Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261013T023822Z-pbisc-yellow-scsemi-blue/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/method_palette_20261013/yellow_revision/before_palette.json

@@ -0,0 +1,306 @@
+{
+  "schema": "pbisc.method-palette.v1",
+  "user_directive": "PBISC red family, scSemiProfiler blue family throughout current results review",
+  "primary": {
+    "PBISC": "#B2182B",
+    "scSemiProfiler": "#2166AC"
+  },
+  "rules": {
+    "fig1b_overview_nature": {
+      "mapping": {
+        "#5B6E8F": "#B2182B",
+        "#E6EAF1": "#F8E8EB",
+        "#A5AFC1": "#D5808A",
+        "#8E6C9E": "#9E1B32"
+      },
+      "reason": "PBISC model blocks, model arrows and optional adaptation only; preserve expression, detection, bulk and cell-type encodings.",
+      "source_pdf": "/home/sj_server_1/PBISC/paper-release-20260928/manuscript/figures/fig01_architecture_nature.pdf"
+    },
+    "sec02_bulk_conditioning": {
+      "mapping": {},
+      "reason": "Orange denotes bulk-conditioning/context, not a method comparison; preserve semantic diagram colors."
+    },
+    "fig1d_corpus": {
+      "mapping": {},
+      "reason": "Cell types and cohort composition, not method colors."
+    },
+    "review_style_20261013_ablation": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#009E73": "#E57F88",
+        "#332288": "#7F0000",
+        "#0072B2": "#D96874",
+        "#A6CBE3": "#F1A6AE",
+        "#882255": "#8F2638"
+      },
+      "reason": "PBISC architecture and gene-panel variants; black real-R2 reference retained."
+    },
+    "sec03_panel": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#0072B2": "#D96874",
+        "#A6CBE3": "#F1A6AE",
+        "#882255": "#8F2638"
+      },
+      "reason": "PBISC gene panels in red shades; scSemi reference line blue."
+    },
+    "review_style_20261013_rank": {
+      "mapping": {
+        "#E07B39": "#7F0000",
+        "#D55E00": "#B2182B"
+      },
+      "reason": "Full fine-tuning dark red and LoRA red; rank markers preserve their shapes."
+    },
+    "fig3_healthy_pairs": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874"
+      },
+      "reason": "PBISC and its raw/calibrated inputs in red shades; scSemi blue. Neutral reference/errorbar styles retained."
+    },
+    "fig3_healthy_fewpair": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874"
+      },
+      "reason": "PBISC and its raw/calibrated inputs in red shades; scSemi blue. Neutral reference/errorbar styles retained."
+    },
+    "fig3_healthy_copy": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874"
+      },
+      "reason": "PBISC and its raw/calibrated inputs in red shades; scSemi blue. Neutral reference/errorbar styles retained."
+    },
+    "fig2_healthy_pairs": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874"
+      },
+      "reason": "PBISC and its raw/calibrated inputs in red shades; scSemi blue. Neutral reference/errorbar styles retained."
+    },
+    "fig4_disease_fewpair": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874"
+      },
+      "reason": "PBISC and its raw/calibrated inputs in red shades; scSemi blue. Neutral reference/errorbar styles retained."
+    },
+    "sec09_fewpair_cells": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874"
+      },
+      "reason": "PBISC and its raw/calibrated inputs in red shades; scSemi blue. Neutral reference/errorbar styles retained."
+    },
+    "sec09_extra_ksens": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874"
+      },
+      "reason": "PBISC and its raw/calibrated inputs in red shades; scSemi blue. Neutral reference/errorbar styles retained."
+    },
+    "sec09_same_pairs": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874"
+      },
+      "reason": "PBISC and its raw/calibrated inputs in red shades; scSemi blue. Neutral reference/errorbar styles retained."
+    },
+    "sec04_sle_benchmark": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874"
+      },
+      "reason": "PBISC and its raw/calibrated inputs in red shades; scSemi blue. Neutral reference/errorbar styles retained."
+    },
+    "review_style_20261013_healthy_composition": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#0072B2": "#2166AC"
+      },
+      "reason": "Two-method paired comparison; this asset already used blue for scSemi."
+    },
+    "sec05_condition_space": {
+      "mapping": {},
+      "reason": "Input-space diagnostic: measured/shift-corrected inputs, no method comparison."
