Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/sec07_ft_modes.py

@@ -0,0 +1,327 @@
+"""Section 7, slide (a): full FT vs LoRA vs condition-encoder-only vs sc-only FT vs zero-shot,
+plus the compute cost (trainable parameters, wall-clock time per training step).
+
+Every arm is conditioned on the real bulk as measured (score arms *__raw): the zero-shot baseline is the
+pretrained model fed real bulk (zs__raw), and the sc-only control is the sc-only fine-tune fed real bulk
+(pseudo_full__raw). The condition-encoder-only arm exists only as real_condenc__raw (fold-0 call).
+Accuracy: ImMicro and sepsis, 5-fold cross-fitted, pooled scoring (one call over all donors).
+The condition-encoder-only arm exists only in the fold-0 pilot, so the L1 and coverage panels also show
+every arm from the fold-0 scoring call (open diamonds). Metrics are only comparable within one scoring call,
+so filled circles (pooled call) and open diamonds (fold-0 call) are never compared with each other.
+The sepsis-vs-control logFC r is shown for the pooled call only: the fold-0 contrast rests on 5 controls
+and the source README calls the per-fold contrast too noisy to use.
+
+Compute: time per training step (median over runs), not minutes per run, because runs early-stop at
+different steps (sepsis full 5.75k-20k, LoRA 5.75k-15k). All three modes use batch 256 and validate every
+250 steps (asserted from RUN.json), so wall seconds / steps_run is a like-for-like per-step cost.
+
+Every number is read from the score summaries / RUN.json receipts below. CPU only.
+"""
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from scipy.stats import wilcoxon
+
+DECK = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/"
+            "2026-09-30--paper-summary--r00")
+sys.path.insert(0, str(DECK / "tools"))
+import style  # noqa: E402
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.lines import Line2D  # noqa: E402
+
+PLAN = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/"
+            "paired_finetune_intradataset_20260927")
+FIGDIR, DATADIR = DECK / "figures", DECK / "data"
+DATASETS = {"immicro": "ImMicro", "ye2026_sepsis": "Sepsis"}
+CELLTYPES = ["T cell", "B cell", "Monocyte", "NK cell"]
+
+# rows top -> bottom: (score arm key, run dir name or None, label, colour)
+ARMS = [
+    ("real_full__raw", "real_full", "full FT", style.ARM["full_ft"]),
+    ("real_lora__raw", "real_lora", "LoRA FT (r 16)", style.ARM["lora"]),
+    ("real_condenc__raw", "real_condenc", "cond.-encoder-only FT", style.ARM["other"]),
+    ("pseudo_full__raw", None, "sc-only FT", style.ARM["sconly"]),
+    ("zs__raw", None, "zero-shot, real bulk", style.ARM["zs"]),
+]
+COMPUTE_ARMS = ARMS[:3]
+ZS = "zs__raw"
+
+
+def fit_lim(vals, lo, hi, step, pad_frac=0.05):
+    """Default axis limits, widened (to a multiple of step) only when a value would sit on or past an edge."""
+    vals = np.asarray(vals, float)
+    pad = pad_frac * (hi - lo)
+    if vals.min() - pad < lo:
+        lo = round(float(np.floor((vals.min() - pad) / step) * step), 10)
+    if vals.max() + pad > hi:
+        hi = round(float(np.ceil((vals.max() + pad) / step) * step), 10)
+    return lo, hi
+
+
+def load_call(ds, call):
+    root = PLAN / "scores" / ds / call
+    summ = json.loads((root / "summary.json").read_text())
+    con = pd.read_csv(root / "contrast.csv") if (root / "contrast.csv").exists() else None
+    per = pd.read_csv(root / "per_donor.csv")
+    return root, summ, con, per
+
+
+def arm_metrics(ds, call):
+    root, summ, con, per = load_call(ds, call)
+    rows = []
+    for arm, _, label, _ in ARMS:
+        if arm not in summ["arms"]:
+            continue
+        h = summ["headline"][arm]
+        r = {"dataset": ds, "call": call, "arm": arm, "label": label, "n_donors": summ["n_donors"],
+             "comp_l1": h["comp_l1"], "coverage": h["coverage"], "precision": h["precision"],
+             "source_summary": str(root / "summary.json")}
+        if con is not None:
+            c = con[(con["arm"] == arm) & (con["statistic"] == "mean_cp10k") & con["celltype"].isin(CELLTYPES)]
+            assert set(c["celltype"]) == set(CELLTYPES), (ds, call, arm)
+            r["logfc_r"] = float(c["r_generated"].mean())
+            assert abs(r["logfc_r"] - h["contrast_mean_cp10k_r_gen"]) < 1e-9
+            r["n_case"] = int(c["n_case_donors_obs"].iloc[0])
+            r["n_control"] = int(c["n_control_donors_obs"].iloc[0])
+            r["source_contrast"] = str(root / "contrast.csv")
+        rows.append(r)
+    return pd.DataFrame(rows), per
+
+
+def compute_receipts():
+    rows = []
+    for ds in DATASETS:
+        for arm, run, label, _ in COMPUTE_ARMS:
+            for f in sorted((PLAN / "runs" / ds).glob(f"fold[0-9]/{run}/RUN.json")):
+                d = json.loads(f.read_text())
+                rows.append({"dataset": ds, "arm": arm, "label": label, "fold": d["fold"], "mode": d["mode"],
+                             "lr": d["lr"], "batch_size": d["batch_size"], "eval_every": d["eval_every"],
+                             "n_trainable": d["trainable"]["n_trainable"],
+                             "n_total_base": d["trainable"]["n_total_base"],
+                             "wall_seconds": d["wall_seconds"], "steps_run": d["steps_run"],
+                             "sec_per_step": d["wall_seconds"] / d["steps_run"],
+                             "selected_step": d["selected_step"],
+                             "selected_inner_val_loss": d["selected_inner_val_loss"],
+                             "pretrained_inner_val_loss": d["pretrained_inner_val_loss"],
+                             "source_checkpoint": d["source_checkpoint"], "run_json": str(f)})
+    return pd.DataFrame(rows)
+
+
+def main():
+    acc, pers = [], {}
+    for ds in DATASETS:
+        for call in ("pooled", "fold0"):
+            df, per = arm_metrics(ds, call)
+            acc.append(df)
+            pers[(ds, call)] = per
+    acc = pd.concat(acc, ignore_index=True)
+    # sanity: the condition-encoder arm is only in the fold-0 call; every plotted arm is a real-bulk (raw) arm
+    assert not ((acc["call"] == "pooled") & (acc["arm"] == "real_condenc__raw")).any()
+    assert acc["arm"].str.endswith("__raw").all()
+    # every plotted arm is present in both calls (except cond.-encoder-only, fold-0 call only)
+    for (ds, call), g in acc.groupby(["dataset", "call"]):
+        want = {a[0] for a in ARMS} - ({"real_condenc__raw"} if call == "pooled" else set())
+        assert set(g["arm"]) == want, (ds, call, set(g["arm"]))
+
+    # paired Wilcoxon of donor-level L1 (uncorrected): full vs LoRA and each fine-tune vs zero-shot (pooled call);
+    # cond.-encoder-only vs LoRA / zero-shot (fold-0 call, the only call with that arm)
+    wil = []
+    tests = [("pooled", "real_full__raw", "real_lora__raw", "full - LoRA"),
+             ("pooled", "real_full__raw", ZS, "full - zero-shot"),
+             ("pooled", "real_lora__raw", ZS, "LoRA - zero-shot"),
+             ("fold0", "real_condenc__raw", "real_lora__raw", "cond.-encoder-only - LoRA"),
+             ("fold0", "real_condenc__raw", ZS, "cond.-encoder-only - zero-shot")]
+    for ds in DATASETS:
+        for call, a, b, name in tests:
+            per = pers[(ds, call)]
+            for arm in (a, b):  # donor mean must reproduce the headline L1
+                h = float(acc[(acc["dataset"] == ds) & (acc["call"] == call) & (acc["arm"] == arm)]["comp_l1"].iloc[0])
+                assert abs(per.loc[per["arm"] == arm, "comp_l1"].mean() - h) < 1e-9, (ds, call, arm)
+            p = per.pivot(index="donor", columns="arm", values="comp_l1")
+            x = (p[a] - p[b]).dropna()
+            wil.append({"dataset": ds, "call": call, "comparison": f"{name} comp_l1", "arm_a": a, "arm_b": b,
+                        "mean_diff": float(x.mean()), "frac_donors_a_lower": float((x < 0).mean()),
+                        "wilcoxon_p": float(wilcoxon(x).pvalue), "n": int(len(x)),
+                        "source": str(PLAN / "scores" / ds / call / "per_donor.csv")})
+    wil = pd.DataFrame(wil)
+
+    rec = compute_receipts()
+    # per-step cost is like-for-like only if batch size and validation cadence match across modes
+    assert rec["batch_size"].nunique() == 1 and rec["eval_every"].nunique() == 1
+    agg = dict(n_runs=("fold", "size"), folds=("fold", lambda s: ",".join(map(str, sorted(s)))),
+               n_trainable=("n_trainable", "first"), n_trainable_nunique=("n_trainable", "nunique"),
+               lr=("lr", "first"),
+               wall_min_median=("wall_seconds", lambda s: float(np.median(s)) / 60.0),
+               wall_min_min=("wall_seconds", lambda s: float(np.min(s)) / 60.0),
+               wall_min_max=("wall_seconds", lambda s: float(np.max(s)) / 60.0),
+               steps_run_median=("steps_run", "median"), steps_run_min=("steps_run", "min"),
+               steps_run_max=("steps_run", "max"),
+               sec_per_step_median=("sec_per_step", "median"))
+    comp = rec.groupby(["dataset", "arm", "label"]).agg(**agg).reset_index()
+    comp_all = rec.groupby(["arm", "label"]).agg(**agg).reset_index().assign(dataset="both")
+    comp = pd.concat([comp, comp_all], ignore_index=True)
+    assert (comp["n_trainable_nunique"] == 1).all()
+    n_full = int(rec.loc[rec["arm"] == "real_full__raw", "n_trainable"].iloc[0])
+    n_lora = int(rec.loc[rec["arm"] == "real_lora__raw", "n_trainable"].iloc[0])
+    ratio = n_full / n_lora
+    sps = comp[comp["dataset"] == "both"].set_index("arm")["sec_per_step_median"]
+    step_ratio = float(sps["real_full__raw"] / sps["real_lora__raw"])
+
+    acc.to_csv(DATADIR / "sec07_ft_modes_accuracy.csv", index=False)
+    rec.to_csv(DATADIR / "sec07_ft_modes_compute_runs.csv", index=False)
+    comp.to_csv(DATADIR / "sec07_ft_modes_compute.csv", index=False)
+    wil.to_csv(DATADIR / "sec07_ft_modes_wilcoxon.csv", index=False)
+
+    # ------------------------------------------------------------------ figure
+    W, H = 12.2, 4.95
+    fig = plt.figure(figsize=(W, H))
+
+    def ax_in(x, y, w, h):
+        return fig.add_axes([x / W, y / H, w / W, h / H])
+
+    ny = len(ARMS)
+    ypos = {a[0]: ny - 1 - i for i, a in enumerate(ARMS)}
+    x0, gap, ggap = 2.3, 0.36, 0.62
+    pw = (W - x0 - ggap - 3 * gap) / 5
+    top_y, top_h = 2.3, 2.1
+    # (dataset, metric, x label, xlim, ticks, show fold-0 diamonds)
+    # last tick sits inside the axis so it never touches the next panel's "0" (logFC r is the last panel).
+    # Default limits are widened from the data (fit_lim) so the zero-shot real-bulk point is never clipped;
+    # the L1 limit is shared by both cohorts so the two L1 panels stay comparable.
+    def vals(ds_list, met, calls):
+        s = acc[acc["dataset"].isin(ds_list) & acc["call"].isin(calls)][met]
+        return s.dropna().to_numpy(float)
+    l1_lim = fit_lim(vals(list(DATASETS), "comp_l1", ["pooled", "fold0"]), 0, 0.5, 0.1)
+    cov_lim = fit_lim(vals(list(DATASETS), "coverage", ["pooled", "fold0"]), 0, 1.0, 0.1)
+    r_lim = fit_lim(vals(["ye2026_sepsis"], "logfc_r", ["pooled"]), 0, 0.65, 0.05)
+    panels = [("immicro", "comp_l1", "L1 ↓", l1_lim, [0, 0.2, 0.4], True),
+              ("immicro", "coverage", "coverage ↑", cov_lim, [0, 0.4, 0.8], True),
+              ("ye2026_sepsis", "comp_l1", "L1 ↓", l1_lim, [0, 0.2, 0.4], True),
+              ("ye2026_sepsis", "coverage", "coverage ↑", cov_lim, [0, 0.4, 0.8], True),
+              ("ye2026_sepsis", "logfc_r", "logFC r ↑", r_lim, [0, 0.2, 0.4, 0.6], False)]
+    for _, _, _, lim, ticks, _ in panels:
+        assert lim[0] <= min(ticks) and max(ticks) < lim[1], (lim, ticks)
+    xs = []
+    x = x0
+    for i in range(len(panels)):
+        xs.append(x)
+        x += pw + (ggap if i == 1 else gap)
+    for i, (ds, met, xlabel, xlim, ticks, show_f0) in enumerate(panels):
+        ax = ax_in(xs[i], top_y, pw, top_h)
+        for arm, _, label, col in ARMS:
+            y = ypos[arm]
+            ax.axhline(y, color="#E6E6E6", lw=0.8, zorder=0)
+            calls = (("pooled", 0.14, "o", True), ("fold0", -0.14, "D", False)) if show_f0 else \
+                (("pooled", 0.0, "o", True),)
+            drawn = False
+            for call, dy, mk, fill in calls:
+                s = acc[(acc["dataset"] == ds) & (acc["call"] == call) & (acc["arm"] == arm)]
+                if len(s) == 0:
+                    continue
+                v = float(s[met].iloc[0])
+                ax.plot([v], [y + dy], marker=mk, ms=9 if mk == "o" else 7.5, mec=col,
+                        mfc=col if fill else "white", mew=1.8, ls="none", zorder=3)
+                drawn = True
+            if not drawn:
+                x_txt = max(0.0, (0 - xlim[0]) / (xlim[1] - xlim[0])) + 0.02  # right of the r = 0 line
+                ax.text(x_txt, y, "no pooled contrast", ha="left", va="center", fontsize=style.FS - 3,
+                        color="#6E6E6E", transform=ax.get_yaxis_transform())
+        if xlim[0] < 0:  # the axis extends below 0 (zero-shot logFC r); mark r = 0
+            ax.axvline(0, color="#BDBDBD", lw=0.8, zorder=1)
+        ax.set_xlim(*xlim)
+        ax.set_xticks(ticks)
+        ax.set_xticklabels([f"{t:g}" for t in ticks])
+        ax.set_ylim(-0.6, ny - 0.4)
+        ax.set_yticks([ypos[a[0]] for a in ARMS])
+        if i == 0:
+            ax.set_yticklabels([a[2] for a in ARMS], fontsize=style.FS - 1)
+        else:
+            ax.set_yticklabels([])
+        ax.tick_params(axis="y", length=0)
+        ax.spines["left"].set_visible(False)
+        ax.set_xlabel(xlabel, labelpad=3)
+    # dataset headers (group labels, not figure titles)
+    im_n = int(acc[(acc["dataset"] == "immicro") & (acc["call"] == "pooled")]["n_donors"].iloc[0])
+    im_n0 = int(acc[(acc["dataset"] == "immicro") & (acc["call"] == "fold0")]["n_donors"].iloc[0])
+    se = acc[(acc["dataset"] == "ye2026_sepsis") & (acc["call"] == "pooled")].iloc[0]
+    se_n0 = int(acc[(acc["dataset"] == "ye2026_sepsis") & (acc["call"] == "fold0")]["n_donors"].iloc[0])
+    hy = (top_y + top_h + 0.12) / H
+    fig.text((xs[0] + (xs[1] + pw)) / 2 / W, hy, f"ImMicro, healthy (n {im_n})", ha="center", va="bottom",
+             fontsize=style.FS, color=style.INK)
+    fig.text((xs[2] + (xs[4] + pw)) / 2 / W, hy,
+             f"Sepsis {int(se['n_case'])} vs control {int(se['n_control'])} (n {int(se['n_donors'])})",
+             ha="center", va="bottom", fontsize=style.FS, color=style.INK)
+    for xa, xb in ((xs[0], xs[1] + pw), (xs[2], xs[4] + pw)):
+        fig.add_artist(Line2D([xa / W, xb / W], [hy - 0.005, hy - 0.005], color=style.GREY, lw=0.8))
+
+    # ------------------------------------------------------------------ compute row
+    cy, ch = 0.45, 1.0
+    carms = COMPUTE_ARMS
+    cpos = {a[0]: len(carms) - 1 - i for i, a in enumerate(carms)}
+    bx = [x0, x0 + 2.75]
+    bw = [2.15, 2.0]
+    # trainable params (log)
+    ax = ax_in(bx[0], cy, bw[0], ch)
+    for arm, _, label, col in carms:
+        n = int(comp.loc[comp["arm"] == arm, "n_trainable"].iloc[0])
+        ax.barh(cpos[arm], n / 1e6, height=0.62, color=col, lw=0)
+        ax.text(n / 1e6 * 1.12, cpos[arm], f"{n / 1e6:.1f} M", va="center", ha="left", fontsize=style.FS - 2)
+    ax.set_xscale("log")
+    ax.set_xlim(0.5, 400)
+    ax.set_xticks([1, 10, 100])
+    ax.set_xticklabels(["1 M", "10 M", "100 M"])
+    ax.minorticks_off()
+    ax.set_yticks([cpos[a[0]] for a in carms])
+    ax.set_yticklabels([a[2] for a in carms], fontsize=style.FS - 1)
+    ax.tick_params(axis="y", length=0)
+    ax.spines["left"].set_visible(False)
+    ax.set_ylim(-0.6, len(carms) - 0.4)
+    ax.set_xlabel("trainable params (log)", labelpad=3)
+    # wall-clock ms per training step, median over all runs of both cohorts
+    ax = ax_in(bx[1], cy, bw[1], ch)
+    both = comp[comp["dataset"] == "both"].set_index("arm")
+    for arm, _, label, col in carms:
+        s = both.loc[arm]
+        v = 1000.0 * s["sec_per_step_median"]
+        ax.barh(cpos[arm], v, height=0.62, color=col, lw=0)
+        txt = f"{v:.0f} ms" if s["n_runs"] > 2 else f"{v:.0f} ms (fold 0)"
+        ax.text(v + 1.2, cpos[arm], txt, va="center", ha="left", fontsize=style.FS - 2)
+    ax.set_xlim(0, 80)
+    ax.set_xticks([0, 25, 50])
+    ax.set_yticks([cpos[a[0]] for a in carms])
+    ax.set_yticklabels([])
+    ax.tick_params(axis="y", length=0)
+    ax.spines["left"].set_visible(False)
+    ax.set_ylim(-0.6, len(carms) - 0.4)
+    ax.set_xlabel("wall ms / step", labelpad=3)
+
+    # legend for the two scoring calls (bottom right)
+    lx = (bx[1] + bw[1] + 0.75) / W
+    handles = [Line2D([], [], marker="o", ms=9, mfc=style.INK, mec=style.INK, ls="none",
+                      label=f"5-fold pooled (n {im_n} / {int(se['n_donors'])})"),
+               Line2D([], [], marker="D", ms=7.5, mfc="white", mec=style.INK, mew=1.8, ls="none",
+                      label=f"fold-0 pilot (n {im_n0} / {se_n0})")]
+    fig.legend(handles=handles, loc="lower left", bbox_to_anchor=(lx, (cy + 0.42) / H), frameon=False,
+               fontsize=style.FS - 2, handletextpad=0.4, labelspacing=0.5)
+    fig.text(lx + 0.004, (cy - 0.02) / H,
+             "compare same marker only\ncond.-encoder-only: fold 0 only\nlogFC r: pooled only",
+             ha="left", va="bottom", fontsize=style.FS - 3, color="#4D4D4D", linespacing=1.3)
+
+    style.save(fig, "sec07_ft_modes", FIGDIR)
+
+    # print the numbers used on the slide
+    print(f"trainable full {n_full} lora {n_lora} ratio {ratio:.2f}; per-step full/LoRA {step_ratio:.2f}")
+    print(comp[["dataset", "label", "n_runs", "folds", "lr", "n_trainable", "wall_min_median", "wall_min_min",
+                "wall_min_max", "steps_run_min", "steps_run_max", "sec_per_step_median"]]
+          .round(4).to_string(index=False))
+    print(acc[["dataset", "call", "label", "n_donors", "comp_l1", "coverage", "precision", "logfc_r",
+               "n_case", "n_control"]].round(3).to_string(index=False))
+    print(f"x limits: L1 {l1_lim}, coverage {cov_lim}, logFC r {r_lim}")
+    print(wil.drop(columns=["source"]).to_string(index=False))
+
+
+if __name__ == "__main__":
+    main()

paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/sec08_pairs_needed.py

@@ -0,0 +1,525 @@
+"""Section 8 of the 2026-09-30 paper-summary deck: how many bulk-sc pairs LoRA needs (learning curve).
