.is_file() and "work/" not in str(p.relative_to(T)) + "/" and p.name != "MANIFEST.json" and not str(p.relative_to(T)).startswith("tools/"):
+            outputs[str(p.relative_to(T))] = sha(p)
+    work = {str(p.relative_to(T)): sha(p) for p in sorted((T / "work").rglob("*")) if p.is_file()}
+    manifest = {
+        "schema_version": "pbisc_sdd.util6k.behcet.manifest.v1",
+        "created_at": _dt.datetime.now(tz=_dt.timezone.utc).isoformat(),
+        "task": "Behcet C1Q-high table for the final model, group-level tests, selection-rule audit (CPU only)",
+        "inputs": inputs,
+        "tools": tools,
+        "work_files": work,
+        "outputs": outputs,
+    }
+    (T / "MANIFEST.json").write_text(json.dumps(manifest, indent=2) + "\n")
+    missing = [k for k, v in inputs.items() if v.get("missing")]
+    print(f"wrote {T / 'MANIFEST.json'}; {len(inputs)} inputs, missing: {missing}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

paper-releases/pbisc-diffusion/updates/20260929T062828Z-manuscript-relocation/diffusion/metadata/request.json

@@ -0,0 +1,73 @@
+{
+  "assets": [
+    {
+      "host": "server_1",
+      "kind": "maintenance",
+      "source": "/home/sj_server_1/PBISC/paper-release-20260928/maintenance/POLICY.md",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "maintenance",
+      "source": "/home/sj_server_1/PBISC/paper-release-20260928/maintenance/SCOPE.json",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "maintenance",
+      "source": "/home/sj_server_1/PBISC/paper-release-20260928/maintenance/update.py",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "document",
+      "source": "/home/sj_server_1/PBISC/paper-release-20260928/manuscript/AGENTS.md",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "document",
+      "source": "/home/sj_server_1/PBISC/paper-release-20260928/manuscript/CLAUDE.md",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "document",
+      "source": "/home/sj_server_1/PBISC/paper-release-20260928/manuscript/README.md",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "document",
+      "source": "/home/sj_server_1/PBISC/paper-release-20260928/manuscript/editorial/narrative_evidence.md",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "document",
+      "source": "/home/sj_server_1/PBISC/paper-release-20260928/manuscript/editorial/RELOCATION_20260929.json",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "code",
+      "source": "/home/sj_server_1/PBISC/paper-release-20260928/manuscript/scripts/cohort_calibration/package_analysis.py",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "code",
+      "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/threshold_rescore_20260922/tools/write_architecture_table.py",
+      "state": "completed"
+    },
+    {
+      "host": "server_1",
+      "kind": "code",
+      "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/utility_studies_final6k_20260925/behcet/tools/make_manifest.py",
+      "state": "completed"
+    }
+  ],
+  "inference_locations": [],
+  "project_id": "pbisc-diffusion",
+  "summary": "Relocation of the PBISC-Diffusion manuscript at the user's request (2026-09-29): the Git working tree of junidude/PBISC-diffusion-overleaf moved by same-filesystem rename from /home/sj_server_1/PBISC/immun-pbmc-fm/bulk2sc-paper/nature-communications-overleaf to /home/sj_server_1/PBISC/paper-release-20260928/manuscript (691 files incl. .git, byte-identical; inventory sha256 ad556080...), and its superseded drafts from bulk2sc-paper/archived-analyses to paper-release-20260928/archived-analyses (79 files, byte-identical; not an archive payload root). This delta carries the scope/tooling change (SCOPE.json core_roots.manuscript and update.py CORE_ROOTS, POLICY.md), the manuscript files edited for the new location (README working-copy path, location-independent packaging script, one absolute reference to the BIBM draft that stayed behind, AGENTS.md policy path) with the relocation record, and the two Model 8 scripts that write into the manuscript. Unchanged manuscript files keep their baseline archive copies (same bytes, old source path)."
+}