+    },
+    "sec09_time_to_apply": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#F2B48C": "#F1B4BC",
+        "#F5D08A": "#67A9CF"
+      },
+      "reason": "PBISC adaptation/generation red shades; scSemi training/inference blue shades."
+    },
+    "sec09_cohorts": {
+      "mapping": {},
+      "reason": "Cohort metadata table, no method colors."
+    },
+    "fig4_disease_fivefold": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874",
+        "#CC79A7": "#8F4666"
+      },
+      "reason": "PBISC LoRA, zero-shot/calibrated and sc-only fine-tuning red family; scSemi blue."
+    },
+    "sec09_cat_curves": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874",
+        "#CC79A7": "#8F4666"
+      },
+      "reason": "PBISC LoRA, zero-shot/calibrated and sc-only fine-tuning red family; scSemi blue."
+    },
+    "secS_deconv_disease": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#CC79A7": "#6A51A3"
+      },
+      "reason": "PBISC red, scSemi blue where present; CIBERSORTx purple and BayesPrism green."
+    },
+    "secS_deconv_composition": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#CC79A7": "#6A51A3"
+      },
+      "reason": "PBISC red, scSemi blue where present; CIBERSORTx purple and BayesPrism green."
+    },
+    "review_style_20261013_rare_observed": {
+      "mapping": {},
+      "reason": "Observed-frequency heatmap encodes abundance, not a method."
+    },
+    "review_style_20261013_rare_corrected_comparison": {
+      "mapping": {},
+      "reason": "Points encode cohorts; x/y axes explicitly name methods. Preserve cohort color legend."
+    },
+    "review_style_20261013_behcet": {
+      "mapping": {
+        "#0072B2": "#D96874",
+        "#D55E00": "#B2182B"
+      },
+      "reason": "PBISC raw/calibrated and LoRA predictions in red family; observed native/depth-matched references black."
+    },
+    "sec05_cells_umap": {
+      "mapping": {},
+      "reason": "UMAP points and legend encode cell types. Preserve images and biological labels."
+    },
+    "sec05_cells_metrics": {
+      "mapping": {
+        "#A6CBE3": "#F1A6AE",
+        "#5E8FB5": "#D96874"
+      },
+      "reason": "Raw-bulk and pseudobulk-input PBISC reconstructions red; observed real cells grey. Scope black replacement to four verified rectangles only.",
+      "rectangle_rules": [
+        {
+          "rect": [
+            66.389,
+            247.078,
+            91.925,
+            361.296
+          ],
+          "old": "#1A1A1A",
+          "new": "#B2182B"
+        },
+        {
+          "rect": [
+            207.736,
+            98.028,
+            233.272,
+            361.296
+          ],
+          "old": "#1A1A1A",
+          "new": "#B2182B"
+        },
+        {
+          "rect": [
+            349.082,
+            217.017,
+            374.618,
+            361.296
+          ],
+          "old": "#1A1A1A",
+          "new": "#B2182B"
+        },
+        {
+          "rect": [
+            172.368,
+            8.734,
+            185.568,
+            17.134
+          ],
+          "old": "#1A1A1A",
+          "new": "#B2182B"
+        }
+      ]
+    },
+    "sec08_cross_cohort": {
+      "mapping": {
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874",
+        "#44AA99": "#E57F88",
+        "#117733": "#8F2638",
+        "#332288": "#5F0F24",
+        "#D55E00": "#B2182B"
+      },
+      "reason": "All PBISC adaptation/cohort variants in distinct red shades; existing line styles preserved."
+    },
+    "sec08_cross_cohort_steps": {
+      "mapping": {
+        "#A6CBE3": "#F1A6AE",
+        "#0072B2": "#D96874",
+        "#44AA99": "#E57F88",
+        "#117733": "#8F2638",
+        "#332288": "#5F0F24",
+        "#D55E00": "#B2182B"
+      },
+      "reason": "All PBISC adaptation/cohort variants in distinct red shades; existing line styles preserved."
+    },
+    "sec07_no_pretraining": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#882255": "#7F0000",
+        "#0072B2": "#D96874"
+      },
+      "reason": "Pretrained LoRA red, same-architecture scratch ablation dark red, zero-shot rose."