+
+CPU only. Reads the pooled 5-fold scoring calls of the paired fine-tune plan and the per-fold RUN.json files,
+then draws two slide figures:
+  figures/sec08_pairs_realism.{png,pdf}  FULL: composition L1, coverage, selected step x (ImMicro, sepsis)
+  figures/sec08_pairs_signal.{png,pdf}   TALL: ImMicro NK fraction r, sepsis logFC r, sepsis top-50 DEG
+Every plotted/quoted number is read from a score file here and written with its source to data/sec08_*.csv.
+Model: final f01 6k 1x epoch-16 checkpoint + LoRA (checked against every RUN.json).
+x axis: "0" = zero-shot, the pretrained model fed the real bulk as measured (score arm zs__raw; uses no pair);
+LoRA n = training pairs, and every LoRA run also early-stops on 7-8 paired inner-val donors
+(RUN.json inner_val_rows). Every plotted arm is conditioned on the same real bulk (*__raw), so they share one
+real-bulk broadcast. Cap-limited = ran to max_steps before the patience window closed; its late inner-val loss
+drop is read from history.json.
+When the zero-shot point lies far outside every other value of a panel (BREAK_RATIO), that panel gets a y-axis
+break with a short segment at the same data scale, instead of squeezing the LoRA curve.
+"""
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from scipy.stats import wilcoxon
+
+DECK = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/"
+            "2026-09-30--paper-summary--r00")
+sys.path.insert(0, str(DECK / "tools"))
+import style  # noqa: E402
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.lines import Line2D  # noqa: E402
+
+PLAN = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/"
+            "paired_finetune_intradataset_20260927")
+FINAL_CKPT = ("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/runs/"
+              "f01_6k_1x_seed20260627/checkpoint_epoch_016.pt")
+FIGDIR, DATADIR = DECK / "figures", DECK / "data"
+FIGDIR.mkdir(exist_ok=True)
+DATADIR.mkdir(exist_ok=True)
+
+DSETS = ["immicro", "ye2026_sepsis"]
+DS_LABEL = {"immicro": "ImMicro", "ye2026_sepsis": "Sepsis"}
+LC_N = [5, 10, 20, 40]
+LORA_ARMS = [f"lora_n{n}__raw" for n in LC_N] + ["real_lora__raw"]
+RUN_ARMS = [f"lora_n{n}" for n in LC_N] + ["real_lora"]
+ZS = "zs__raw"                              # zero-shot on the real bulk as measured: 0 pairs
+X_LAB = ["0", "5", "10", "20", "40", "all"]
+KEY_LAB = ["zero_shot", "5", "10", "20", "40", "all"]  # the same points, as keys in sec08_slide_numbers.csv
+X_POS = [0.0, 1.3, 2.3, 3.3, 4.3, 5.3]    # small gap after the zero-shot point (no fine-tuning)
+BREAK_RATIO = 1.5  # y-axis break when the zero-shot point sits > 1.5 spans of the other values beyond them
+CELLTYPES = ["T cell", "B cell", "Monocyte", "NK cell"]
+DEG_TYPES = ["T cell", "B cell", "Monocyte", "NK cell", "DC"]
+RAW_ARMS = LORA_ARMS + ["zs__raw", "real_full__raw", "pseudo_full__raw"]  # all conditioned on raw real bulk
+
+REF = {  # reference-line styles; pseudo/broadcast copied from sec09 REFSTYLE so the deck is consistent
+    "pseudo": dict(ls=(0, (4, 2)), color=style.ARM["pseudo"], lw=1.6, label="Zero-shot from sc pseudobulk"),
+    "broadcast": dict(ls=(0, (1, 1.2)), color=style.ARM["broadcast"], lw=2.4, label="Real-bulk broadcast"),
+    "null": dict(ls=(0, (6, 2, 1.5, 2)), color=style.GREY, lw=1.4, label="Cohort-mean null"),
+}
+LEG_ZS = "Zero-shot, real bulk"
+
+
+def xlabel(st):
+    """x-axis label; the inner-val range is read from the RUN.json files."""
+    return f"LoRA training pairs (each run +{rng_txt(st.n_inner_val)} inner-val pairs for early stopping)"
+
+
+def close(a, b, tol=1e-6):
+    return abs(float(a) - float(b)) <= tol
+
+
+class Call:
+    """One pooled score_ft.py scoring call."""
+
+    def __init__(self, path: Path):
+        self.path = path
+        self.summary = json.loads((path / "summary.json").read_text())
+        self.contrast = pd.read_csv(path / "contrast.csv") if (path / "contrast.csv").exists() else None
+        self.per_donor = pd.read_csv(path / "per_donor.csv")
+        self.src_summary = str(path / "summary.json")
+        self.src_contrast = str(path / "contrast.csv")
+        self.src_per_donor = str(path / "per_donor.csv")
+
+    def head(self, arm, key):
+        return float(self.summary["headline"][arm][key])
+
+    def contrast_r(self, arm, col="r_generated"):
+        c = self.contrast
+        c = c[(c.arm == arm) & (c.statistic == "mean_cp10k") & (c.celltype.isin(CELLTYPES))]
+        assert set(c.celltype) == set(CELLTYPES), (self.path, arm)
+        return float(c[col].mean())
+
+    def deg(self, arm, col="top50_overlap"):
+        d = self.contrast
+        d = d[(d.arm == arm) & (d.statistic == "welch_top50_deg")]
+        assert len(d) == 5, (self.path, arm, len(d))  # T/B/Mono/NK/DC
+        return float(d[col].sum())
+
+    def deg_type(self, arm, ct, col="top50_overlap"):
+        d = self.contrast
+        d = d[(d.arm == arm) & (d.statistic == "welch_top50_deg") & (d.celltype == ct)]
+        assert len(d) == 1, (self.path, arm, ct, len(d))
+        return float(d[col].iloc[0])
+
+    def frac_r(self, arm, ct):
+        d = self.per_donor[self.per_donor.arm == arm]
+        return float(np.corrcoef(d[f"gen_{ct}"].astype(float), d[f"obs_{ct}"].astype(float))[0, 1])
+
+
+def load():
+    calls = {ds: Call(PLAN / "scores" / ds / "pooled") for ds in DSETS}
+    deg_call = Call(PLAN / "scores" / "ye2026_sepsis" / "pooled_deg")
+    for ds, c in calls.items():
+        for a in LORA_ARMS + [ZS, "zs__pseudo"]:
+            assert a in c.summary["arms"], (ds, a)
+            assert c.summary["per_arm"][a].get("receipt_verified", True) is not False, (ds, a)
+            # headline composition L1 equals the donor mean of per_donor.csv
+            assert close(c.per_donor.loc[c.per_donor.arm == a, "comp_l1"].mean(), c.head(a, "comp_l1"), 1e-9), a
+    # headline contrast means equal the per-type mean of contrast.csv
+    se = calls["ye2026_sepsis"]
+    for a in LORA_ARMS + [ZS, "zs__pseudo"]:
+        assert close(se.contrast_r(a), se.head(a, "contrast_mean_cp10k_r_gen"), 1e-9), a
+        assert close(deg_call.deg(a), deg_call.head(a, "deg_top50_sum"), 1e-9), a
+    # broadcast is the same for every arm conditioned on raw real bulk
+    bc = {a: se.contrast_r(a, "r_broadcast") for a in RAW_ARMS}
+    assert max(bc.values()) - min(bc.values()) < 1e-12, bc
+    bcd = {a: deg_call.deg(a, "top50_overlap_broadcast") for a in RAW_ARMS}
+    assert len(set(bcd.values())) == 1, bcd
+    assert close(bcd["real_lora__raw"], deg_call.head("real_lora__raw", "deg_top50_sum_broadcast"), 1e-9)
+    # non-DEG metrics agree between the two sepsis scoring calls (k-NN ties allow 1e-3)
+    for a in LORA_ARMS + [ZS]:
+        for k in ("comp_l1", "coverage"):
+            assert abs(se.head(a, k) - deg_call.head(a, k)) <= 1e-3, (a, k)
+    return calls, deg_call
+
+
+def load_steps():
+    rows = []
+    for ds in DSETS:
+        for k in range(5):
+            for arm in RUN_ARMS:
+                p = PLAN / "runs" / ds / f"fold{k}" / arm / "RUN.json"
+                r = json.loads(p.read_text())
+                assert r["source_checkpoint"] == FINAL_CKPT, (p, r["source_checkpoint"])
+                assert r["mode"] == "lora", p
+                assert r["inner_val_rows"], p  # every LoRA run early-stops on paired inner-val donors
+                hist = json.loads((p.parent / "history.json").read_text())
+                h = {int(e["step"]): float(e["inner_val_loss"]) for e in hist}
+                # cap-limited = ran to max_steps without the patience window closing
+                cap_limited = (r["steps_run"] >= r["max_steps"]
+                               and r["steps_run"] - r["selected_step"] < r["patience"] * r["eval_every"])
+                last = max(h)
+                win = r["patience"] * r["eval_every"]  # last patience window (2,000 steps)
+                drop_total = h[0] - h[last]
+                drop_last = h[last - win] - h[last]
+                rows.append({"dataset": ds, "fold": k, "arm": arm, "n_train": len(r["train_rows"]),
+                             "n_inner_val": len(r["inner_val_rows"]), "selected_step": r["selected_step"],
+                             "max_steps": r["max_steps"], "steps_run": r["steps_run"],
+                             "patience_steps": win, "cap_limited": bool(cap_limited),
+                             "val_loss_step0": h[0], "val_loss_last": h[last],
+                             "val_loss_last_window_drop_frac": drop_last / drop_total,
+                             "wall_seconds": r.get("wall_seconds"), "source": str(p),
+                             "history": str(p.parent / "history.json")})
+    st = pd.DataFrame(rows)
+    # nested subsets: n_train equals the arm size for lora_n*
+    for n in LC_N:
+        assert (st[st.arm == f"lora_n{n}"].n_train == n).all(), n
+    return st
+
+
+def load_fold_sizes():
+    """Test donors per fold (PREPARE.json); their sum must equal the donors in the pooled scoring call."""
+    out = {}
+    for ds in DSETS:
+        p = PLAN / "data" / ds / "PREPARE.json"
+        fs = json.loads(p.read_text())["fold_sizes"]
+        out[ds] = dict(test=[int(v["test"]) for v in fs.values()], source=str(p))
+    return out
+
+
+def rng_txt(v):
+    v = sorted(int(x) for x in v)
+    return f"{v[0]}" if v[0] == v[-1] else f"{v[0]}–{v[-1]}"
+
+
+def break_axis(ax, v0, lo, hi):
+    """Shrink ax to lo..hi and add a short segment above it that shows v0 at the same data scale."""
+    p = ax.get_position()
+    h_top, gap = 0.22 * p.height, 0.08 * p.height
+    h_main = p.height - h_top - gap
+    ax.set_position([p.x0, p.y0, p.width, h_main])
+    ax.set_ylim(lo, hi)
+    half = (hi - lo) * h_top / h_main / 2
+    top = ax.figure.add_axes([p.x0, p.y0 + h_main + gap, p.width, h_top], sharex=ax)
+    top.set_ylim(v0 - half, v0 + half)
+    # one tick in the short segment: the nearest value on the main axis' tick grid that fits inside it
+    step = float(np.diff(ax.get_yticks()[:2])[0]) / 2
+    tick = round(round(v0 / step) * step, 10)
+    top.set_yticks([tick] if abs(tick - v0) < half else [])
+    if not len(top.get_yticks()):  # no grid value fits: print the value next to the point instead
+        top.text(X_POS[0] + 0.3, v0, f"{v0:.2f}", va="center", ha="left", fontsize=12)
+    top.spines["bottom"].set_visible(False)
+    top.tick_params(axis="x", bottom=False, labelbottom=False)
+    mk = dict(marker=[(-1, -0.5), (1, 0.5)], markersize=10, ls="none", color=style.INK, mec=style.INK, mew=1.2,
+              clip_on=False)
+    top.plot([0], [0], transform=top.transAxes, **mk)
+    ax.plot([0], [1], transform=ax.transAxes, **mk)
+    return top
+
+
+def curve(ax, v0, vl, refs, rows, panel, ds, src0, srcl, ylim0=None):
+    """Zero-shot (real bulk) point at x=0, LoRA curve at 5..all, reference lines. Returns the axis for the title."""
+    lo, hi = min(vl), max(vl)
+    for key, val, src in refs:
+        lo, hi = min(lo, val), max(hi, val)
+    span = hi - lo
+    broken = ylim0 is None and span > 0 and (v0 - hi > BREAK_RATIO * span or lo - v0 > BREAK_RATIO * span)
+    if broken:
+        assert v0 > hi, (panel, ds, "axis break below the curve is not implemented")
+    else:
+        lo, hi = min(lo, v0), max(hi, v0)
+    pad = (hi - lo) * 0.15
+    axes = [ax]
+    if broken:
+        top = break_axis(ax, v0, lo - pad, hi + pad)
+        axes.append(top)
+    else:
+        ax.set_ylim(ylim0 if ylim0 is not None else lo - pad, hi + pad)
+    (axes[-1] if broken else ax).scatter([X_POS[0]], [v0], s=70, color=style.ARM["zs"], edgecolor=style.INK,
+                                         lw=0.6, zorder=5, clip_on=False)
+    ax.plot(X_POS[1:], vl, color=style.ARM["lora"], lw=2.0, zorder=3)
+    ax.scatter(X_POS[1:], vl, s=70, color=style.ARM["lora"], edgecolor=style.INK, lw=0.6, zorder=5)
+    rows.append(dict(panel=panel, dataset=ds, series="zero_shot", pairs=X_LAB[0], value=v0, source=src0,
+                     y_axis_break=bool(broken)))
+    for lab, v, s in zip(X_LAB[1:], vl, srcl):
+        rows.append(dict(panel=panel, dataset=ds, series="lora", pairs=lab, value=v, source=s,
+                         y_axis_break=bool(broken)))
+    for key, val, src in refs:
+        r = REF[key]
+        for a in axes:
+            a.axhline(val, ls=r["ls"], color=r["color"], lw=r["lw"], zorder=2)
+        rows.append(dict(panel=panel, dataset=ds, series=f"ref_{key}", pairs="", value=val, source=src,
+                         y_axis_break=bool(broken)))
+    ax.set_xlim(-0.5, X_POS[-1] + 0.5)
+    ax.set_xticks(X_POS)
+    ax.set_xticklabels(X_LAB)
+    # nothing is clipped: every plotted value is inside the y range of the axis that draws it
+    for v in list(vl) + [val for _, val, _ in refs] + ([] if broken else [v0]):
+        assert ax.get_ylim()[0] <= v <= ax.get_ylim()[1], (panel, ds, v, ax.get_ylim())
+    if broken:
+        assert axes[-1].get_ylim()[0] <= v0 <= axes[-1].get_ylim()[1]
+    return axes[-1]
+
+
+def fig_realism(calls, st, fold_sizes, rows, steprows):
+    fig, axs = plt.subplots(2, 3, figsize=style.FULL,
+                            gridspec_kw=dict(left=0.168, right=0.99, top=0.83, bottom=0.12,
+                                             hspace=0.36, wspace=0.30))
+    heads = ["Composition L1 ↓", "Coverage ↑", "Selected step (×1,000)"]
+    pos0 = {ri: axs[ri, 2].get_position() for ri in range(len(DSETS))}  # row extents before any y-axis break
+    title_ax = {ci: axs[0, ci] for ci in range(3)}
+    for ri, ds in enumerate(DSETS):
+        c = calls[ds]
+        for ci, key in enumerate(["comp_l1", "coverage"]):
+            ax = axs[ri, ci]
+            refs = [("pseudo", c.head("zs__pseudo", key), c.src_summary)]
+            if key == "comp_l1":
+                refs.append(("null", c.head("real_lora__raw", "comp_l1_obs_loo_null"), c.src_summary))
+            tax = curve(ax, c.head(ZS, key), [c.head(a, key) for a in LORA_ARMS], refs, rows,
+                        f"{key}", ds, c.src_summary, [c.src_summary] * 5)
+            if ri == 0:
+                title_ax[ci] = tax
+        # selected early-stopping step per fold
+        ax = axs[ri, 2]
+        s = st[st.dataset == ds]
+        for xi, arm in zip(X_POS[1:], RUN_ARMS):
+            f = s[s.arm == arm].sort_values("fold")
+            v = f.selected_step.astype(float).values / 1000
+            jit = np.linspace(-0.22, 0.22, len(v))
+            ax.scatter(xi + jit, v, s=36, color=style.ARM["lora"], edgecolor=style.INK, lw=0.5, zorder=4)
+            med = float(np.median(v))
+            ax.plot([xi - 0.36, xi + 0.36], [med, med], color=style.INK, lw=2.0, zorder=5)
+            steprows.append(dict(dataset=ds, arm=arm, pairs=X_LAB[X_POS.index(xi)], median_step=med * 1000,
+                                 min_step=v.min() * 1000, max_step=v.max() * 1000,
+                                 n_cap_limited=int(f.cap_limited.sum()),
+                                 max_last_window_drop_frac=(float(f[f.cap_limited].val_loss_last_window_drop_frac.max())
+                                                            if f.cap_limited.any() else np.nan),
+                                 n_train=rng_txt(f.n_train), n_inner_val=rng_txt(f.n_inner_val),
+                                 source=";".join(f.source)))
+        cap = float(s.max_steps.max()) / 1000
+        ax.axhline(cap, color=style.GREY, lw=1.0, ls=(0, (2, 2)), zorder=1)
+        ax.text(X_POS[0] - 0.4, cap + 0.6, f"cap {cap:g}k", color=style.GREY, fontsize=12, va="bottom")
+        ax.text(X_POS[0], cap * 0.45, "no training", color=style.GREY, fontsize=12, ha="center",
+                va="center", rotation=90)
+        ax.set_ylim(0, cap * 1.18)
+        ax.set_xlim(-0.5, X_POS[-1] + 0.5)
+        ax.set_xticks(X_POS)
+        ax.set_xticklabels(X_LAB)
+        n_all = rng_txt(s[s.arm == "real_lora"].n_train)
+        n_iv = rng_txt(s.n_inner_val)
+        n_scored = int(c.summary["n_donors"])
+        assert n_scored == sum(fold_sizes[ds]["test"]), (ds, n_scored, fold_sizes[ds]["test"])
+        ymid = (pos0[ri].y0 + pos0[ri].y1) / 2
+        fig.text(0.005, ymid + 0.075, DS_LABEL[ds], fontsize=15, fontweight="bold", va="bottom", ha="left")
+        fig.text(0.005, ymid + 0.065, f"{n_scored} donors\nall = {n_all} pairs\ninner-val = {n_iv}",
+                 fontsize=12.5, va="top", ha="left", linespacing=1.35)
+        rows.append(dict(panel="meta", dataset=ds, series="n_scored_donors", pairs="", value=n_scored,
+                         source=c.src_summary))
+        rows.append(dict(panel="meta", dataset=ds, series="n_pairs_all", pairs="all", value=n_all,
+                         source="RUN.json train_rows: " + ";".join(s[s.arm == "real_lora"].source)))
+        rows.append(dict(panel="meta", dataset=ds, series="n_inner_val", pairs="", value=n_iv,
+                         source="RUN.json inner_val_rows (all LoRA arms, 5 folds)"))
+        rows.append(dict(panel="meta", dataset=ds, series="n_test_donors_per_fold", pairs="",
+                         value="/".join(map(str, fold_sizes[ds]["test"])),
+                         source=fold_sizes[ds]["source"] + " fold_sizes test"))
+    for ci, h in enumerate(heads):
+        title_ax[ci].set_title(h, fontsize=14, pad=8)
+    x0, x1 = axs[1, 0].get_position().x0, axs[1, 2].get_position().x1
+    fig.text((x0 + x1) / 2, 0.005, xlabel(st), ha="center", va="bottom", fontsize=14)
+    handles = [
+        Line2D([0], [0], marker="o", color=style.ARM["zs"], lw=0, markeredgecolor=style.INK, markersize=8,
+               label=LEG_ZS),
+        Line2D([0], [0], marker="o", color=style.ARM["lora"], lw=2.0, markeredgecolor=style.INK, markersize=8,
+               label="LoRA fine-tune (dots in step panel = folds)"),
+        Line2D([0], [0], color=style.INK, lw=2.0, label="Median of 5 folds"),
+        Line2D([0], [0], ls=REF["pseudo"]["ls"], color=REF["pseudo"]["color"], lw=REF["pseudo"]["lw"],
+               label=REF["pseudo"]["label"]),
+        Line2D([0], [0], ls=REF["null"]["ls"], color=REF["null"]["color"], lw=REF["null"]["lw"],
+               label=REF["null"]["label"]),
+    ]
+    fig.legend(handles=handles, loc="upper left", bbox_to_anchor=(0.005, 1.0), ncol=3, frameon=False,
+               fontsize=12, handlelength=2.4, columnspacing=1.6, labelspacing=0.4, borderaxespad=0.0)
+    style.save(fig, "sec08_pairs_realism", FIGDIR)
+
+
+def fig_signal(calls, deg_call, st, rows):
+    fig, axs = plt.subplots(1, 3, figsize=style.TALL,
+                            gridspec_kw=dict(left=0.085, right=0.99, top=0.79, bottom=0.115, wspace=0.42))
+    im, se = calls["immicro"], calls["ye2026_sepsis"]
+    tax = [None] * 3
+    # a: ImMicro across-donor NK fraction r
+    tax[0] = curve(axs[0], im.frac_r(ZS, "NK cell"), [im.frac_r(a, "NK cell") for a in LORA_ARMS],
+                   [("pseudo", im.frac_r("zs__pseudo", "NK cell"), im.src_per_donor)], rows, "nk_fraction_r",
+                   "immicro", im.src_per_donor, [im.src_per_donor] * 5)
+    # other cell types, recorded for the footnote
+    for ct in ["T cell", "B cell", "Monocyte"]:
+        for a, lab in zip([ZS] + LORA_ARMS, X_LAB):
+            rows.append(dict(panel="other_fraction_r", dataset="immicro", series=ct, pairs=lab,
+                             value=im.frac_r(a, ct), source=im.src_per_donor))
+    # b: sepsis disease logFC r (mean of T/B/Mono/NK, mean_cp10k)
+    tax[1] = curve(axs[1], se.contrast_r(ZS), [se.contrast_r(a) for a in LORA_ARMS],
+                   [("pseudo", se.contrast_r("zs__pseudo"), se.src_contrast),
+                    ("broadcast", se.contrast_r("real_lora__raw", "r_broadcast"), se.src_contrast)],
+                   rows, "logfc_r", "ye2026_sepsis", se.src_contrast, [se.src_contrast] * 5)
+    # c: sepsis top-50 DEG (Welch, sum over T/B/Mono/NK/DC)
+    tax[2] = curve(axs[2], deg_call.deg(ZS), [deg_call.deg(a) for a in LORA_ARMS],
+                   [("pseudo", deg_call.deg("zs__pseudo"), deg_call.src_contrast),
+                    ("broadcast", deg_call.deg("real_lora__raw", "top50_overlap_broadcast"), deg_call.src_contrast)],
+                   rows, "top50_deg", "ye2026_sepsis", deg_call.src_contrast, [deg_call.src_contrast] * 5, ylim0=0)
+    # top-50 DEG overlap per cell type (footnote: which types carry the gain), incl. broadcast
+    for ct in DEG_TYPES:
+        for a, lab in zip([ZS] + LORA_ARMS + ["zs__pseudo"], X_LAB + ["pseudo"]):
+            rows.append(dict(panel="top50_deg_by_type", dataset="ye2026_sepsis", series=ct, pairs=lab,
+                             value=deg_call.deg_type(a, ct), source=deg_call.src_contrast))
+        rows.append(dict(panel="top50_deg_by_type", dataset="ye2026_sepsis", series=ct, pairs="broadcast",
+                         value=deg_call.deg_type("real_lora__raw", ct, "top50_overlap_broadcast"),
+                         source=deg_call.src_contrast))
+    cd = se.summary["contrast_definition"]
+    grp = f"Sepsis {cd['n_case_donors']} vs {cd['n_control_donors']}"
+    heads = [f"ImMicro, {int(im.summary['n_donors'])} donors\ndonor NK fraction r",
+             f"{grp}\ndisease logFC r", f"{grp}\ntop-50 DEG overlap"]
+    for ax, h in zip(tax, heads):
+        ax.set_title(h, fontsize=14, pad=8, linespacing=1.3)
+    x0, x1 = axs[0].get_position().x0, axs[2].get_position().x1
+    fig.text((x0 + x1) / 2, 0.005, xlabel(st), ha="center", va="bottom", fontsize=13)
+    rows.append(dict(panel="meta", dataset="ye2026_sepsis", series="n_case_vs_control", pairs="",
+                     value=f"{se.summary['contrast_definition']['n_case_donors']} vs "
+                           f"{se.summary['contrast_definition']['n_control_donors']}", source=se.src_summary))
+    handles = [  # column-wise fill: [zero-shot, pseudo] | [LoRA, broadcast]
+        Line2D([0], [0], marker="o", color=style.ARM["zs"], lw=0, markeredgecolor=style.INK, markersize=8,
+               label=LEG_ZS),
+        Line2D([0], [0], ls=REF["pseudo"]["ls"], color=REF["pseudo"]["color"], lw=REF["pseudo"]["lw"],
+               label=REF["pseudo"]["label"]),
+        Line2D([0], [0], marker="o", color=style.ARM["lora"], lw=2.0, markeredgecolor=style.INK, markersize=8,
+               label="LoRA fine-tune"),
+        Line2D([0], [0], ls=REF["broadcast"]["ls"], color=REF["broadcast"]["color"], lw=REF["broadcast"]["lw"],
+               label=REF["broadcast"]["label"]),
+    ]
+    fig.legend(handles=handles, loc="upper left", bbox_to_anchor=(0.005, 1.0), ncol=2, frameon=False,
+               fontsize=12, handlelength=2.4, columnspacing=1.2, labelspacing=0.4, borderaxespad=0.0)
+    style.save(fig, "sec08_pairs_signal", FIGDIR)
+
+
+def slide_numbers(calls, deg_call, st):
+    """Numbers quoted in the slide keywords/footnote, each with its source file."""
+    im, se = calls["immicro"], calls["ye2026_sepsis"]
+    out = []
+
+    def add(key, value, src):
+        out.append(dict(key=key, value=value, source=src))
+
+    for ds, c in calls.items():
+        for key in ("coverage", "comp_l1"):
+            add(f"{ds}.{key}.zero_shot", round(c.head(ZS, key), 3), c.src_summary)
+            add(f"{ds}.{key}.lora_n5", round(c.head("lora_n5__raw", key), 3), c.src_summary)
+            add(f"{ds}.{key}.lora_all", round(c.head("real_lora__raw", key), 3), c.src_summary)
+            add(f"{ds}.{key}.zs_pseudo", round(c.head("zs__pseudo", key), 3), c.src_summary)
+        add(f"{ds}.comp_l1.null", round(c.head("real_lora__raw", "comp_l1_obs_loo_null"), 3), c.src_summary)
+        # donor-level composition L1, each LoRA size vs zero-shot (paired Wilcoxon, uncorrected)
+        piv = c.per_donor.pivot(index="donor", columns="arm", values="comp_l1")
+        for a, lab in zip(LORA_ARMS, X_LAB[1:]):
+            x = (piv[a] - piv[ZS]).dropna()
+            add(f"{ds}.comp_l1.lora_{lab}_minus_zero_shot.mean", round(float(x.mean()), 3), c.src_per_donor)
+            add(f"{ds}.comp_l1.lora_{lab}_minus_zero_shot.frac_donors_lower", round(float((x < 0).mean()), 2),
+                c.src_per_donor)
+            add(f"{ds}.comp_l1.lora_{lab}_vs_zero_shot.wilcoxon_p", float(f"{wilcoxon(x).pvalue:.2g}"),
+                c.src_per_donor)
+        add(f"{ds}.n_scored_donors", int(c.summary["n_donors"]), c.src_summary)
+        s = st[st.dataset == ds]
+        add(f"{ds}.selected_step.lora_n5.range", rng_txt(s[s.arm == "lora_n5"].selected_step), "RUN.json: "
+            + ";".join(s[s.arm == "lora_n5"].source))
+        add(f"{ds}.n_train.all", rng_txt(s[s.arm == "real_lora"].n_train), "RUN.json: "
+            + ";".join(s[s.arm == "real_lora"].source))
+        add(f"{ds}.n_inner_val", rng_txt(s.n_inner_val), "RUN.json inner_val_rows (all LoRA arms, 5 folds)")
+        for arm in RUN_ARMS:
+            f = s[s.arm == arm]
+            add(f"{ds}.folds_cap_limited.{arm}", f"{int(f.cap_limited.sum())}/{len(f)}",
+                "RUN.json steps_run == max_steps and steps_run - selected_step < patience*eval_every: "
+                + ";".join(f.source))
+        capped = s[s.cap_limited]
+        if len(capped):
+            add(f"{ds}.cap_limited.max_val_loss_drop_frac_last_2k_steps",
+                round(float(capped.val_loss_last_window_drop_frac.max()), 4),
+                "history.json inner_val_loss (drop over last patience window / drop from step 0): "
+                + ";".join(capped.history))
+    for a, lab in zip([ZS] + LORA_ARMS + ["zs__pseudo"], KEY_LAB + ["pseudo"]):
+        add(f"immicro.nk_fraction_r.{lab}", round(im.frac_r(a, "NK cell"), 2), im.src_per_donor)
+        add(f"sepsis.logfc_r.{lab}", round(se.contrast_r(a), 2), se.src_contrast)
+        add(f"sepsis.top50_deg.{lab}", int(deg_call.deg(a)), deg_call.src_contrast)
+    add("sepsis.logfc_r.broadcast", round(se.contrast_r("real_lora__raw", "r_broadcast"), 2), se.src_contrast)
+    add("sepsis.top50_deg.broadcast", int(deg_call.deg("real_lora__raw", "top50_overlap_broadcast")),
+        deg_call.src_contrast)
+
+    def first_sustained(v, thr, strict=True):
+        """First LoRA n from which every later n is above (or >=) thr."""
+        for i in range(len(v)):
+            if all((x > thr) if strict else (x >= thr) for x in v[i:]):
+                return X_LAB[1 + i]
+        return "never"
+    bc_r = se.contrast_r("real_lora__raw", "r_broadcast")
+    bc_d = deg_call.deg("real_lora__raw", "top50_overlap_broadcast")
+    vr = [se.contrast_r(a) for a in LORA_ARMS]
+    vd = [deg_call.deg(a) for a in LORA_ARMS]
+    add("sepsis.logfc_r.first_n_above_broadcast", first_sustained(vr, bc_r), se.src_contrast)
+    add("sepsis.logfc_r.first_n_above_zero_shot", first_sustained(vr, se.contrast_r(ZS)), se.src_contrast)
+    add("sepsis.top50_deg.first_n_at_or_above_broadcast", first_sustained(vd, bc_d, strict=False),
+        deg_call.src_contrast)
+    add("sepsis.top50_deg.first_n_above_zero_shot", first_sustained(vd, deg_call.deg(ZS)), deg_call.src_contrast)
+    # across-donor fraction r per cell type: first n from which LoRA stays at/above zero-shot (footnote)
+    for ct, key in [("T cell", "T"), ("B cell", "B"), ("Monocyte", "Mono"), ("NK cell", "NK")]:
+        v = [im.frac_r(a, ct) for a in LORA_ARMS]
+        add(f"immicro.fraction_r.{key}.first_n_at_or_above_zero_shot",
+            first_sustained(v, im.frac_r(ZS, ct), strict=False), im.src_per_donor)
+        add(f"immicro.fraction_r.{key}.zero_shot", round(im.frac_r(ZS, ct), 2), im.src_per_donor)
+        add(f"immicro.fraction_r.{key}.lora_5_10_20_40_all", "/".join(f"{x:.2f}" for x in v), im.src_per_donor)
+    # top-50 DEG overlap by cell type (which types carry the gain)
+    for ct, key in [("T cell", "T"), ("B cell", "B"), ("Monocyte", "Mono"), ("NK cell", "NK"), ("DC", "DC")]:
+        v = [deg_call.deg_type(a, ct) for a in [ZS] + LORA_ARMS]
+        add(f"sepsis.top50_deg.{key}.zero_shot_5_10_20_40_all", "/".join(f"{x:g}" for x in v),
+            deg_call.src_contrast)
+        add(f"sepsis.top50_deg.{key}.broadcast", int(deg_call.deg_type("real_lora__raw", ct,
+                                                                       "top50_overlap_broadcast")),
+            deg_call.src_contrast)
+    add("sepsis.n_case", int(se.summary["contrast_definition"]["n_case_donors"]), se.src_summary)
+    add("sepsis.n_control", int(se.summary["contrast_definition"]["n_control_donors"]), se.src_summary)
+    add("final_checkpoint", FINAL_CKPT, "RUN.json source_checkpoint (all 50 LoRA runs)")
+    return pd.DataFrame(out)
+
+
+def check_against_fig3(rows):
+    """Cross-check the LoRA curve and reference lines with the plan's own fig3 source data (same numbers,
+    independent code path). The plan's fig3 n=0 point is a different zero-shot arm, so the zero-shot real-bulk
+    point is checked in load() instead (headline vs per_donor.csv / contrast.csv)."""