+    },
+    "secA_scsemi_readout": {
+      "mapping": {
+        "#D55E00": "#B2182B",
+        "#E69F00": "#2166AC",
+        "#F5D08A": "#67A9CF"
+      },
+      "reason": "PBISC red; official scSemi blue and modified readout pale blue with original hatch."
+    }
+  },
+  "reference_policy": "Biological/cohort colors and neutral observed/oracle diagnostic references are separate encodings, not model identity.",
+  "unused_fill_policy": "Matplotlib #1F77B4 fill on neutral errorbar caps is not painted; preserve.",
+  "preserve_geometry_text_data": true
+}

paper-releases/pbisc-diffusion/updates/20261013T023822Z-pbisc-yellow-scsemi-blue/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/data/method_palette_20261013/yellow_revision/before_results_review_style_plots.py

@@ -0,0 +1,303 @@
+"""Restyle three saved-table review figures with the original deck's visual style.
+
+CPU plotting only. Healthy composition uses paired absolute errors before averaging held-out draws.
+Rare-state estimates are preserved. Only saved CSVs are read; no new experiment.
+build() returns ordered PDF paths, all export paths and source provenance.
+"""
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+from pathlib import Path
+
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+
+import numpy as np
+import pandas as pd
+import style as ST
+import matplotlib.pyplot as plt
+from matplotlib.lines import Line2D
+from matplotlib.patches import Patch
+from matplotlib.ticker import MaxNLocator
+
+D = Path(__file__).resolve().parents[1]
+M = D.parents[1]
+P = M / "plans/paper_figures_strictpairs_20261012"
+O = M / "plans/paired_finetune_intradataset_20260927"
+FIG = D / "figures"
+STEM = "review_style_20261013_"
+
+TYPES = ["T cell", "B cell", "Monocyte", "NK cell", "DC", "Other"]
+# These are cell-type colours from the saved-table review, not method colours.
+TYPE_COLORS = ["#56B4E9", "#E69F00", "#009E73", "#CC79A7", "#0072B2", "#B1B7BE"]
+COHORTS = ["ye2026_sepsis", "imperial_eskd", "giroux2022", "ji2025_chili"]
+COHORT_NAMES = ["Sepsis", "Imperial", "Giroux", "Ji"]
+# Identical to the original deck's tools/fig5_rare_states.py COH mapping.
+COHORT_COLORS = {"ye2026_sepsis": "#CC79A7", "imperial_eskd": "#0072B2",
+                 "giroux2022": "#009E73", "ji2025_chili": "#E69F00"}
+STATE_LABELS = {
+    "C1Q-high monocyte (reference)": "C1Q-high monocyte",
+    "Cycling activated T (CD38/HLA-DR)": "Cycling activated T",
+    "Cycling monocyte (Ye 2026)": "Cycling monocyte",
+    "ISG-high monocyte (positive control)": "ISG-high monocyte (+ control)",
+    "Platelet signal in monocytes (negative control)": "Platelet signal in mono (− control)",
+}
+
+
+def _source(path: Path, role: str, inputs: dict[str, dict]) -> Path:
+    path = path.resolve()
+    with path.open("rb") as stream:
+        digest = hashlib.file_digest(stream, "sha256").hexdigest()
+    inputs[str(path)] = {"path": str(path), "sha256": digest, "role": role}
+    return path
+
+
+def _save(fig, name: str, assets: list[str]) -> str:
+    """Use the deck's on-slide sizes/fonts; export the same artwork in three formats."""
+    for ext in ("png", "pdf", "svg"):
+        path = FIG / f"{STEM}{name}.{ext}"
+        fig.savefig(path, dpi=ST.DPI, facecolor="white", bbox_inches="tight", pad_inches=0.05)
+        assets.append(str(path))
+    plt.close(fig)
+    return str(FIG / f"{STEM}{name}.pdf")
+
+
+def _healthy(inputs: dict[str, dict], assets: list[str]) -> tuple[str, dict]:
+    frames = []
+    for draw in range(5):
+        for folder in (f"draw{draw}", f"draw{draw}_scsemi5"):
+            path = _source(O / "scores/immicro/fewshot_n5" / folder / "per_donor.csv",
+                           "existing healthy per-donor scores", inputs)
+            frame = pd.read_csv(path)
+            frame["draw"] = draw
+            frames.append(frame)
+    table = pd.concat(frames, ignore_index=True)
+    order = ["zs_raw", "zs_cal", "scsemi5", "lora5"]
+    table = table[table.arm.isin(order)].copy()
+    assert table.groupby("arm").size().to_dict() == {arm: 510 for arm in order}
+    assert not table.duplicated(["arm", "draw", "donor"]).any()
+    keys = {arm: set(map(tuple, table[table.arm == arm][["draw", "donor"]].to_numpy())) for arm in order}
+    assert all(keys[arm] == keys["lora5"] for arm in order)
+    observed = table[table.arm == "lora5"].set_index(["draw", "donor"])[[f"obs_{c}" for c in TYPES]].sort_index()
+    for arm in order:
+        got = table[table.arm == arm].set_index(["draw", "donor"])[observed.columns].sort_index()
+        assert np.allclose(got.to_numpy(), observed.to_numpy(), atol=1e-12, rtol=0)
+    # Identical descriptive estimator: first average held-out draws within each
+    # donor, then average donors. Every scatter point is the same donor mean.
+    grouped = table.groupby(["arm", "donor"])[[f"{pre}_{c}" for pre in ("obs", "gen") for c in TYPES]].mean()
+    observed_values = grouped.loc["lora5"][[f"obs_{c}" for c in TYPES]].mean().to_numpy()
+    values = np.vstack([observed_values] + [grouped.loc[arm][[f"gen_{c}" for c in TYPES]].mean().to_numpy()
+                                         for arm in order])
+    assert np.allclose(values.sum(1), 1, atol=1e-8)
+
+    # Compare errors before averaging repeated draws; averaging signed errors
+    # first can cancel mistakes. Donors, not donor-draw rows, receive equal weight.
+    errors = {}
+    for arm in ("lora5", "scsemi5"):
+        rows = table[table.arm == arm].set_index(["draw", "donor"]).sort_index()
+        err = pd.DataFrame({typ: 100 * (rows[f"gen_{typ}"] - rows[f"obs_{typ}"]).abs()
+                            for typ in TYPES}, index=rows.index)
+        assert np.allclose(err.sum(axis=1) / 100, rows["comp_l1"], atol=1e-12, rtol=0)
+        errors[arm] = err.groupby(level="donor").mean().sort_index()
+    assert errors["lora5"].index.equals(errors["scsemi5"].index)
+    assert len(errors["lora5"]) == 107
+    assert all(np.isfinite(x.to_numpy()).all() for x in errors.values())
+    delta = errors["scsemi5"] - errors["lora5"]
+    tolerance = 1e-10
+    wins = (delta > tolerance).sum()
+    losses = (delta < -tolerance).sum()
+    ties = (delta.abs() <= tolerance).sum()
+    assert ((wins + losses + ties) == 107).all()
+    means = {arm: data.mean() for arm, data in errors.items()}
+    totals = {arm: data.sum(axis=1) / 100 for arm, data in errors.items()}
+    total_delta = totals["scsemi5"] - totals["lora5"]
+    overall_wins = int((total_delta > tolerance / 100).sum())
+    reduction = 100 * (1 - totals["lora5"].mean() / totals["scsemi5"].mean())
+    summary = pd.DataFrame({"cell_type": TYPES,
+        "pbisc_mae_pp": means["lora5"].to_numpy(),
+        "scsemi_mae_pp": means["scsemi5"].to_numpy(),
+        "scsemi_minus_pbisc_pp": delta.mean().to_numpy(),
+        "pbisc_wins": wins.to_numpy(), "scsemi_wins": losses.to_numpy(),
+        "ties": ties.to_numpy(), "donors": 107})
+    rows = []
+    for donor in errors["lora5"].index:
+        for typ in TYPES:
+            rows.append({"donor": donor, "cell_type": typ,
+                "pbisc_mae_pp": errors["lora5"].loc[donor,typ],
+                "scsemi_mae_pp": errors["scsemi5"].loc[donor,typ],
+                "scsemi_minus_pbisc_pp": delta.loc[donor,typ]})
+    audit = D / "data/composition_error_20261013"
+    audit.mkdir(exist_ok=True)
+    summary.to_csv(audit/"cell_type_summary.csv",index=False)
+    pd.DataFrame(rows).to_csv(audit/"paired_donor_errors.csv",index=False)
+    _source(audit/"cell_type_summary.csv", "derived donor-weighted absolute-error summary", inputs)
+    _source(audit/"paired_donor_errors.csv", "derived paired donor errors; repeated draws averaged within donor", inputs)
+
+    # High contrast is deliberate for this two-method comparison; other slides'
+    # palettes are preserved. Separate bars + labels carry the result, not colour alone.