+    ref = pd.read_csv(PLAN / "figures" / "source_data" / "fig3_learning_curve.csv")
+    pmap = {("comp_l1", "immicro"): "3a", ("comp_l1", "ye2026_sepsis"): "3b",
+            ("coverage", "immicro"): "3c", ("coverage", "ye2026_sepsis"): "3d",
+            ("nk_fraction_r", "immicro"): "3e", ("logfc_r", "ye2026_sepsis"): "3f",
+            ("top50_deg", "ye2026_sepsis"): "3h"}
+    n = 0
+    for r in rows:
+        p = pmap.get((r["panel"], r["dataset"]))
+        if p is None or r["series"] == "zero_shot":
+            continue
+        if r["series"] == "lora":
+            m = ref[(ref.panel == p) & (ref.series == r["series"]) & (ref.draw == f"n={r['pairs']}")]
+        else:
+            m = ref[(ref.panel == p) & (ref.series == r["series"])]
+        assert len(m) == 1, (r, len(m))
+        assert abs(float(m.value.iloc[0]) - float(r["value"])) < 1e-5, (r, float(m.value.iloc[0]))
+        n += 1
+    return n
+
+
+def main():
+    calls, deg_call = load()
+    st = load_steps()
+    fold_sizes = load_fold_sizes()
+    rows, steprows = [], []
+    fig_realism(calls, st, fold_sizes, rows, steprows)
+    rows_sig = []
+    fig_signal(calls, deg_call, st, rows_sig)
+    n_checked = check_against_fig3(rows + rows_sig)
+    pd.DataFrame(rows).to_csv(DATADIR / "sec08_pairs_realism.csv", index=False, float_format="%.6g")
+    pd.DataFrame(steprows).to_csv(DATADIR / "sec08_selected_steps.csv", index=False, float_format="%.6g")
+    st.to_csv(DATADIR / "sec08_selected_steps_per_fold.csv", index=False)
+    pd.DataFrame(rows_sig).to_csv(DATADIR / "sec08_pairs_signal.csv", index=False, float_format="%.6g")
+    nums = slide_numbers(calls, deg_call, st)
+    nums.to_csv(DATADIR / "sec08_slide_numbers.csv", index=False)
+    print(nums.to_string(index=False, max_colwidth=70))
+    print(f"cross-checked {n_checked} plotted values against fig3_learning_curve.csv")
+    brk = [(r["panel"], r["dataset"]) for r in rows + rows_sig if r.get("series") == "zero_shot" and r["y_axis_break"]]
+    print(f"y-axis breaks: {brk}")
+
+
+if __name__ == "__main__":
+    main()

paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/sec09_paired_validation.py

@@ -0,0 +1,1251 @@
+#!/usr/bin/env python3
+"""Section 9 of the 2026-09-30 paper-summary deck: the final method in five paired cohorts,
+and scSemiProfiler trained on the same 5 donors.
+
+CPU only: reads score files, draws slide-sized figures. No torch, no CUDA, no generation, no scANVI.
+
+Data loading is imported unchanged from the benchmark's own figure script
+(plans/paired_finetune_intradataset_20260927/figures/make_figures.py: load_fewpair, load_pooled),
+so every plotted number and every scSemiProfiler acceptance check (same donor set, same reference
+sha, receipt verified) is the same as in the paper figures. Only the drawing is new.
+
+Outputs (deck folder):
+  figures/sec09_cohorts.{png,pdf}          slide a: five paired cohorts (table)
+  figures/sec09_fewpair_cells.{png,pdf}    slide b: 5 donors, cell-level metrics
+  figures/sec09_fewpair_disease.{png,pdf}  slide c: 5 donors, disease signal
+  figures/sec09_fivefold.{png,pdf}         slide d: all pairs, 5-fold pooled
+  data/sec09_*.csv                         every plotted number with its source file
+  data/sec09_slide_text.json               slide titles, keywords, footnotes (long "footnote" for the speaker
+                                           notes + one-line "footnote_short") and quoted numbers (+ sources),
+                                           all built by this script from the same data as the figures
+  data/sec09_quoted_numbers.csv            the quoted numbers as a flat table
+
+Baselines (paper story, user decision 2026-09-30): the no-fine-tune baseline is zero-shot on the real bulk as
+measured (few-pair arm zs_raw, pooled arm zs__raw); the sc-only control is the sc-only fine-tune fed the real bulk
+as measured (pooled arm pseudo_full__raw). The benchmark's input-adjusted arms (zs_cal, zs__cal, pseudo_full__cal)
+are dropped right after loading and never reach a figure, table or slide text; OFF_STORY guards the outputs.
+
+RERUN CHECKLIST (all 25 scSemiProfiler draws are scored; rerun whenever a score file changes):
+  1. /home/sj_server_1/miniforge3/bin/python3 tools/sec09_paired_validation.py   (no edits needed; must exit 0)
+  2. Rebuild the four slides from data/sec09_slide_text.json: every title, keyword line, footnote and quoted
+     number there is regenerated. Do not keep hand-copied slide text.
+  3. Look at the four PNGs again.
+"""
+from __future__ import annotations
+
+import os
+
+os.environ["CUDA_VISIBLE_DEVICES"] = ""  # plotting only
+
+import gzip
+import importlib
+import importlib.util
+import json
+import re
+import sys
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from scipy.stats import wilcoxon
+
+DECK = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/"
+            "2026-09-30--paper-summary--r00")
+T = Path("/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/plans/"
+         "paired_finetune_intradataset_20260927")
+SCORES = T / "scores"
+DATA = T / "data"
+FIGDIR = DECK / "figures"
+OUTDATA = DECK / "data"
+SCRIPT = DECK / "tools" / "sec09_paired_validation.py"
+DATALAKE = Path("/home/sj_server_1/gblee/datalake/paired-pbmc-bulk-sc")
+
+# 1) the benchmark's loaders (this import sets print-size rcParams) ...
+_spec = importlib.util.spec_from_file_location("pf_make_figures", T / "figures" / "make_figures.py")
+mf = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(mf)
+# 2) ... then the deck style, which must win the rcParams
+sys.path.insert(0, str(DECK / "tools"))
+import style  # noqa: E402
+
+importlib.reload(style)
+import matplotlib.pyplot as plt  # noqa: E402
+from matplotlib.lines import Line2D  # noqa: E402
+from matplotlib.patches import Patch  # noqa: E402
+from matplotlib.ticker import MaxNLocator  # noqa: E402
+
+# deck order: healthy first, whole-blood cohort last
+ORDER = ["immicro", "ye2026_sepsis", "imperial_eskd", "giroux2022", "ji2025_chili"]
+DISEASE = [d for d in ORDER if d != "immicro"]
+NAME = {"immicro": "ImMicro", "ye2026_sepsis": "Sepsis", "imperial_eskd": "Imperial",
+        "giroux2022": "Giroux", "ji2025_chili": "Ji"}
+ROWNAME = {**NAME, "ji2025_chili": "Ji (whole blood)"}
+COL = {"zs_raw": style.ARM["zs"], "scsemi5": style.ARM["scsemi"], "lora": style.ARM["lora"],
+       "sconly": style.ARM["sconly"]}
+LAB = {"zs_raw": "Zero-shot, real bulk", "scsemi5": "scSemiProfiler, 5 donors",
+       "lora": "LoRA, 5 donors", "sconly": "sc-only fine-tune",
+       "lora_all": "LoRA, all training pairs"}
+# few-pair arms kept from load_fewpair (canonical key -> arm name in summary.json); everything else is dropped
+ARM_FILE = {"lora": "lora5", "scsemi5": "scsemi5", "zs_raw": "zs_raw", "pseudo": "zs_pseudo"}
+# pooled 5-fold arms (plot key -> arm name in scores/<cohort>/pooled/summary.json)
+POOLED_ARMS = {"zs_raw": "zs__raw", "sconly": "pseudo_full__raw", "lora": "real_lora__raw"}
+# text that must not appear in any figure, CSV label or slide text of this section
+OFF_STORY = re.compile(r"calibrat|refcal|\bcal\b|_cal\b|__cal\b|α|alpha|paired cal", re.I)
+KW_MAX_CHARS = 45        # keyword line length on the slide
+FOOT_SHORT_MAX = 110     # one-line on-slide footnote
+REFSTYLE = {"pseudo": dict(ls=(0, (4, 2)), color=style.ARM["pseudo"], lw=1.6, label="Zero-shot, sc pseudobulk input"),
+            "broadcast": dict(ls=(0, (1, 1.2)), color=style.ARM["broadcast"], lw=2.4, label="Real-bulk broadcast")}
+EDGE = "#333333"
+SUB = "#595959"
+FS = style.FS
+HILITE = "#FFF3D1"
+LEFT_IN = 2.25          # room for the row labels (cohort + train/test pairs)
+WIN_FRAC = 0.8          # a cohort verdict needs one method better in >= 80 % of the shared draws
+QC = {d: DATA / d / "QC_REPORT.md" for d in DISEASE}
+ARGS = T / "tools" / "contrast_args.sh"
+README = T / "README.md"
+SCORE_FT = T / "tools" / "score_ft.py"
+PREPARE = T / "tools" / "prepare_dataset.py"
+INV = DATALAKE / "docs" / "dataset_inventory.csv"
+SOFT_SC = DATALAKE / "datasets" / "GSE314416" / "metadata" / "GSE314416_family.soft.gz"
+SOFT_BULK = DATALAKE / "datasets" / "GSE314922" / "metadata" / "GSE314922_family.soft.gz"
+
+CKPT = "runs/f01_6k_1x_seed20260627/checkpoint_epoch_016.pt"
+NUMS = {k: [] for k in "abcd"}   # slide -> [{"value", "source"}] quoted on the slide
+EXTRA = []                       # further numbers written to the CSV only
+
+
+def num(slide, value, source):
+    NUMS[slide].append({"value": str(value), "source": str(source)})
+
+
+def extra(slide, item, value, source):
+    EXTRA.append({"slide": slide, "item": item, "value": str(value), "source": str(source)})
+
+
+def rng(vals, fmt="{:d}"):
+    vals = sorted(set(vals))
+    return fmt.format(vals[0]) if len(vals) == 1 else f"{fmt.format(vals[0])}–{fmt.format(vals[-1])}"
+
+
+# ------------------------------------------------------------------------------------------
+# Evidence: every label on the cohort table must match ONE line of its source file
+# ------------------------------------------------------------------------------------------
+_TEXT_CACHE: dict = {}
+
+
+def _read_text(p: Path) -> str:
+    if p not in _TEXT_CACHE:
+        if p.suffix == ".gz":
+            with gzip.open(p, "rt", encoding="utf-8", errors="replace") as fh:
+                _TEXT_CACHE[p] = fh.read()
+        else:
+            _TEXT_CACHE[p] = p.read_text(encoding="utf-8", errors="replace")
+    return _TEXT_CACHE[p]
+
+
+def evidence(p: Path, pattern: str) -> str:
+    """The full source line (with its number) that contains a match of `pattern`; the match may not span lines."""
+    text = _read_text(p)
+    m = re.search(pattern, text, flags=re.M)
+    if not m:
+        raise ValueError(f"evidence not found: /{pattern}/ in {p}")
+    if "\n" in m.group(0):
+        raise ValueError(f"evidence /{pattern}/ spans lines in {p}")
+    a = text.rfind("\n", 0, m.start()) + 1
+    b = text.find("\n", m.end())
+    line = text[a:(len(text) if b < 0 else b)].strip()
+    return f"L{text.count(chr(10), 0, m.start()) + 1}: {line[:240]}"
+
+
+def match_text(p: Path, pattern: str) -> str:
+    """Only the matched text (one line) of `pattern` in p, after the same checks as evidence()."""
+    evidence(p, pattern)
+    return re.search(pattern, _read_text(p), flags=re.M).group(0)
+
+
+def evidences(pairs) -> tuple[str, str]:
+    ev = [evidence(p, pat) for p, pat in pairs]
+    return " || ".join(ev), " ; ".join(sorted({str(p) for p, _ in pairs}))
+
+
+# display text -> [(file, one-line regex), ...]; each regex must match a single line of its file
+COHORT_TEXT = {
+    "immicro": {
+        "study": ("ImMicro 2026", [(INV, r"Babdor et al\., Cell 2026 ImMicro")]),
+        "disease": ("healthy", [(INV, r'Babdor et al\., Cell 2026 ImMicro",Healthy humans')]),
+        "bulk": ("PBMC, mRNA-Seq", [(SOFT_BULK, r"^!Sample_source_name_ch1 = PBMC *$"),
+                                    (SOFT_BULK, r"Universal Plus mRNA-Seq with NuQuant kit")]),
+        "sc": ("10x 5′ v1.1 CITE-seq", [(SOFT_SC, r"Chromium Next GEM Single Cell 5.{1,2} v1\.1 Library Kit"),
+                                         (SOFT_SC, r"Library strategy: CITE-seq")]),
+    },
+    "ye2026_sepsis": {
+        "study": ("Ye 2026", [(QC["ye2026_sepsis"], r"Ye Q et al\., Nat Immunol 2026")]),
+        "disease": ("sepsis", [(QC["ye2026_sepsis"], r"^# Ye 2026 sepsis")]),
+        # extraction is total RNA (Trizol) but the library (polyA vs rRNA depletion) is not stated -> show PBMC only
+        "bulk": ("PBMC", [(QC["ye2026_sepsis"], r"GEO protocol: PBMC total RNA \(Trizol\).*polyA vs rRNA depletion not stated")]),
+        "sc": ("10x 5′", [(QC["ye2026_sepsis"], r"Chemistry: GEO says .*but the data show 5'")]),
+    },
+    "imperial_eskd": {
+        "study": ("Imperial ESKD", [(QC["imperial_eskd"], r"patients with COVID-19, Imperial cohort")]),
+        "disease": ("COVID-19, dialysis", [(QC["imperial_eskd"], r"haemodialysis \(end-stage kidney disease\) patients with COVID-19")]),
+        "bulk": ("PBMC", [(QC["imperial_eskd"], r"^- Bulk: .*PBMC RNA-seq")]),
+        "sc": ("10x 5′ v2 CITE-seq", [(QC["imperial_eskd"], r"^- scRNA: .*PBMC CITE-seq"),
+                                       (QC["imperial_eskd"], r"Chemistry: 10x 5′ v2")]),
+    },
+    "giroux2022": {
+        "study": ("Giroux 2022", [(QC["giroux2022"], r"Giroux NS et al\., Sci Rep 2022")]),
+        "disease": ("COVID-19, outpatient", [(QC["giroux2022"], r"Duke outpatients, early COVID-19")]),
+        "bulk": ("PBMC, total RNA (SoLo)", [(QC["giroux2022"], r"^# Giroux 2022 COVID-19: paired PBMC bulk and scRNA"),
+                                            (QC["giroux2022"], r"GEO protocol: Zymo Direct-zol total RNA, NuGEN Ovation SoLo")]),
+        "sc": ("10x 5′ v1", [(QC["giroux2022"], r"10x library per subject-day \(5' v1")]),
+    },
+    "ji2025_chili": {
+        "study": ("Ji 2025", [(QC["ji2025_chili"], r"Ji C et al\., Cancer Immunol Immunother 2025")]),
+        "disease": ("ICI liver injury", [(QC["ji2025_chili"], r"liver injury on checkpoint inhibitors \(ChILI\)")]),
+        "bulk": ("whole blood, 3′ QuantSeq", [(QC["ji2025_chili"], r"\*\*PAXgene whole-blood tubes\*\*"),
+                                              (QC["ji2025_chili"], r"Protocol: Lexogen QuantSeq 3′ FWD")]),
+        "sc": ("10x 5′", [(QC["ji2025_chili"], r"Chemistry: 10x 5′ GEX")]),
+    },
+}
+# display labels of the scorer's case / control groups; the group lists are asserted against summary.json and
+# contrast_args.sh; `neither` = groups scored in neither arm of the contrast
+CONTRAST = {
+    "ye2026_sepsis": dict(case="sepsis", control="control", case_groups=["adult_sepsis", "ped_sepsis"],
+                          control_groups=["adult_HC", "adult_ICU_ctrl", "ped_HC", "ped_ICU_ctrl"], neither=[],
+                          ev=[(ARGS, r'ye2026_sepsis\) echo "--case-groups adult_sepsis,ped_sepsis '
+                                     r'--control-groups adult_HC,adult_ICU_ctrl,ped_HC,ped_ICU_ctrl"')]),
+    "imperial_eskd": dict(case="COVID+", control="COVID−", case_groups=["covid_positive"],
+                          control_groups=["covid_negative"], neither=["recovery"],
+                          ev=[(ARGS, r"COVID-positive vs COVID-negative haemodialysis patients \(recovery samples in neither group\)"),
+                              (ARGS, r'imperial_eskd\) echo "--case-groups covid_positive --control-groups covid_negative"')]),
+    # the GEO label is "severe"; the paper and the author sheet call this group "moderate" (all are outpatients)
+    "giroux2022": dict(case="moderate", control="mild", case_groups=["severe"], control_groups=["mild"], neither=[],
+                       ev=[(QC["giroux2022"], r'The paper and the author sheet call the "severe" group "moderate"'),
+                           (ARGS, r'giroux2022\) echo "--case-groups severe --control-groups mild"')]),
+    "ji2025_chili": dict(case="ChILI", control="ICI control", case_groups=["TOL", "FU1"], control_groups=["BC", "AC"],
+                         neither=[],
+                         ev=[(ARGS, r"ChILI patients \(liver injury and its follow-up\) vs checkpoint-inhibitor controls"),
+                             (ARGS, r'ji2025_chili\) echo "--case-groups TOL,FU1 --control-groups BC,AC"')]),
+}
+
+
+def usable_folds(ds):
+    f = pd.read_csv(DATA / ds / "folds.csv")
+    f["donor_id"] = f["donor_id"].astype(str)
+    f["sample_id"] = f["sample_id"].astype(str)
+    return f[f.fold >= 0].copy()
+
+
+def min_cells() -> int:
+    return int(re.search(r"^MIN_CELLS = (\d+)", _read_text(PREPARE), flags=re.M).group(1))
+
+
+def cohort_rows():
+    mc = min_cells()
+    rows = []
+    for ds in ORDER:
+        r = {"dataset": ds}
+        for key, (txt, pairs) in COHORT_TEXT[ds].items():
+            r[key] = txt
+            r[f"{key}_evidence"], r[f"{key}_source"] = evidences(pairs)
+        dj = DATA / ds / "fewshot_n5" / "DRAWS.json"
+        draws = json.loads(dj.read_text())
+        used = usable_folds(ds)
+        if len(used) != draws["n_usable_rows"] or used.donor_id.nunique() != draws["n_usable_donors"]:
+            raise ValueError(f"{ds}: folds.csv and DRAWS.json disagree on usable rows/donors")
+        # exclusion tally, re-derived with prepare_dataset.py's rule: identity_pass & n_cells >= MIN_CELLS
+        feats = np.load(DATA / ds / "features.npz", allow_pickle=False)
+        samples = pd.read_csv(DATA / ds / "samples.csv")
+        n_all = int(feats["bulk_raw"].shape[0])
+        idp = feats["identity_pass"].astype(bool) if "identity_pass" in feats.files else np.ones(n_all, bool)
+        chk = (pd.Series(feats["identity_check"].astype(str)).str.split(":").str[0].str.strip()
+               if "identity_check" in feats.files else pd.Series(["(no flag)"] * n_all))
+        low = (samples["n_cells"].to_numpy() < mc)
+        n_fail, n_border = int((chk == "fail").sum()), int((chk == "borderline").sum())
+        if int((~idp).sum()) != n_fail + n_border:
+            raise ValueError(f"{ds}: identity_pass False rows != fail + borderline")
+        n_low = int((idp & low).sum())
+        if n_all - n_fail - n_border - n_low != draws["n_usable_rows"]:
+            raise ValueError(f"{ds}: exclusion tally does not reproduce the usable rows")
+        pooled_sum = json.loads((SCORES / ds / "pooled" / "summary.json").read_text())
+        if pooled_sum["n_donors"] != draws["n_usable_rows"]:
+            raise ValueError(f"{ds}: pooled scored n {pooled_sum['n_donors']} != usable rows {draws['n_usable_rows']}")
+        r.update(n_pairs_all=n_all, n_identity_fail=n_fail, n_identity_borderline=n_border,
+                 n_low_cells=n_low, min_cells=mc, identity_flag=("identity_pass" in feats.files),
+                 n_rows=draws["n_usable_rows"], n_donors=draws["n_usable_donors"], n_rows_source=str(dj),
+                 exclusion_source=f"{DATA / ds / 'features.npz'} (identity_pass, identity_check); "
+                                  f"{DATA / ds / 'samples.csv'} (n_cells); {PREPARE} (MIN_CELLS)")
+        cd = pooled_sum["contrast_definition"]
+        psrc = str(SCORES / ds / "pooled" / "summary.json")
+        if ds in CONTRAST:
+            C = CONTRAST[ds]
+            if list(cd["case_groups"]) != C["case_groups"] or list(cd["control_groups"]) != C["control_groups"]:
+                raise ValueError(f"{ds}: contrast groups changed: {cd['case_groups']} / {cd['control_groups']}")
+            neither = sorted({g for g, a, b in zip(cd["groups"], cd["case"], cd["ctrl"]) if not a and not b})
+            if neither != sorted(C["neither"]) or (int(cd.get("n_donors_in_neither", 0)) > 0) != bool(neither):
+                raise ValueError(f"{ds}: groups in neither arm {neither} != expected {C['neither']}")
+            ev, evsrc = evidences(C["ev"])
+            r.update(case_label=C["case"], control_label=C["control"], n_case=int(cd["n_case_donors"]),
+                     n_control=int(cd["n_control_donors"]), n_neither=int(cd.get("n_donors_in_neither", 0)),
+                     neither_label=",".join(neither), contrast_source=psrc, contrast_label_evidence=ev,
+                     contrast_label_source=evsrc)
+        else:
+            if cd.get("enabled"):
+                raise ValueError(f"{ds}: expected no contrast")
+            r.update(case_label="", control_label="", n_case=np.nan, n_control=np.nan, n_neither=0,
+                     neither_label="", contrast_source=psrc)
+        rows.append(r)
+    return pd.DataFrame(rows)
+
+
+def fig_cohorts(tab: pd.DataFrame):
+    """Plain table: columns are placed by measuring the rendered text, so nothing overlaps."""