+    colors = {"lora5": ST.ARM["lora"], "scsemi5": "#0072B2"}
+    labels = ["T cells", "B cells", "Monocytes", "NK cells", "DC", "Other"]
+    fig, axes = plt.subplots(1, 2, figsize=ST.FULL,
+                            gridspec_kw={"width_ratios": [1.15, 1], "wspace": 0.43})
+    fig.subplots_adjust(left=0.10, right=0.98, bottom=0.15, top=0.70)
+    fig.text(0.10,0.975,
+        f"Overall composition L1: {totals['lora5'].mean():.3f} vs {totals['scsemi5'].mean():.3f}",
+        fontsize=17,weight="bold",ha="left",va="top",color=ST.INK)
+    fig.text(0.10,0.895,
+        f"PBISC: {reduction:.1f}% lower mean error; closer in {overall_wins}/107 donors",
+        fontsize=14,ha="left",va="top",color=ST.INK)
+    fig.legend([Patch(facecolor=colors[a]) for a in ("lora5","scsemi5")],
+        ["PBISC LoRA", "scSemiProfiler"], loc="upper right",bbox_to_anchor=(0.99,1.005),
+        ncol=1,fontsize=12,handlelength=1.1,handletextpad=0.55)
+    y = np.arange(len(TYPES)); height=0.29
+    for offset,arm in [(-0.17,"lora5"),(0.17,"scsemi5")]:
+        v=means[arm].to_numpy()
+        axes[0].barh(y+offset,v,height=height,color=colors[arm],zorder=3)
+        for yy,value in zip(y+offset,v):
+            axes[0].text(value+0.10,yy,f"{value:.2f}",va="center",ha="left",fontsize=11,color=ST.INK)
+    axes[0].set(yticks=y,yticklabels=labels,xlim=(0,8.05),xlabel="Mean absolute error (percentage points)")
+    axes[0].set_title("Error by cell type  ↓",loc="left",fontsize=14,pad=13)
+    axes[0].set_xticks([0,2,4,6,8]);axes[0].invert_yaxis()
+    axes[0].grid(axis="x",color="#E9E9E9",linewidth=.7,zorder=0)
+    for ax in axes:
+        ax.tick_params(axis="y",length=0,pad=8,labelsize=12)
+        ax.spines["left"].set_visible(False)
+    width=wins.to_numpy()/107*100
+    tied=ties.to_numpy()/107*100
+    other=losses.to_numpy()/107*100
+    axes[1].barh(y,width,height=.62,color=colors["lora5"],zorder=2)
+    axes[1].barh(y,tied,left=width,height=.62,color="#AAAAAA",zorder=2)
+    axes[1].barh(y,other,left=width+tied,height=.62,color=colors["scsemi5"],zorder=2)
+    for i,(nw,nl,nt) in enumerate(zip(wins,losses,ties)):
+        # Counts identify small segments without shrinking their text.
+        axes[1].text(width[i]/2,i,str(nw),ha="center",va="center",color="white",fontsize=12,weight="bold")
+        axes[1].text(width[i]+tied[i]+other[i]/2,i,str(nl),ha="center",va="center",color="white",fontsize=12,weight="bold")
+        if nt:
+            axes[1].text(101,i,f"{nt} tie",ha="left",va="center",fontsize=10)
+    axes[1].axvline(50,color=ST.INK,linestyle=(0,(3,3)),linewidth=.8,zorder=1)
+    axes[1].set(yticks=y,yticklabels=labels,xlim=(0,100),xlabel="Donors with lower error (%)")
+    axes[1].set_title("Which method is closer?",loc="left",fontsize=14,pad=13)
+    axes[1].set_xticks([0,25,50,75,100]);axes[1].invert_yaxis()
+    fig.text(.10,.012,"Absolute error per draw, then averaged within donor. Equal donor weights; descriptive comparison.",
+             fontsize=10.5,color=ST.SUBINK,ha="left")
+    fig.text(.98,.805,"Numbers in bars = donors (n = 107)",fontsize=11,color=ST.SUBINK,ha="right")
+    return _save(fig, "healthy_composition", assets), {
+        "estimator": "100*abs(generated-observed) per donor-draw; average held-out draws within donor; equal donor weight",
+        "comparison": "paired errors, not error after averaging proportions",
+        "rows_per_arm":510,"donors":107,"cell_types":TYPES,
+        "summary":summary.to_dict(orient="records"),
+        "composition_l1":{arm:float(data.mean()) for arm,data in totals.items()},
+        "overall_pbisc_win_donors":overall_wins,"relative_l1_reduction_percent":float(reduction),
+        "tie_tolerance_pp":tolerance,"method_colors":{"PBISC LoRA":colors["lora5"],"scSemiProfiler":colors["scsemi5"]},
+        "unchanged_mean_composition_fractions":values.tolist(),
+        "raw_predictions_changed":False,"derived_summary_changed":True,
+        "inference":"descriptive only; overlapping support draws are not independent training replicates"}
+
+
+def _rare_observed(inputs: dict[str, dict], assets: list[str]) -> tuple[str, dict]:
+    path = _source(P / "source_data/rare_all_states.csv", "existing corrected rare-state summaries", inputs)
+    table = pd.read_csv(path)
+    assert set(table.evidence_tier) == {"diagnostic"} and set(table.design) == {"legacy_5donor"}
+    states = list(dict.fromkeys(table.state))
+    values = table[table.arm == "lora5"].groupby(["state", "cohort"]).obs_pct.median().unstack().reindex(
+        index=states, columns=COHORTS)
+    fig, ax = plt.subplots(figsize=ST.FULL)
+    fig.subplots_adjust(left=0.305, right=0.87, bottom=0.09, top=0.95)
+    cmap = plt.get_cmap("YlGnBu").copy()
+    cmap.set_bad("#F2F2F2")
+    image = ax.imshow(values.to_numpy(), cmap=cmap, aspect="auto", vmin=0,
+                      vmax=np.nanmax(values.to_numpy()), interpolation="nearest")
+    ax.set_xticks(range(4), COHORT_NAMES)
+    ax.set_yticks(range(len(states)), [STATE_LABELS.get(state, state) for state in states])
+    ax.xaxis.tick_top()
+    ax.tick_params(axis="both", length=0, pad=8, labelsize=ST.FS - 2)
+    ax.set_xticks(np.arange(-0.5, 4, 1), minor=True)
+    ax.set_yticks(np.arange(-0.5, len(states), 1), minor=True)
+    ax.grid(which="minor", color="white", linewidth=1.4)
+    ax.tick_params(which="minor", length=0)
+    for spine in ax.spines.values():
+        spine.set_visible(False)
+    for i, state in enumerate(states):
+        for j, cohort in enumerate(COHORTS):
+            value = values.iloc[i, j]
+            rgb = image.cmap(image.norm(value))[:3] if pd.notna(value) else (1, 1, 1)
+            luminance = 0.2126 * rgb[0] + 0.7152 * rgb[1] + 0.0722 * rgb[2]
+            ax.text(j, i, "NA" if pd.isna(value) else f"{value:.2g}", ha="center", va="center",
+                    fontsize=ST.FS - 2, color="white" if luminance < 0.45 else ST.INK)
+    color_ax = fig.add_axes([0.895, 0.14, 0.018, 0.70])
+    bar = fig.colorbar(image, cax=color_ax)
+    bar.set_label("Observed share within parent (%)", fontsize=ST.FS - 1, labelpad=10)