+    headers = ["Cohort", "Condition", "Bulk", "scRNA-seq", "Pairs (donors)", "Case vs control"]
+    center = {4}
+    body = []
+    for r in tab.itertuples():
+        pd_txt = f"{r.n_rows}" if r.n_rows == r.n_donors else f"{r.n_rows} ({r.n_donors})"
+        ctxt = (f"{r.case_label} {int(r.n_case)} vs {r.control_label} {int(r.n_control)}" if r.case_label
+                else "none (healthy)")
+        csub = f"+{r.n_neither} {r.neither_label} (neither)" if r.n_neither else None
+        body.append([(r.study, {}, None), (r.disease, {}, None),
+                     (r.bulk, {"fontweight": "bold"} if r.dataset == "ji2025_chili" else {}, None),
+                     (r.sc, {}, None), (pd_txt, {}, None), (ctxt, {} if r.case_label else {"color": SUB}, csub)])
+    W, H = style.FULL[0], 3.7
+    fig = plt.figure(figsize=(W, H))
+    ax = fig.add_axes([0.0, 0.0, 1.0, 1.0])
+    ax.set_axis_off()
+    ax.set_xlim(0, 1)
+    ax.set_ylim(0, 1)
+    n = len(body)
+    top, rowh = 0.87, 0.84 / n
+    texts = [[] for _ in headers]
+    for j, h in enumerate(headers):
+        texts[j].append(ax.text(0, top + 0.055, h, ha="left", va="center", fontsize=FS, fontweight="bold",
+                                color=style.INK))
+    for i, row in enumerate(body):
+        y = top - rowh * (i + 0.5)
+        for j, (txt, kw, sub) in enumerate(row):
+            if sub:
+                texts[j].append(ax.text(0, y + 0.03, txt, ha="left", va="center", fontsize=FS,
+                                        **{"color": style.INK, **kw}))
+                texts[j].append(ax.text(0, y - 0.035, sub, ha="left", va="center", fontsize=FS - 3, color=SUB))
+            else:
+                texts[j].append(ax.text(0, y, txt, ha="left", va="center", fontsize=FS,
+                                        **{"color": style.INK, **kw}))
+    fig.canvas.draw()
+    rend = fig.canvas.get_renderer()
+    widths = [max(t.get_window_extent(rend).width for t in col) / (fig.dpi * W) for col in texts]
+    gap = (1.0 - sum(widths)) / (len(headers) - 1)
+    if gap < 0.012:
+        raise ValueError(f"cohort table too wide for the slide (gap {gap:.3f})")
+    x = 0.0
+    for j, col in enumerate(texts):
+        for t in col:
+            if j in center:
+                t.set_x(x + widths[j] / 2)
+                t.set_ha("center")
+            else:
+                t.set_x(x)
+        x += widths[j] + gap
+    ax.plot([0, 1], [top + 0.005] * 2, color=style.INK, lw=1.2)
+    for i in range(1, n):
+        ax.plot([0, 1], [top - rowh * i] * 2, color="#C8C8C8", lw=0.7)
+    ax.plot([0, 1], [top - rowh * n] * 2, color=style.INK, lw=1.2)
+    style.save(fig, "sec09_cohorts", FIGDIR)
+
+
+# ------------------------------------------------------------------------------------------
+# Shared drawing helpers: rows = cohorts, columns = metrics, horizontal bars
+# ------------------------------------------------------------------------------------------
+BAR_H = 0.72
+
+
+def jitter(n, seed):
+    r = np.random.default_rng(seed)
+    return (r.random(n) - 0.5) * BAR_H * 0.6
+
+
+def hbars(ax, arms, vals_of, seed=0, dots=True):
+    """arms drawn top to bottom at y = 0..na-1. vals_of(arm) -> list of floats. Returns stats."""
+    out = {}
+    for i, arm in enumerate(arms):
+        vals = np.asarray([v for v in vals_of(arm) if not np.isnan(v)], float)
+        if len(vals) == 0:
+            ax.text(0.01, i, "not run", transform=ax.get_yaxis_transform(), va="center", fontsize=FS - 3,
+                    color=SUB)
+            out[arm] = (np.nan, np.nan, 0)
+            continue
+        m = float(vals.mean())
+        sd = float(vals.std(ddof=1)) if len(vals) > 1 else 0.0
+        ax.barh(i, m, height=BAR_H, color=COL[arm], edgecolor=EDGE, linewidth=0.6, zorder=2)
+        if dots and len(vals) > 1:
+            ax.errorbar(m, i, xerr=sd, fmt="none", ecolor="#111111", elinewidth=1.0, capsize=3, capthick=1.0,
+                        zorder=4)
+            # dots unclipped so a draw at an axis end (precision 1.0) stays whole; check_inside() asserts that every
+            # dot and bar lies within the final x-limits. SD whiskers stay clipped (a count's SD can reach below 0).
+            ax.scatter(vals, i + jitter(len(vals), seed + i), s=16, facecolor="white", edgecolor="#111111",
+                       linewidth=0.7, zorder=5, clip_on=False)
+        out[arm] = (m, sd, len(vals))
+    ax.set_ylim(len(arms) - 0.5, -0.5)
+    ax.set_yticks([])
+    return out
+
+
+def check_inside(fig, name):
+    """Every draw dot, bar end and reference line lies within its panel's final x-limits (nothing cut off)."""
+    from matplotlib.collections import PathCollection
+    from matplotlib.patches import Rectangle
+    for ax in fig.axes:
+        lo, hi = sorted(ax.get_xlim())
+        eps = 1e-9 * max(1.0, abs(hi - lo))
+        xs = [float(x) for c in ax.collections if isinstance(c, PathCollection) for x in c.get_offsets()[:, 0]]
+        xs += [x for p in ax.patches if isinstance(p, Rectangle) for x in (p.get_x(), p.get_x() + p.get_width())]
+        xs += [float(x) for ln in ax.lines for x in np.atleast_1d(ln.get_xdata()) if ln.get_linestyle() != "None"]
+        bad = [x for x in xs if not (lo - eps <= x <= hi + eps)]
+        if bad:
+            raise ValueError(f"{name}: value(s) {bad[:3]} outside the x-limits ({lo:.3g}, {hi:.3g})")
+
+
+def ref_line(ax, x, kind, na):
+    """Reference line above the bars (zorder 2) but below error bars (4) and draw dots (5)."""
+    if x is None or np.isnan(x):
+        return
+    s = REFSTYLE[kind]
+    ax.plot([x, x], [-0.5, na - 0.5], ls=s["ls"], color=s["color"], lw=s["lw"], zorder=3,
+            solid_capstyle="butt")
+
+
+def row_label(ax, name, subs):
+    """Cohort name (FS) with grey sub-lines (FS-3) stacked below it, the block centred on the row."""
+    lines = [(name, FS, style.INK)] + [(s, FS - 3, SUB) for s in subs]
+    y = sum(fs * 1.3 for _, fs, _ in lines) / 2
+    for txt, fs, col in lines:
+        h = fs * 1.3
+        ax.annotate(txt, xy=(0, 0.5), xycoords="axes fraction", xytext=(-12, y - h / 2), textcoords="offset points",
+                    ha="right", va="center", fontsize=fs, color=col)
+        y -= h
+
+
+def label_header(ax, text):
+    ax.annotate(text, xy=(0, 1), xycoords="axes fraction", xytext=(-12, 10), textcoords="offset points",
+                ha="right", va="bottom", fontsize=FS - 3, color=SUB, style="italic")
+
+
+def blank(ax, text):
+    for sp in ax.spines.values():
+        sp.set_visible(False)
+    ax.set_xticks([])
+    ax.set_yticks([])
+    ax.text(0.5, 0.5, text, transform=ax.transAxes, ha="center", va="center", fontsize=FS - 3, color=SUB,
+            style="italic")
+
+
+def grid(nrows, ncols, figsize, left, right, top, bottom, hspace, wspace):
+    fig = plt.figure(figsize=figsize)
+    gs = fig.add_gridspec(nrows, ncols, left=left / figsize[0], right=1 - right / figsize[0],
+                          top=1 - top / figsize[1], bottom=bottom / figsize[1], hspace=hspace, wspace=wspace)
+    axs = np.empty((nrows, ncols), dtype=object)
+    for i in range(nrows):
+        for j in range(ncols):
+            axs[i, j] = fig.add_subplot(gs[i, j])
+    return fig, axs
+
+
+def shared_x(axs_col, lo, hi, xlabel, ticks=None):
+    for k, ax in enumerate(axs_col):
+        ax.set_xlim(lo, hi)
+        if ticks is not None:
+            ax.set_xticks(ticks)
+        if k < len(axs_col) - 1:
+            ax.tick_params(axis="x", labelbottom=False, length=0)
+            ax.spines["bottom"].set_visible(False)
+    axs_col[-1].set_xlabel(xlabel)
+
+
+def legend_bottom(fig, handles, ncol, y=0.0):
+    fig.legend(handles=handles, loc="lower center", bbox_to_anchor=(0.5, y), ncol=ncol, frameon=False,
+               fontsize=FS - 2, handlelength=2.2, columnspacing=1.6, borderaxespad=0.0)
+
+
+def patch(key, label=None):
+    return Patch(facecolor=COL[key], edgecolor=EDGE, linewidth=0.6, label=label or LAB[key])
+
+
+def refh(kind):
+    s = REFSTYLE[kind]
+    return Line2D([0], [0], ls=s["ls"], color=s["color"], lw=s["lw"], label=s["label"])
+
+
+def dot_handle(n_draws):
+    lab = "one draw; bar = mean ± SD" + (f" of {n_draws}" if n_draws else "")
+    return Line2D([0], [0], marker="o", lw=0, markerfacecolor="white", markeredgecolor="#111111",
+                  markersize=5, label=lab)
+
+
+# ------------------------------------------------------------------------------------------
+# Few-pair (5 donors) data
+# ------------------------------------------------------------------------------------------
+_JSON: dict = {}
+
+
+def jload(p: Path):
+    if p not in _JSON:
+        _JSON[p] = json.loads(Path(p).read_text())
+    return _JSON[p]
+
+
+def fewpair():
+    few, status = mf.load_fewpair(SCORES)
+    few = few[few.arm.isin(ARM_FILE)].reset_index(drop=True)  # zero-shot baseline = raw real bulk (zs_raw) only
+    status.to_csv(OUTDATA / "sec09_scsemi_status.csv", index=False)
+    # depth-matched precision / coverage are in each call's summary.json headline (not in load_fewpair's metrics)
+    for k in ("precision_dm", "coverage_dm"):
+        few[k] = [float(jload(T / s)["headline"][ARM_FILE[a]][k]) for s, a in zip(few.source, few.arm)]
+    return few, status
+
+
+def few_vals(few, ds, arm, metric):
+    s = few[(few.dataset == ds) & (few.arm == arm)].sort_values("draw")
+    return s[metric].astype(float).tolist(), s["draw"].tolist(), s["source"].tolist()
+
+
+def few_ref(few, ds, kind, metric):
+    """pseudo = zero-shot from sc pseudobulk (zs_pseudo arm); broadcast = raw real bulk logFC/DEG; mean over draws."""
+    if kind == "pseudo":
+        s = few[(few.dataset == ds) & (few.arm == "pseudo")]
+        v = s[metric]
+    else:
+        s = few[(few.dataset == ds) & (few.arm == "lora")]
+        v = s["broadcast_r" if metric == "contrast_r" else "deg_broadcast"]
+    v = v.dropna()
+    return (float(v.mean()) if len(v) else np.nan), ";".join(sorted(set(s["source"])))
+
+
+def n_draws(few, ds, arm):
+    return int(few[(few.dataset == ds) & (few.arm == arm)]["draw"].nunique())
+
+
+def draw_info(few, ds):
+    """Training / test pairs of the scored LoRA draws (DRAWS.json), and test subjects per contrast arm."""
+    dj = DATA / ds / "fewshot_n5" / "DRAWS.json"
+    D = jload(dj)["draw"]
+    u = usable_folds(ds)
+    lora = few[(few.dataset == ds) & (few.arm == "lora")].sort_values("draw")
+    tr, te, n_sub, sub_cc = [], [], [], []
+    for r in lora.itertuples():
+        v = D[str(int(r.draw))] if str(int(r.draw)) in D else D[int(r.draw)]
+        if int(r.n_scored) != int(v["n_test_rows"]):
+            raise ValueError(f"{ds} draw {r.draw}: scored n {r.n_scored} != DRAWS n_test_rows {v['n_test_rows']}")
+        train = {str(x) for x in v["train_donors"]}
+        if int(u[u.donor_id.isin(train)].shape[0]) != int(v["n_train_rows"]):
+            raise ValueError(f"{ds} draw {r.draw}: train rows in folds.csv != DRAWS n_train_rows")
+        tr.append(int(v["n_train_rows"]))
+        te.append(int(r.n_scored))
+        test = u[~u.donor_id.isin(train)]
+        n_sub.append(int(test.donor_id.nunique()))
+        if ds in CONTRAST:
+            C = CONTRAST[ds]
+            cs = set(test[test.group.isin(C["case_groups"])].donor_id)
+            ct = set(test[test.group.isin(C["control_groups"])].donor_id)
+            sub_cc.append((len(cs), len(ct)))
+    return {"train": rng(tr), "test": rng(te), "n_train": tr, "n_test": te, "n_test_subjects": n_sub,
+            "subjects_cc": sub_cc, "n_draws": len(tr), "source": str(dj)}
+
+
+def subjects_text(info):
+    """'2 vs 2' when a cohort's test set holds few subjects (< 10) per draw; else ''."""
+    if not info["subjects_cc"] or max(info["n_test_subjects"]) >= 10:
+        return ""
+    a = rng([c for c, _ in info["subjects_cc"]])
+    b = rng([t for _, t in info["subjects_cc"]])
+    return f"{a} vs {b}"
+
+
+def row_subs(info, per_row_draws):
+    subs = [f"train {info['train']} / test {info['test']}"]
+    st = subjects_text(info)
+    if st:
+        subs.append(f"test = {st} subjects")
+    if per_row_draws:
+        subs.append(f"{info['n_draws']} draws")
+    return subs
+
+
+def long_rows(fig, few, datasets, arms, metrics, refs):
+    rows = []
+    for ds in datasets:
+        for metric in metrics:
+            for arm in arms:
+                vals, draws, srcs = few_vals(few, ds, arm, metric)
+                for v, d, s in zip(vals, draws, srcs):
+                    rows.append({"figure": fig, "dataset": ds, "metric": metric, "series": arm, "kind": "draw",
+                                 "draw": d, "value": v, "source": s})
+                vv = [v for v in vals if not np.isnan(v)]
+                if vv:
+                    rows.append({"figure": fig, "dataset": ds, "metric": metric, "series": arm, "kind": "mean",
+                                 "draw": "", "value": float(np.mean(vv)),
+                                 "sd": float(np.std(vv, ddof=1)) if len(vv) > 1 else 0.0, "n_draws": len(vv),
+                                 "source": "derived: mean/SD of the draw rows"})
+            for kind in refs.get(metric, []):
+                v, s = few_ref(few, ds, kind, metric)
+                rows.append({"figure": fig, "dataset": ds, "metric": metric, "series": f"ref_{kind}",
+                             "kind": "reference (mean over draws)", "draw": "", "value": v, "source": s})
+    return rows
+
+
+def same_draw_compare(few, status, datasets, metric, lower_better=False, min_margin=0.0, other="scsemi5"):
+    """LoRA-5 vs another arm (scSemi-5 by default, or zero-shot zs_raw) draw by draw, on the draws where both
+    exist (same test pairs; scSemi-5 draws only when load_fewpair accepted them).
+
+    Per draw: LoRA better / other better / tie (equal values). Cohort verdict:
+      'tie'    if at least half the shared draws tie, or |mean difference| < min_margin;
+      'lora'   if LoRA is better in >= WIN_FRAC of the shared draws;
+      'other'  if the other arm is better in >= WIN_FRAC of the shared draws;
+      'split'  otherwise.
+    """
+    used = status[status.status == "used"]
+    res = {}
+    for ds in datasets:
+        a = few[(few.dataset == ds) & (few.arm == "lora")].set_index("draw")[metric].astype(float)
+        b = few[(few.dataset == ds) & (few.arm == other)].set_index("draw")[metric].astype(float)
+        cand = sorted(used[used.dataset == ds]["draw"]) if other == "scsemi5" else sorted(a.index)
+        dr = [d for d in cand if d in a.index and d in b.index and not (np.isnan(a[d]) or np.isnan(b[d]))]
+        if not dr:
+            continue
+        d = (a[dr] - b[dr]) * (-1.0 if lower_better else 1.0)  # > 0: LoRA better
+        nL, nO = int((d > 0).sum()), int((d < 0).sum())
+        nT = len(dr) - nL - nO
+        mdiff = float(a[dr].mean() - b[dr].mean())
+        if 2 * nT >= len(dr) or abs(mdiff) < min_margin:
+            verdict = "tie"
+        elif nL >= WIN_FRAC * len(dr):
+            verdict = "lora"
+        elif nO >= WIN_FRAC * len(dr):
+            verdict = "other"
+        else:
+            verdict = "split"
+        res[ds] = dict(lora=float(a[dr].mean()), other=float(b[dr].mean()), n=len(dr), n_lora=nL, n_other=nO,
+                       n_tie=nT, verdict=verdict, draws=dr)
+    return res
+
+
+def l1_vs_zeroshot_per_donor(ds):
+    """LoRA-5 minus zero-shot (raw real bulk, arm zs_raw) composition L1, averaged per donor over every draw the
+    donor was tested in, then a Wilcoxon signed-rank test over donors (units repeat across draws, so donor-draw
+    rows are not independent). The donor-draw row test is returned too, for the CSV only."""
+    s2d = dict(zip(usable_folds(ds).sample_id, usable_folds(ds).donor_id))
+    parts = []
+    for j, p in mf.draw_dirs(SCORES / ds / "fewshot_n5"):
+        pdn = pd.read_csv(p / "per_donor.csv", usecols=["arm", "donor", "comp_l1"])
+        w = pdn[pdn.arm.isin(["lora5", "zs_raw"])].pivot(index="donor", columns="arm", values="comp_l1")
+        if w[["lora5", "zs_raw"]].isna().any().any():
+            raise ValueError(f"{p}: per_donor.csv lacks lora5 or zs_raw rows for some donors")
+        x = (w["lora5"] - w["zs_raw"]).rename("diff").reset_index()
+        x["draw"] = j
+        parts.append(x)
+    x = pd.concat(parts)
+    x["donor_id"] = x["donor"].astype(str).map(s2d)
+    if x["donor_id"].isna().any():
+        raise ValueError(f"{ds}: per_donor units missing from folds.csv")
+    per = x.groupby("donor_id")["diff"].mean()
+    return dict(n_donors=int(len(per)), n_rows=int(len(x)), mean_diff=float(per.mean()),
+                frac_better=float((per < 0).mean()), p=float(wilcoxon(per).pvalue),
+                row_mean_diff=float(x["diff"].mean()), row_p=float(wilcoxon(x["diff"]).pvalue),
+                source=f"{SCORES / ds / 'fewshot_n5'}/draw<j>/per_donor.csv (comp_l1, arms lora5 / zs_raw); "
+                       f"{DATA / ds / 'folds.csv'} (sample -> donor)")
+
+
+# ------------------------------------------------------------------------------------------
+# Slide b: 5 donors, cells
+# ------------------------------------------------------------------------------------------
+def fig_fewpair_cells(few, info):
+    arms = ["zs_raw", "scsemi5", "lora"]
+    metrics = [("coverage", "Coverage ↑", (0, 1.0), [0, 0.5, 1.0]),
+               ("precision", "Precision ↑", (0, 1.0), [0, 0.5, 1.0]),
+               ("comp_l1", "Composition L1 ↓", None, None)]
+    nd_all = {info[d]["n_draws"] for d in ORDER}
+    per_row = len(nd_all) > 1
+    fig, axs = grid(len(ORDER), len(metrics), style.FULL, left=LEFT_IN, right=0.15, top=0.45, bottom=1.25,
+                    hspace=0.22, wspace=0.10)
+    l1max = 0.0
+    for i, ds in enumerate(ORDER):
+        nl = n_draws(few, ds, "lora")
+        ns = n_draws(few, ds, "scsemi5")
+        for j, (mk, _, _, _) in enumerate(metrics):
+            ax = axs[i, j]
+            st = hbars(ax, arms, lambda a: few_vals(few, ds, a, mk)[0], seed=17 * i + j)
+            ref = few_ref(few, ds, "pseudo", mk)[0]
+            ref_line(ax, ref, "pseudo", len(arms))
+            if mk == "comp_l1":
+                allv = [v for a in arms for v in few_vals(few, ds, a, mk)[0] if not np.isnan(v)] + [ref]
+                l1max = max([l1max, np.nanmax(allv)] + [m + sd for m, sd, n in st.values() if n])
+            elif max([v for a in arms for v in few_vals(few, ds, a, mk)[0] if not np.isnan(v)] + [ref]) > 1.0:
+                raise ValueError(f"{ds} {mk}: value above the fixed 0-1 axis")
+            if j == 0 and 0 < ns < nl:
+                m, sd, n = st["scsemi5"]
+                ax.text(max(m + sd, max(few_vals(few, ds, "scsemi5", mk)[0])) + 0.03, 1, f"{ns} of {nl} draws",
+                        va="center", ha="left", fontsize=FS - 3, color="#8A5A00")
+        row_label(axs[i, 0], ROWNAME[ds], row_subs(info[ds], per_row))
+    label_header(axs[0, 0], "pairs per draw")
+    for j, (mk, lab, lim, ticks) in enumerate(metrics):
+        if lim is None:  # raw zero-shot L1 reaches ~0.8: the axis follows the data (every draw dot and SD bar)
+            lim = (0, np.ceil(l1max * 1.08 * 10) / 10)
+            ticks = [round(t, 1) for t in np.arange(0, lim[1] + 1e-9, 0.2)]
+        shared_x(axs[:, j], lim[0], lim[1], lab, ticks)
+    handles = [patch("zs_raw"), patch("scsemi5"), patch("lora"), refh("pseudo"),
+               dot_handle(None if per_row else nd_all.pop())]
+    legend_bottom(fig, handles, ncol=3)
+    check_inside(fig, "sec09_fewpair_cells")
+    style.save(fig, "sec09_fewpair_cells", FIGDIR)
+    rows = long_rows("sec09_fewpair_cells", few, ORDER, arms, [m[0] for m in metrics] + ["precision_dm", "coverage_dm"],
+                     {"coverage": ["pseudo"], "precision": ["pseudo"], "comp_l1": ["pseudo"]})
+    pd.DataFrame(rows).to_csv(OUTDATA / "sec09_fewpair_cells.csv", index=False, float_format="%.6g")
+
+
+# ------------------------------------------------------------------------------------------
+# Slide c: 5 donors, disease signal
+# ------------------------------------------------------------------------------------------
+def fmt_deg(v):
+    return f"{v:.1f}" if v < 10 else f"{v:.0f}"
+
+
+def fig_fewpair_disease(few, info):
+    arms = ["zs_raw", "scsemi5", "lora"]
+    nd_all = {info[d]["n_draws"] for d in DISEASE}
+    per_row = len(nd_all) > 1
+    fig, axs = grid(len(DISEASE), 2, style.TALL, left=LEFT_IN, right=0.25, top=0.45, bottom=1.55,
+                    hspace=0.45, wspace=0.12)
+    rmin, rmax = 0.0, 0.0
+    for i, ds in enumerate(DISEASE):
+        nl = n_draws(few, ds, "lora")
+        ns = n_draws(few, ds, "scsemi5")
+        # logFC r
+        ax = axs[i, 0]
+        hbars(ax, arms, lambda a: few_vals(few, ds, a, "contrast_r")[0], seed=31 * i)
+        bc = few_ref(few, ds, "broadcast", "contrast_r")[0]
+        ref_line(ax, bc, "broadcast", len(arms))
+        ax.axvline(0, color="#555555", lw=0.8, zorder=1)
+        ax.spines["left"].set_visible(False)
+        allv = [v for a in arms for v in few_vals(few, ds, a, "contrast_r")[0] if not np.isnan(v)] + [bc]
+        rmin, rmax = min(rmin, min(allv)), max(rmax, max(allv))
+        if ns < nl:
+            ax.text(1.0, 1, f"{ns} of {nl} draws", transform=ax.get_yaxis_transform(), va="center", ha="right",
+                    fontsize=FS - 3, color="#8A5A00")
+        # top-50 DEG, own scale per cohort
+        ax = axs[i, 1]
+        st = hbars(ax, arms, lambda a: few_vals(few, ds, a, "deg")[0], seed=37 * i)
+        bcd = few_ref(few, ds, "broadcast", "deg")[0]
+        ref_line(ax, bcd, "broadcast", len(arms))
+        hi = max([bcd] + [v for a in arms for v in few_vals(few, ds, a, "deg")[0] if not np.isnan(v)])
+        hi = max(hi, 4.0)
+        ax.set_xlim(0, hi * 1.40)
+        ax.xaxis.set_major_locator(MaxNLocator(nbins=3, integer=True))
+        for k, a in enumerate(arms):
+            m, sd, n = st[a]
+            if n:
+                ax.text(0.995, k, fmt_deg(m), transform=ax.get_yaxis_transform(), va="center", ha="right",
+                        fontsize=FS - 3, color=SUB)
+        row_label(axs[i, 0], ROWNAME[ds], row_subs(info[ds], per_row))
+    label_header(axs[0, 0], "pairs per draw")
+    lo = min(-0.1, np.floor(rmin * 10) / 10)
+    hi = max(0.9, np.ceil(rmax * 10) / 10 + 0.05)
+    shared_x(axs[:, 0], lo, hi, "Disease logFC r ↑", [0, 0.4, 0.8])
+    axs[-1, 1].set_xlabel("Top-50 DEG overlap ↑")
+    for ax in axs[:, 1]:
+        ax.tick_params(axis="x", labelsize=FS - 3)
+    handles = [patch("zs_raw"), patch("scsemi5"), patch("lora"), refh("broadcast"),
+               dot_handle(None if per_row else nd_all.pop())]
+    legend_bottom(fig, handles, ncol=2)
+    check_inside(fig, "sec09_fewpair_disease")
+    style.save(fig, "sec09_fewpair_disease", FIGDIR)
+    rows = long_rows("sec09_fewpair_disease", few, DISEASE, arms, ["contrast_r", "deg"],
+                     {"contrast_r": ["broadcast"], "deg": ["broadcast"]})
+    pd.DataFrame(rows).to_csv(OUTDATA / "sec09_fewpair_disease.csv", index=False, float_format="%.6g")
+
+
+# ------------------------------------------------------------------------------------------
+# Slide d: all pairs, 5-fold pooled
+# ------------------------------------------------------------------------------------------
+def baselines_text(vals):
+    """'0.11' when both baselines round to the same 2 decimals, else '0.10–0.12'."""
+    s = sorted({f"{v:.2f}" for v in vals})
+    return s[0] if len(s) == 1 else f"{s[0]}–{s[-1]}"
+
+
+def fig_fivefold(pooled, tab):
+    arms = ["zs_raw", "sconly", "lora"]
+    metrics = [("coverage", "Coverage ↑"), ("contrast_r", "Disease logFC r ↑"), ("deg", "Top-50 DEG overlap ↑")]
+    fig, axs = grid(len(ORDER), 3, style.FULL, left=LEFT_IN, right=0.35, top=0.45, bottom=1.25,
+                    hspace=0.40, wspace=0.12)
+    rows = []
+    ji_note = None
+    tabi = tab.set_index("dataset")
+
+    def src(m, mk):
+        return m.get("deg_source", m["source"]) if mk == "deg" else m["source"]
+
+    for i, ds in enumerate(ORDER):
+        P = pooled[ds]
+        for j, (mk, _) in enumerate(metrics):
+            ax = axs[i, j]
+            if ds == "ji2025_chili":
+                ax.set_facecolor(HILITE)
+            if mk in ("contrast_r", "deg") and ds == "immicro":
+                blank(ax, "healthy cohort: no contrast")
+                continue
+            hbars(ax, arms, lambda a: [P[POOLED_ARMS[a]][mk]], dots=False)
+            for a in arms:
+                rows.append({"figure": "sec09_fivefold", "dataset": ds, "metric": mk, "series": a,
+                             "arm_file": POOLED_ARMS[a], "value": P[POOLED_ARMS[a]][mk],
+                             "n_scored": P[POOLED_ARMS[a]]["n_scored"], "source": src(P[POOLED_ARMS[a]], mk)})
+            if mk == "coverage":
+                ref = P["zs__pseudo"]["coverage"]
+                ref_line(ax, ref, "pseudo", len(arms))
+                rows.append({"figure": "sec09_fivefold", "dataset": ds, "metric": mk, "series": "ref_pseudo",
+                             "arm_file": "zs__pseudo", "value": ref, "source": P["zs__pseudo"]["source"]})
+            else:
+                m = P["real_lora__raw"]
+                ref = m["broadcast_r"] if mk == "contrast_r" else m["deg_broadcast"]
+                ref_line(ax, ref, "broadcast", len(arms))
+                rows.append({"figure": "sec09_fivefold", "dataset": ds, "metric": mk, "series": "ref_broadcast",
+                             "arm_file": "real_lora__raw (raw real bulk)", "value": ref, "source": src(m, mk)})
+            if mk == "deg":
+                hi = max([ref] + [P[POOLED_ARMS[a]]["deg"] for a in arms] + [4.0])
+                ax.set_xlim(0, hi * 1.25)
+                ax.xaxis.set_major_locator(MaxNLocator(nbins=3, integer=True))
+                ax.tick_params(axis="x", labelsize=FS - 3)
+                for k, a in enumerate(arms):
+                    v = P[POOLED_ARMS[a]]["deg"]
+                    ax.text(v + hi * 0.03, k, f"{v:.0f}", va="center", ha="left", fontsize=FS - 3, color=SUB)
+            if mk == "contrast_r":
+                ax.axvline(0, color="#555555", lw=0.8, zorder=1)
+                ax.spines["left"].set_visible(False)
+            if ds == "ji2025_chili" and mk == "coverage":
+                v_zs = P[POOLED_ARMS["zs_raw"]]["coverage"]
+                v_sc = P[POOLED_ARMS["sconly"]]["coverage"]
+                v_lo = P[POOLED_ARMS["lora"]]["coverage"]
+                txt = ax.text(max(v_zs, v_sc) + 0.03, 0.5, f"baselines {baselines_text([v_zs, v_sc])}",
+                              va="center", ha="left", fontsize=FS - 2, color="#5A5A5A")
+                ji_note = (txt, ax, P["zs__pseudo"]["coverage"])
+                ax.text(v_lo + 0.02, 2, f"{v_lo:.2f}", va="center", ha="left", fontsize=FS - 3, color=SUB)
+        n = P["real_lora__raw"]["n_scored"]
+        nd = int(tabi.loc[ds, "n_donors"])
+        if int(tabi.loc[ds, "n_rows"]) != int(n):
+            raise ValueError(f"{ds}: pooled n {n} != usable rows")
+        row_label(axs[i, 0], ROWNAME[ds], [f"n={n}" + (f" ({nd} donors)" if nd != n else "")])
+    shared_x(axs[:, 0], 0, 1.0, metrics[0][1], [0, 0.5, 1.0])
+    # the Ji annotation must end left of the pseudobulk reference line (checked on the final x scale)
+    txt, ax, ref = ji_note
+    fig.canvas.draw()
+    x1 = ax.transData.inverted().transform(txt.get_window_extent(fig.canvas.get_renderer()))[1, 0]
+    if x1 >= ref - 0.01:
+        raise ValueError(f"Ji annotation ends at {x1:.2f}, over the pseudobulk reference line at {ref:.2f}")
+    rvals = [r["value"] for r in rows if r["metric"] == "contrast_r"]
+    shared_x(axs[1:, 1], min(-0.1, np.floor(min(rvals) * 10) / 10), max(0.9, np.ceil(max(rvals) * 10) / 10),
+             metrics[1][1], [0, 0.4, 0.8])
+    axs[-1, 2].set_xlabel(metrics[2][1])
+    handles = [patch("zs_raw"), patch("sconly"), patch("lora", LAB["lora_all"]), refh("pseudo"),
+               refh("broadcast")]
+    legend_bottom(fig, handles, ncol=3)
+    check_inside(fig, "sec09_fivefold")
+    style.save(fig, "sec09_fivefold", FIGDIR)
+    pd.DataFrame(rows).to_csv(OUTDATA / "sec09_fivefold.csv", index=False, float_format="%.6g")
+    return rows
+
+
+# ------------------------------------------------------------------------------------------
+# Slide text: every keyword / footnote / quoted number is built here from the same data
+# ------------------------------------------------------------------------------------------
+BANNED = ("novel", "powerful", "robust", "state-of-the-art", "seamless", "leverage")
+
+
+def n_words(s):
+    return sum(1 for t in s.split() if re.search(r"[A-Za-z0-9]", t))
+
+
+def check_text(s, what):
+    if any(b in s.lower() for b in BANNED) or re.search(r"[가-힣]", s):
+        raise ValueError(f"banned word / Korean in {what}: {s!r}")
+    if OFF_STORY.search(s):
+        raise ValueError(f"off-story (input adjustment) wording in {what}: {s!r}")
+    return s
+
+
+def check_keywords(kws):
+    if len(kws) > 3:
+        raise ValueError(f"more than 3 keyword lines: {kws}")
+    for k in kws:
+        if n_words(k) > 7:
+            raise ValueError(f"keyword over 7 words: {k!r}")
+        if len(k) > KW_MAX_CHARS:
+            raise ValueError(f"keyword over {KW_MAX_CHARS} characters ({len(k)}): {k!r}")
+        check_text(k, "keyword")
+    return kws
+
+
+def check_foot_short(s):
+    """One on-slide line: <= FOOT_SHORT_MAX characters, no file paths."""