+    bar.ax.tick_params(labelsize=ST.FS - 2)
+    bar.outline.set_visible(False)
+    return _save(fig, "rare_observed", assets), {
+        "estimator": "median obs_pct across draws from the lora5 observed reference; no additional state filtering",
+        "states": states, "display_labels": [STATE_LABELS.get(state, state) for state in states],
+        "cohorts": COHORTS, "state_cohort_combinations": int(values.notna().sum().sum()),
+        "percentages": [[None if pd.isna(v) else float(v) for v in row] for row in values.to_numpy()]}
+
+
+def _rare_comparison(inputs: dict[str, dict], assets: list[str]) -> tuple[str, dict]:
+    path = _source(P / "source_data/rare_all_states_aggregate.csv", "existing zero-aware rare-state metrics", inputs)
+    table = pd.read_csv(path)
+    errors = table.pivot(index=["cohort", "state"], columns="arm", values="abundance_error")
+    ranks = table.pivot(index=["cohort", "state"], columns="arm", values="ranking")
+    assert len(errors) == 59
+    fig, axes = plt.subplots(1, 2, figsize=ST.FULL, gridspec_kw={"wspace": 0.34})
+    fig.subplots_adjust(left=0.085, right=0.975, bottom=0.20, top=0.88)
+    for cohort, label in zip(COHORTS, COHORT_NAMES):
+        for ax, data in zip(axes, (errors, ranks)):
+            points = data.loc[cohort]
+            ax.scatter(points.scsemi5, points.lora5, color=COHORT_COLORS[cohort], s=34,
+                       edgecolor=ST.INK, linewidth=0.4, alpha=0.85, label=label, zorder=3)
+    limit = float(np.nanmax(errors[["scsemi5", "lora5"]].to_numpy())) * 1.05
+    axes[0].plot([0, limit], [0, limit], color=ST.GREY, linewidth=0.9, linestyle="--", zorder=1)
+    axes[0].set(xlim=(0, limit), ylim=(0, limit))
+    axes[0].xaxis.set_major_locator(MaxNLocator(nbins=5))
+    axes[0].yaxis.set_major_locator(MaxNLocator(nbins=5))
+    axes[0].set_title("Abundance error (pp)", loc="left", fontsize=ST.FS, pad=12)
+    axes[1].plot([-1, 1], [-1, 1], color=ST.GREY, linewidth=0.9, linestyle="--", zorder=1)
+    axes[1].set(xlim=(-1, 1), ylim=(-1, 1))
+    axes[1].set_xticks([-1, -0.5, 0, 0.5, 1]); axes[1].set_yticks([-1, -0.5, 0, 0.5, 1])
+    axes[1].set_title("Donor ranking (Spearman ρ)", loc="left", fontsize=ST.FS, pad=12)
+    for ax in axes:
+        ax.set_xlabel("scSemiProfiler", fontsize=ST.FS)
+        ax.set_ylabel("PBISC LoRA", fontsize=ST.FS)
+        ax.set_aspect("equal")
+        ax.grid(color="#EEEEEE", linewidth=0.7, zorder=0)
+        ax.tick_params(labelsize=ST.FS - 2)
+    handles = [Line2D([], [], linestyle="none", marker="o", color=COHORT_COLORS[cohort],
+                      markeredgecolor=ST.INK, markeredgewidth=0.4, markersize=7) for cohort in COHORTS]
+    fig.legend(handles, COHORT_NAMES, loc="lower center", bbox_to_anchor=(0.53, 0.025), ncol=4,
+               fontsize=ST.FS - 2, columnspacing=1.9, handletextpad=0.4)
+    return _save(fig, "rare_corrected_comparison", assets), {
+        "estimator": "saved abundance_error and ranking columns; pivot only, no recomputation or imputation",
+        "state_cohort_combinations": len(errors), "cohort_colors": COHORT_COLORS,
+        "finite_abundance_pairs": int(np.isfinite(errors[["scsemi5", "lora5"]]).all(axis=1).sum()),
+        "finite_ranking_pairs": int(np.isfinite(ranks[["scsemi5", "lora5"]]).all(axis=1).sum())}
+
+
+def build() -> dict:
+    """Return figures=[composition, observed, comparison] PDFs plus exports/provenance."""
+    inputs: dict[str, dict] = {}
+    assets: list[str] = []
+    _source(Path(ST.__file__), "original 76-slide matplotlib style and method palette", inputs)
+    _source(D / "tools/fig5_rare_states.py", "original Fig. 5 cohort-colour mapping", inputs)
+    _source(Path(__file__), "saved-table figure restyling code", inputs)
+    # An importing deck builder may already have changed matplotlib defaults.