+    if len(s) > FOOT_SHORT_MAX or "\n" in s:
+        raise ValueError(f"footnote_short over {FOOT_SHORT_MAX} characters ({len(s)}): {s!r}")
+    if re.search(r"/home|\b(scores|data|runs|tools|figures)/|\.(json|csv|npz|md|py|pt)\b", s):
+        raise ValueError(f"file path in footnote_short: {s!r}")
+    return check_text(s, "footnote_short")
+
+
+def keyword_or_short(long, short):
+    """The long wording when it fits a keyword line, else the short one (both computed from the same numbers)."""
+    return long if len(long) <= KW_MAX_CHARS else short
+
+
+def names(dss):
+    return ", ".join(NAME[d] for d in dss)
+
+
+def fmt_p(p):
+    return f"{p:.0e}" if p < 0.01 else f"{p:.2f}"
+
+
+def main():
+    FIGDIR.mkdir(parents=True, exist_ok=True)
+    OUTDATA.mkdir(parents=True, exist_ok=True)
+
+    tab = cohort_rows()
+    tab.to_csv(OUTDATA / "sec09_cohorts.csv", index=False)
+    fig_cohorts(tab)
+
+    few, status = fewpair()
+    info = {ds: draw_info(few, ds) for ds in ORDER}
+    pooled, calls = mf.load_pooled(SCORES)
+    fig_fewpair_cells(few, info)
+    fig_fewpair_disease(few, info)
+    fig_fivefold(pooled, tab)
+
+    used = status[status.status == "used"]
+    run_json = T / "runs" / "immicro" / "fewshot_n5" / "draw0" / "lora" / "RUN.json"
+    evidence(run_json, re.escape(CKPT))  # the few-pair LoRA runs start from the final checkpoint
+    steps = int(re.search(r'"fixed_steps": (\d+)', evidence(run_json, r'"fixed_steps": \d+')).group(1))
+    lora_r = re.search(r"LoRA r = (\d+)", evidence(README, r"LoRA r = \d+ \(1\.3 M")).group(1)
+    sampling = match_text(README, r"test donors only, 1,500 cells each, seed \d+, DDIM 50, EMA, Bernoulli read-out")
+    early = match_text(README, r"About 10 % of each fold's training donors are inner-validation, used only for early stopping")
+    pc_def = match_text(SCORE_FT, r"precision/coverage/density \(PCA-50 \+ k = 10 balls fitted on the scored donors' observed R1 cells\)")
+    pca_n = {int(jload(SCORES / ds / "fewshot_n5" / "draw0" / "summary.json")["reference"]["pca"]["n_components"])
+             for ds in ORDER}
+    if pca_n != {50}:
+        raise ValueError(f"PCA components differ from the footnote: {pca_n}")
+    ddef = jload(SCORES / "ji2025_chili" / "fewshot_n5" / "draw0" / "summary.json")["deg_definition"]
+    if not ("welch_bh_deg" in ddef["functions"] and ddef["top_n"] == 50 and len(ddef["types"]) == 5):
+        raise ValueError("DEG definition differs from the footnote")
+    n_types = len(ddef["types"])
+    base_src = f"{run_json}; {README}"
+    slides = []
+
+    # ---------------- slide a ----------------
+    tot_rows, tot_donors = int(tab.n_rows.sum()), int(tab.n_donors.sum())
+    n_all = int(tab.n_pairs_all.sum())
+    n_fail, n_bord, n_low = int(tab.n_identity_fail.sum()), int(tab.n_identity_borderline.sum()), int(tab.n_low_cells.sum())
+    ji = tab.set_index("dataset").loc["ji2025_chili"]
+    if not ji.bulk.startswith("whole blood"):
+        raise ValueError("Ji bulk label changed")
+    noflag = [NAME[d] for d in tab[~tab.identity_flag].dataset]
+    multi = [NAME[d] for d in tab[tab.n_rows != tab.n_donors].dataset]
+    kw_a = check_keywords([f"{tot_rows} paired samples, {tot_donors} donors",
+                           "Only identity-pass pairs; donor-level splits",
+                           "Ji bulk: whole blood, not PBMC"])
+    foot_a = (f"Usable pair = passed the pairing identity check and >= {int(tab.min_cells.iloc[0])} cells "
+              f"(fold >= 0 in data/<cohort>/folds.csv); dropped {n_fail} identity-fail, {n_bord} borderline and "
+              f"{n_low} low-cell pairs of {n_all}. {', '.join(noflag)}: no identity flag, all used. "
+              f"{', '.join(multi)}: several pairs per donor. Case vs control = samples in "
+              f"scores/<cohort>/pooled/summary.json. Giroux 'moderate' = GEO 'severe' (paper term). "
+              f"Labels: data/<cohort>/QC_REPORT.md; ImMicro: GEO GSE314922/GSE314416.")
+    num("a", f"{tot_rows} paired samples (sum of usable rows)", f"{DATA}/<cohort>/fewshot_n5/DRAWS.json (n_usable_rows)")
+    num("a", f"{tot_donors} donors (sum of usable donors)", f"{DATA}/<cohort>/fewshot_n5/DRAWS.json (n_usable_donors)")
+    num("a", f"dropped {n_fail} identity-fail, {n_bord} borderline, {n_low} < {int(tab.min_cells.iloc[0])} cells, of {n_all}",
+        f"{DATA}/<cohort>/features.npz (identity_pass, identity_check); samples.csv (n_cells); {PREPARE} (MIN_CELLS)")
+    for r in tab.itertuples():
+        v = f"{NAME[r.dataset]} {r.n_rows} pairs / {r.n_donors} donors"
+        s = r.n_rows_source
+        if r.case_label:
+            v += f"; {r.case_label} {int(r.n_case)} vs {r.control_label} {int(r.n_control)}"
+            if r.n_neither:
+                v += f" (+{r.n_neither} {r.neither_label}, neither)"
+            s += f"; {r.contrast_source}"
+        num("a", v, s)
+    fs_a = check_foot_short(f"Usable pair = passed identity check and ≥ {int(tab.min_cells.iloc[0])} cells: "
+                            f"{tot_rows} of {n_all} pairs; donor-level splits.")
+    slides.append({"title": "Five paired cohorts", "layout": "full",
+                   "figures": [str(FIGDIR / "sec09_cohorts.png")], "keywords": kw_a, "footnote": foot_a,
+                   "footnote_short": fs_a, "numbers": NUMS["a"]})
+
+    # ---------------- slide b ----------------
+    mk_b = (("coverage", False), ("precision", False), ("comp_l1", True), ("coverage_dm", False),
+            ("precision_dm", False))
+    cmp_b = {mk: same_draw_compare(few, status, ORDER, mk, lower) for mk, lower in mk_b}
+    cmp_bz = {mk: same_draw_compare(few, status, ORDER, mk, lower, other="zs_raw") for mk, lower in mk_b}
+    wins = {mk: sum(v["verdict"] == "lora" for v in c.values()) for mk, c in cmp_b.items()}
+    wins_z = {mk: sum(v["verdict"] == "lora" for v in c.values()) for mk, c in cmp_bz.items()}
+    ncoh = {mk: len(c) for mk, c in cmp_b.items()}
+    if ncoh["coverage"] != len(ORDER):
+        raise ValueError("a cohort has no shared scSemiProfiler draw")
+    if any(len(c) != len(ORDER) for c in cmp_bz.values()):
+        raise ValueError("a cohort has no zero-shot draw")
+    nb = len(ORDER)
+    if wins["coverage"] == wins["precision"]:
+        kw_b1 = f"LoRA-5 > scSemi-5: coverage, precision {wins['coverage']}/{nb}"
+    else:
+        kw_b1 = f"Beats scSemi-5: coverage {wins['coverage']}/{nb}, precision {wins['precision']}/{nb}"
+
+    def span(v):
+        return f"{min(v):.2f}–{max(v):.2f}"
+
+    cov_mean = {a: [float(np.mean(few_vals(few, d, a, "coverage")[0])) for d in ORDER]
+                for a in ("lora", "scsemi5", "zs_raw")}
+    kw_b2 = f"LoRA-5 coverage {span(cov_mean['lora'])}; scSemi-5 {span(cov_mean['scsemi5'])}"
+    l1t = {ds: l1_vs_zeroshot_per_donor(ds) for ds in ORDER}
+    gain = [d for d in ORDER if l1t[d]["p"] < 0.05 and l1t[d]["mean_diff"] < 0]
+    others = [d for d in ORDER if d not in gain]
+    if not gain:
+        kw_b3 = "No composition gain vs zero-shot"
+    elif not others:
+        kw_b3 = "Lower L1 than zero-shot: all cohorts"
+    elif len(gain) >= len(others):
+        kw_b3 = keyword_or_short(f"Lower L1 than zero-shot: all but {names(others)}",
+                                 f"Lower L1 than zero-shot: {len(gain)}/{nb}")
+    else:
+        kw_b3 = keyword_or_short(f"Lower L1 than zero-shot: {names(gain)} only",
+                                 f"Lower L1 than zero-shot: {len(gain)}/{nb}")
+    kw_b = check_keywords([kw_b1, kw_b2, kw_b3])
+    l1_split = [d for d, v in cmp_b["comp_l1"].items() if v["verdict"] != "lora"]
+    dm_c = {a: [cmp_b["coverage_dm"][d][k] for d in ORDER] for a, k in (("lora", "lora"), ("scsemi5", "other"))}
+    if wins["coverage_dm"] != wins["coverage"] or wins["precision_dm"] != wins["precision"]:
+        raise ValueError("depth-matched ranking differs from raw P/C: rewrite the 'same ranking' footnote")
+    tr_all = [x for d in ORDER for x in info[d]["n_train"]]
+    missing = status[status.status != "used"]
+    miss_txt = ", ".join(f"{NAME[r.dataset]} draw{r.draw}" for r in missing.itertuples())
+    ns_used, ns_all = len(used), len(status)
+    zs_pc = (f"coverage {wins_z['coverage']}/{nb}, precision {wins_z['precision']}/{nb}, "
+             f"depth-matched {wins_z['coverage_dm']}/{nb} and {wins_z['precision_dm']}/{nb}")
+    foot_b = (f"Base f01 6k·1x·E16 (EMA, DDIM 50) + LoRA r={lora_r}, {steps:,} steps. LoRA-5 / scSemi-5 = LoRA / "
+              f"scSemiProfiler trained on the same 5 donors ({min(tr_all)}–{max(tr_all)} pairs); test = all other usable "
+              f"pairs. Zero-shot = base model without fine-tuning, fed each test pair's real bulk as measured (same "
+              f"call as LoRA-5). scSemi-5 scored in a separate call, same test set and reference (sha-checked); "
+              f"{ns_used}/{ns_all} draws" + (f" (missing: {miss_txt})" if len(missing) else "") + ". "
+              f"Verdict = better in >= 80% of shared draws. LoRA-5 lower L1 than scSemi-5 in {wins['comp_l1']}/{nb}"
+              + (f" ({', '.join(f'{NAME[d]} {cmp_b['comp_l1'][d]['verdict']} {cmp_b['comp_l1'][d]['n_lora']}/{cmp_b['comp_l1'][d]['n']}' for d in l1_split)})"
+                 if l1_split else "")
+              + f". LoRA-5 > zero-shot: {zs_pc} cohorts. Depth-matched P/C vs scSemi-5: same ranking, coverage "
+              f"LoRA-5 {span(dm_c['lora'])} vs scSemi-5 {span(dm_c['scsemi5'])}. L1 LoRA-5 minus zero-shot, per-donor "
+              f"Wilcoxon (gain = p < 0.05): " + ", ".join(f"{NAME[d]} {l1t[d]['mean_diff']:+.2f} (p {fmt_p(l1t[d]['p'])})"
+                                                         for d in ORDER)
+              + ". scANVI labels; P/C = PCA-50, k=10 balls. Scores: scores/<cohort>/fewshot_n5/.")
+    nd_b = sorted({info[d]["n_draws"] for d in ORDER})
+    fs_b = check_foot_short(f"{rng(nd_b)} draws per cohort; test = all other usable pairs. Zero-shot = base model "
+                            f"on real bulk. scANVI labels.")
+    scs = f"{SCORES}/<cohort>/fewshot_n5/draw<j>/summary.json and draw<j>_scsemi5/summary.json"
+    for mk in ("coverage", "precision", "comp_l1", "coverage_dm", "precision_dm"):
+        c = cmp_b[mk]
+        num("b", f"LoRA-5 vs scSemi-5 {mk} (shared-draw means; LoRA-better draws/n; verdict): " +
+            "; ".join(f"{NAME[d]} {v['lora']:.3f} vs {v['other']:.3f} ({v['n_lora']}/{v['n']}, {v['verdict']})"
+                      for d, v in c.items()) + f" -> LoRA verdicts {wins[mk]}/{ncoh[mk]}", scs)
+    for mk in ("coverage", "precision", "comp_l1", "coverage_dm", "precision_dm"):
+        c = cmp_bz[mk]
+        num("b", f"LoRA-5 vs zero-shot (real bulk) {mk} (same-draw means; LoRA-better draws/n; verdict): " +
+            "; ".join(f"{NAME[d]} {v['lora']:.3f} vs {v['other']:.3f} ({v['n_lora']}/{v['n']}, {v['verdict']})"
+                      for d, v in c.items()) + f" -> LoRA verdicts {wins_z[mk]}/{len(c)}",
+            f"{SCORES}/<cohort>/fewshot_n5/draw<j>/summary.json (arms lora5, zs_raw)")
+    num("b", "coverage cohort means over all draws: " + "; ".join(
+        f"{lab} " + ", ".join(f"{NAME[d]} {v:.3f}" for d, v in zip(ORDER, cov_mean[a]))
+        for a, lab in (("lora", "LoRA-5"), ("scsemi5", "scSemi-5"), ("zs_raw", "zero-shot"))), scs)
+    num("b", "LoRA-5 minus zero-shot (real bulk) composition L1, per-donor Wilcoxon (draw-averaged): " +
+        "; ".join(f"{NAME[d]} {v['mean_diff']:+.3f}, n={v['n_donors']}, p={v['p']:.2g}" for d, v in l1t.items()) +
+        f" -> gain in {len(gain)}/{nb}",
+        f"{SCORES}/<cohort>/fewshot_n5/draw<j>/per_donor.csv; {DATA}/<cohort>/folds.csv")
+    num("b", "scSemiProfiler draws used / LoRA draws: " +
+        ", ".join(f"{NAME[d]} {int((used.dataset == d).sum())}/{n_draws(few, d, 'lora')}" for d in ORDER),
+        OUTDATA / "sec09_scsemi_status.csv")
+    num("b", "training pairs per draw: " + ", ".join(f"{NAME[d]} {info[d]['train']}" for d in ORDER) +
+        "; test pairs: " + ", ".join(f"{NAME[d]} {info[d]['test']}" for d in ORDER),
+        f"{DATA}/<cohort>/fewshot_n5/DRAWS.json (n_train_rows, n_test_rows)")
+    num("b", f"LoRA r = {lora_r}; {steps} fixed steps; base {CKPT}; sampling: {sampling}", base_src)
+    num("b", f"P/C definition: {pc_def}; PCA components in every summary.json = 50", SCORE_FT)
+    for ds in ORDER:
+        extra("b", f"{ds} LoRA-5 minus zero-shot (real bulk) L1, Wilcoxon over donor-draw rows (pseudo-replicated; "
+                   f"not quoted, replaced by the per-donor test)",
+              f"{l1t[ds]['row_mean_diff']:+.3f}, n={l1t[ds]['n_rows']}, p {l1t[ds]['row_p']:.2g}", l1t[ds]["source"])
+    slides.append({"title": "5 donors: LoRA vs scSemiProfiler — cells", "layout": "full",
+                   "figures": [str(FIGDIR / "sec09_fewpair_cells.png")], "keywords": kw_b, "footnote": foot_b,
+                   "footnote_short": fs_b, "numbers": NUMS["b"]})
+
+    # ---------------- slide c ----------------
+    cmp_r = same_draw_compare(few, status, DISEASE, "contrast_r")
+    cmp_d = same_draw_compare(few, status, DISEASE, "deg", min_margin=1.0)
+    cmp_rz = same_draw_compare(few, status, DISEASE, "contrast_r", other="zs_raw")
+    cmp_dz = same_draw_compare(few, status, DISEASE, "deg", min_margin=1.0, other="zs_raw")
+    deg_l = [d for d, v in cmp_d.items() if v["verdict"] == "lora"]
+    deg_tie = [d for d, v in cmp_d.items() if v["verdict"] == "tie"]
+    deg_other = [d for d, v in cmp_d.items() if v["verdict"] not in ("lora", "tie")]
+    near0 = [d for d in deg_tie if max(cmp_d[d]["lora"], cmp_d[d]["other"]) < 1.0]
+    kw_c1 = f"Top-50 DEG: LoRA-5 > scSemi-5 {len(deg_l)}/{len(cmp_d)}"
+    kw_c1_short = kw_c1 + (f", tied {len(deg_tie)}" if deg_tie else "")
+    if near0:
+        kw_c1 += f"; {names(near0)} ~0"
+    if [d for d in deg_tie if d not in near0]:
+        kw_c1 += f"; {names([d for d in deg_tie if d not in near0])} tied"
+    if deg_other:
+        kw_c1 += f"; {names(deg_other)} not"
+    kw_c1 = keyword_or_short(kw_c1, kw_c1_short)
+    r_s = [d for d, v in cmp_r.items() if v["verdict"] == "other"]
+    r_split = [d for d, v in cmp_r.items() if v["verdict"] in ("split", "tie")]
+    r_l = [d for d, v in cmp_r.items() if v["verdict"] == "lora"]
+    parts = []
+    if r_s:
+        parts.append(f"scSemi-5 wins {names(r_s)}")
+    if r_split:
+        parts.append(f"{names(r_split)} split")
+    kw_c2_short = f"logFC r: LoRA-5 > scSemi-5 {len(r_l)}/{len(cmp_r)}"
+    kw_c2 = keyword_or_short(("logFC r: " + "; ".join(parts)) if parts else kw_c2_short, kw_c2_short)
+    arms3 = ("zs_raw", "scsemi5", "lora")
+
+    def arm_mean(ds, a, mk):
+        v = [x for x in few_vals(few, ds, a, mk)[0] if not np.isnan(x)]
+        return float(np.mean(v)) if v else np.nan
+
+    bc_all = [d for d in DISEASE
+              if few_ref(few, d, "broadcast", "contrast_r")[0] > np.nanmax([arm_mean(d, a, "contrast_r") for a in arms3])
+              and few_ref(few, d, "broadcast", "deg")[0] > np.nanmax([arm_mean(d, a, "deg") for a in arms3])]
+    kw_c3 = (keyword_or_short(f"{names(bc_all)}: real-bulk broadcast beats all arms",
+                              f"Broadcast beats all arms: {len(bc_all)}/{len(DISEASE)}")
+             if bc_all else "Every cohort: some arm beats broadcast")
+    kw_c = check_keywords([kw_c1, kw_c2, kw_c3])
+    dmax = {d: sorted({int(x) for x in few[(few.dataset == d) & (few.arm == "lora")]["deg_max"].dropna()}) for d in DISEASE}
+    small = [d for d in DISEASE if subjects_text(info[d])]
+
+    def vz(c):
+        return ", ".join(f"{NAME[d]} {v['verdict'].replace('other', 'zero-shot')} {v['n_lora']}/{v['n']}"
+                         for d, v in c.items())
+
+    foot_c = (f"Same draws and verdict rule as the previous slide; DEG tie = half the draws equal or mean gap < 1 gene. "
+              f"logFC r = mean over T/B/Mono/NK of case-vs-control log2FC r (generated vs observed); top-50 DEG = "
+              f"Welch+BH overlap summed over {n_types} types (max 250; "
+              + ", ".join(f"{NAME[d]} {rng(dmax[d])}" for d in DISEASE if dmax[d] != [250]) + " per draw). "
+              + "".join(f"{NAME[d]} test = {subjects_text(info[d])} subjects per draw. " for d in small)
+              + f"LoRA-5 vs zero-shot (real bulk), LoRA-better draws: logFC r {vz(cmp_rz)}; DEG {vz(cmp_dz)}. "
+              + "Broadcast = the input real bulk's own logFC. Scores: scores/<cohort>/fewshot_n5/draw<j>{,_scsemi5}/.")
+    fs_c = check_foot_short(f"Same draws as previous slide. Top-50 DEG = Welch + BH overlap, summed over {n_types} "
+                            f"cell types.")
+    srcc = f"{SCORES}/<cohort>/fewshot_n5/draw<j>{{,_scsemi5}}/summary.json and contrast.csv"
+    num("c", "Top-50 DEG LoRA-5 vs scSemi-5 (shared-draw means; LoRA-better / tied draws; verdict): " +
+        "; ".join(f"{NAME[d]} {v['lora']:.1f} vs {v['other']:.1f} ({v['n_lora']}/{v['n']}, {v['n_tie']} tied, {v['verdict']})"
+                  for d, v in cmp_d.items()), srcc)
+    num("c", "logFC r LoRA-5 vs scSemi-5 (shared-draw means; LoRA-better draws; verdict): " +
+        "; ".join(f"{NAME[d]} {v['lora']:.3f} vs {v['other']:.3f} ({v['n_lora']}/{v['n']}, {v['verdict']})"
+                  for d, v in cmp_r.items()), srcc)
+    num("c", "Top-50 DEG LoRA-5 vs zero-shot (real bulk) (same-draw means; LoRA-better / tied draws; verdict): " +
+        "; ".join(f"{NAME[d]} {v['lora']:.1f} vs {v['other']:.1f} ({v['n_lora']}/{v['n']}, {v['n_tie']} tied, {v['verdict']})"
+                  for d, v in cmp_dz.items()), f"{SCORES}/<cohort>/fewshot_n5/draw<j>/summary.json (arms lora5, zs_raw)")
+    num("c", "logFC r LoRA-5 vs zero-shot (real bulk) (same-draw means; LoRA-better draws; verdict): " +
+        "; ".join(f"{NAME[d]} {v['lora']:.3f} vs {v['other']:.3f} ({v['n_lora']}/{v['n']}, {v['verdict']})"
+                  for d, v in cmp_rz.items()), f"{SCORES}/<cohort>/fewshot_n5/draw<j>/contrast.csv (arms lora5, zs_raw)")
+    for d in bc_all:
+        num("c", f"{NAME[d]} broadcast logFC r {few_ref(few, d, 'broadcast', 'contrast_r')[0]:.3f} vs best arm "
+                 f"{max(arm_mean(d, a, 'contrast_r') for a in arms3):.3f}; broadcast DEG "
+                 f"{few_ref(few, d, 'broadcast', 'deg')[0]:.1f} vs best arm {max(arm_mean(d, a, 'deg') for a in arms3):.1f} "
+                 f"(means over all draws)", f"{SCORES}/{d}/fewshot_n5/draw<j>{{,_scsemi5}}/summary.json")
+    num("c", "DEG maximum per draw: " + ", ".join(f"{NAME[d]} {rng(dmax[d])}" for d in DISEASE),
+        f"{SCORES}/<cohort>/fewshot_n5/draw<j>/summary.json (deg_top50_max)")
+    for d in small:
+        num("c", f"{NAME[d]} test subjects per draw: {subjects_text(info[d])} ({CONTRAST[d]['case']} vs "
+                 f"{CONTRAST[d]['control']}); test pairs {info[d]['test']}",
+            f"{info[d]['source']} (train_donors); {DATA / d / 'folds.csv'} (group)")
+    slides.append({"title": "5 donors: LoRA vs scSemiProfiler — disease signal", "layout": "figure_keywords",
+                   "figures": [str(FIGDIR / "sec09_fewpair_disease.png")], "keywords": kw_c, "footnote": foot_c,
+                   "footnote_short": fs_c, "numbers": NUMS["c"]})
+
+    # ---------------- slide d ----------------
+    L = {d: pooled[d][POOLED_ARMS["lora"]] for d in ORDER}
+    Z = {d: pooled[d][POOLED_ARMS["zs_raw"]] for d in ORDER}
+    S = {d: pooled[d][POOLED_ARMS["sconly"]] for d in ORDER}
+    B = {d: [Z[d], S[d]] for d in ORDER}
+    best = {mk: [d for d in (ORDER if mk == "coverage" else DISEASE) if L[d][mk] >= max(b[mk] for b in B[d])]
+            for mk in ("coverage", "contrast_r", "deg")}
+    kw_d1 = f"LoRA tops coverage {len(best['coverage'])}/{len(ORDER)}, logFC r {len(best['contrast_r'])}/{len(DISEASE)}"
+    jb = baselines_text([b["coverage"] for b in B["ji2025_chili"]])
+    jl = f"{L['ji2025_chili']['coverage']:.2f}"
+    kw_d2 = keyword_or_short(f"Ji whole blood: baselines {jb}, LoRA {jl}", f"Ji whole blood: LoRA {jl}, others {jb}")
+    bc_win = [d for d in DISEASE if L[d]["broadcast_r"] > L[d]["contrast_r"]]
+    kw_d3 = ("; ".join(f"{NAME[d]}: broadcast r {L[d]['broadcast_r']:.2f} > LoRA {L[d]['contrast_r']:.2f}" for d in bc_win)
+             if bc_win else "LoRA above broadcast in every cohort")
+    kw_d3 = keyword_or_short(kw_d3, f"Broadcast r above LoRA: {names(bc_win)}")
+    kw_d = check_keywords([kw_d1, kw_d2, kw_d3])
+    foot_d = (f"Donor-level 5-fold CV; each pair scored once, by the fold model that never trained on its donor "
+              f"(one pooled call per cohort). LoRA r={lora_r} on all training donors' real bulk, early stop on "
+              f"inner-val donors. Zero-shot = base model without fine-tuning; sc-only = full fine-tune on training "
+              f"donors' sc pseudobulk; both are fed the test pair's real bulk as measured. Base f01 6k·1x·E16. "
+              f"Top-50 DEG: LoRA best {len(best['deg'])}/{len(DISEASE)}. Sepsis DEG from pooled_deg. "
+              f"Scores: scores/<cohort>/pooled/summary.json.")
+    fs_d = check_foot_short("Donor-level 5-fold CV; each pair scored by a model that never saw its donor. "
+                            "sc-only: tuned on sc pseudobulk.")
+    psrc = f"{SCORES}/<cohort>/pooled/summary.json (Sepsis DEG: {SCORES}/ye2026_sepsis/pooled_deg/summary.json)"
+    num("d", f"LoRA best of 3 arms: coverage {len(best['coverage'])}/{len(ORDER)}, logFC r "
+             f"{len(best['contrast_r'])}/{len(DISEASE)}, top-50 DEG {len(best['deg'])}/{len(DISEASE)}", psrc)
+    num("d", f"Ji coverage: zero-shot {Z['ji2025_chili']['coverage']:.3f}, sc-only "
+             f"{S['ji2025_chili']['coverage']:.3f}, LoRA {L['ji2025_chili']['coverage']:.3f}",
+        SCORES / "ji2025_chili" / "pooled" / "summary.json")
+    for d in bc_win:
+        num("d", f"{NAME[d]} logFC r: broadcast {L[d]['broadcast_r']:.3f} vs LoRA {L[d]['contrast_r']:.3f}; "
+                 f"DEG broadcast {L[d]['deg_broadcast']:.0f} vs LoRA {L[d]['deg']:.0f}", SCORES / d / "pooled" / "summary.json")
+    num("d", "LoRA logFC r vs broadcast: " + ", ".join(f"{NAME[d]} {L[d]['contrast_r']:.3f} vs {L[d]['broadcast_r']:.3f}"
+                                                         for d in DISEASE), psrc)
+    for mk, lab in (("coverage", "coverage"), ("contrast_r", "logFC r"), ("deg", "top-50 DEG")):
+        dd = ORDER if mk == "coverage" else DISEASE
+        f = "{:.3f}" if mk != "deg" else "{:.0f}"
+        num("d", f"{lab} LoRA / zero-shot / sc-only: " +
+            ", ".join(f"{NAME[d]} {f.format(L[d][mk])} / {f.format(Z[d][mk])} / {f.format(S[d][mk])}" for d in dd), psrc)
+    num("d", "Pooled n (pairs / donors): " + ", ".join(f"{NAME[r.dataset]} {r.n_rows}/{r.n_donors}" for r in tab.itertuples()),
+        f"{SCORES}/<cohort>/pooled/summary.json (n_donors); {DATA}/<cohort>/fewshot_n5/DRAWS.json")
+    num("d", f"early stopping: {early}", README)
+    for d in ORDER:
+        for a, f in POOLED_ARMS.items():
+            extra("d", f"{d} pooled {a} ({f}) coverage / precision / L1 / r / DEG",
+                  f"{pooled[d][f]['coverage']:.3f} / {pooled[d][f]['precision']:.3f} / {pooled[d][f]['comp_l1']:.3f} / "
+                  f"{pooled[d][f]['contrast_r']:.3f} / {pooled[d][f]['deg']:.0f}",
+                  pooled[d][f]["source"] + (f"; DEG: {pooled[d][f]['deg_source']}" if "deg_source" in pooled[d][f] else ""))
+    slides.append({"title": "All pairs, 5-fold", "layout": "full", "figures": [str(FIGDIR / "sec09_fivefold.png")],
+                   "keywords": kw_d, "footnote": foot_d, "footnote_short": fs_d, "numbers": NUMS["d"]})
+
+    for s in slides:
+        for k in ("title", "footnote"):
+            check_text(s[k], k)
+        for n in s["numbers"]:
+            check_text(n["value"], "quoted number")
+    spec = {"section": "09", "slides": slides, "scripts": [str(SCRIPT)],
+            "generated_from": "tools/sec09_paired_validation.py (rerun to refresh; do not edit by hand)",
+            "scsemi_draws_used": f"{ns_used}/{ns_all}"}
+    (OUTDATA / "sec09_slide_text.json").write_text(json.dumps(spec, indent=1, ensure_ascii=False))
+    flat = [{"slide": k, "item": "quoted on slide", "value": n["value"], "source": n["source"]}
+            for k in "abcd" for n in NUMS[k]] + EXTRA
+    pd.DataFrame(flat).to_csv(OUTDATA / "sec09_quoted_numbers.csv", index=False)
+    # nothing written by this script may carry the input-adjustment wording
+    for p in [OUTDATA / "sec09_slide_text.json"] + sorted(OUTDATA.glob("sec09_*.csv")):
+        m = OFF_STORY.search(p.read_text(encoding="utf-8"))
+        if m:
+            raise ValueError(f"{p}: off-story text {m.group(0)!r}")
+
+    print("scSemiProfiler draws used per dataset:",
+          {NAME[d]: f"{int((used.dataset == d).sum())}/{n_draws(few, d, 'lora')}" for d in ORDER})
+    for s in slides:
+        print(f"\n== {s['title']} [{s['layout']}]")
+        for k in s["keywords"]:
+            print("  *", k, f"({n_words(k)} words, {len(k)} chars)")
+        print("  footnote_short:", s["footnote_short"], f"({len(s['footnote_short'])} chars)")
+        print("  footnote:", s["footnote"])
+        for n in s["numbers"]:
+            print("  #", n["value"])
+    for p in sorted(FIGDIR.glob("sec09_*.png")):
+        print("wrote", p)
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

paper-releases/pbisc-diffusion/updates/20260930T090852Z-summary-deck-final-path-story/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/tools/style.py

@@ -0,0 +1,65 @@
+"""Shared matplotlib style for every figure of the 2026-09-30 paper-summary deck.
+
+Figures are drawn at their on-slide size (inches), so font sizes are the sizes seen on the slide.
+Slide: 13.333 x 7.5 in (16:9). Content area below the title: about 12.3 x 5.7 in.
+Use FULL, HALF or TALL for figsize unless a figure needs something else.
+Plain look: white background, no top/right spines, no titles inside figures (the slide title says it),
+no shadows, no gradients, no decorative elements. Colour carries meaning only.
+"""
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+FULL = (12.3, 5.6)     # figure fills the content area
+HALF = (6.0, 5.4)      # one of two side-by-side figures
+TALL = (7.8, 5.6)      # figure + a keyword column on the right
+DPI = 200
+
+FS = 14                # base font size on the slide (pt)
+plt.rcParams.update({
+    "font.family": "sans-serif",
+    "font.sans-serif": ["Arial", "Liberation Sans", "DejaVu Sans"],
+    "font.size": FS, "axes.titlesize": FS, "axes.labelsize": FS,
+    "xtick.labelsize": FS - 2, "ytick.labelsize": FS - 2, "legend.fontsize": FS - 2,
+    "axes.spines.top": False, "axes.spines.right": False,
+    "axes.linewidth": 1.0, "xtick.major.width": 1.0, "ytick.major.width": 1.0,
+    "legend.frameon": False, "figure.facecolor": "white", "axes.facecolor": "white",
+    "savefig.facecolor": "white", "savefig.bbox": "tight", "savefig.pad_inches": 0.05,
+    "pdf.fonttype": 42, "ps.fonttype": 42,
+})
+
+# One colour per method, identical across every figure (Okabe-Ito, same as the paper's paired figures)
+ARM = {
+    "pbisc": "#D55E00",          # PBISC (final model, or LoRA fine-tune)
+    "lora": "#D55E00",
+    "full_ft": "#E07B39",
+    "zs_cal": "#0072B2",          # zero-shot, paired calibration
+    "zs_raw": "#A6CBE3",          # zero-shot, raw bulk (old pale shade; kept for the archived scripts)
+    "zs": "#0072B2",              # zero-shot, real bulk as measured: the no-fine-tune baseline (story without calibration)
+    "pseudo": "#1A1A1A",          # zero-shot from sc pseudobulk (reference line)
+    "scsemi": "#E69F00",          # scSemiProfiler
+    "vae": "#56B4E9",             # PBISC-VAE
+    "sconly": "#CC79A7",          # sc-only fine-tune
+    "broadcast": "#9A9A9A",       # real-bulk broadcast baseline
+    "other": "#009E73",
+    "refcal": "#D55E00",          # zero-shot + reference calibration (PBISC's no-pair real-bulk path; = sec04 PBISC colour)
+    "tpm": "#DDCC77",             # zero-shot, TPM (length-corrected) bulk input (sec05; not used by any other key)
+}
+# Architecture variants (section 3 only). trunk4x / prop avoid the PBISC-VAE (#56B4E9) and
+# scSemiProfiler (#E69F00) reference-line colours drawn on the same slide.
+VARIANT = {"1x": "#D55E00", "4x": "#0072B2", "trunk4x": "#332288", "ctx4x": "#009E73", "16x": "#CC79A7", "prop": "#999933"}
+# Gene panels (section 3 only); 3k avoids the PBISC-VAE line colour
+PANEL = {"2k": "#A6CBE3", "3k": "#882255", "6k": "#D55E00", "15k": "#0072B2"}
+GREY = "#8C8C8C"
+INK = "#1A1A1A"
+SUBINK = "#4D4D4D"             # secondary text (sub-lines under bars)
+PBISC_PALE = "#F3C9A8"         # light PBISC orange: planned / not yet generated (outlined in ARM["pbisc"])
+
+
+def save(fig, name, figdir):
+    """Save PNG (slide) and PDF (vector) with the same stem."""
+    from pathlib import Path
+    figdir = Path(figdir)
+    fig.savefig(figdir / f"{name}.png", dpi=DPI)
+    fig.savefig(figdir / f"{name}.pdf")
+    plt.close(fig)