+    # Reapply the shared style's conventions locally without mutating style.py.
+    with plt.rc_context({"font.family": "sans-serif", "font.sans-serif": ["Arial", "Liberation Sans", "DejaVu Sans"],
+                         "font.size": ST.FS, "axes.titlesize": ST.FS, "axes.labelsize": ST.FS,
+                         "xtick.labelsize": ST.FS - 2, "ytick.labelsize": ST.FS - 2, "legend.fontsize": ST.FS - 2,
+                         "axes.spines.top": False, "axes.spines.right": False, "axes.linewidth": 1.0,
+                         "legend.frameon": False, "figure.facecolor": "white", "axes.facecolor": "white",
+                         "pdf.fonttype": 42, "ps.fonttype": 42, "svg.fonttype": "none"}):
+        results = [_healthy(inputs, assets), _rare_observed(inputs, assets), _rare_comparison(inputs, assets)]
+    return {"figures": [path for path, _ in results], "assets": assets, "inputs": list(inputs.values()),
+            "calculations": {name: record for name, (_, record) in zip(
+                ("healthy_composition", "rare_observed", "rare_corrected_comparison"), results)},
+            "new_training": False, "new_generation": False, "scientific_estimates_changed": True, "underlying_experiment_outputs_changed": False}
+
+
+if __name__ == "__main__":
+    print(json.dumps(build(), indent=2))

paper-releases/pbisc-diffusion/updates/20261013T023822Z-pbisc-yellow-scsemi-blue/diffusion/dataset/server_1/model8/presentations/2026-09-30--paper-summary--r00/figures/method_palette_20261013/secA_scsemi_readout.palette.json

@@ -0,0 +1,93 @@
+{
+  "schema": "pbisc.vector-method-palette.v1",
+  "source": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/secA_scsemi_readout.pdf",
+  "source_sha256": "8d88d3ba4d9e6bab62c739d03830dd5c90a72a580ea6ac1bca679ab6584ccad2",
+  "destination": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/method_palette_20261013/secA_scsemi_readout.pdf",
+  "destination_sha256": "f2993d3ff6b0dffa4f728a6d9055c933a7743f303dfab4f8b1cdf0381158cf5d",
+  "mapping": {
+    "#D55E00": "#C5AA4B",
+    "#E69F00": "#2166AC",
+    "#F5D08A": "#67A9CF"
+  },
+  "rgb_absolute_tolerance": 1e-08,
+  "streams_inspected": 59,
+  "modified_streams": 2,
+  "stream_checks": [
+    {
+      "stream": "page0/Contents0",
+      "noncolor_operator_sha256_before": "d2ba19f4c4ae008977c0bf15a507dfda7f4e9eb2ed9556194607f8b644e39415",
+      "noncolor_operator_sha256_after": "d2ba19f4c4ae008977c0bf15a507dfda7f4e9eb2ed9556194607f8b644e39415",
+      "colour_operators_changed": 107
+    },
+    {
+      "stream": "page0/Pattern//H1",
+      "noncolor_operator_sha256_before": "745e292337bbbaf4c1e0d7de9b706b7e41b61b28f427b614194c152d6e69afbe",
+      "noncolor_operator_sha256_after": "745e292337bbbaf4c1e0d7de9b706b7e41b61b28f427b614194c152d6e69afbe",
+      "colour_operators_changed": 1
+    }
+  ],
+  "colour_counts": [
+    {
+      "old": "#D55E00",
+      "new": "#C5AA4B",
+      "operator": "RG",
+      "count": 1
+    },
+    {
+      "old": "#D55E00",
+      "new": "#C5AA4B",
+      "operator": "rg",
+      "count": 42
+    },
+    {
+      "old": "#E69F00",
+      "new": "#2166AC",
+      "operator": "RG",
+      "count": 1
+    },
+    {
+      "old": "#E69F00",
+      "new": "#2166AC",
+      "operator": "rg",
+      "count": 42
+    },
+    {
+      "old": "#F5D08A",
+      "new": "#67A9CF",
+      "operator": "rg",
+      "count": 22
+    }
+  ],
+  "image_payload_count": 0,
+  "image_payloads": [],
+  "rectangle_rules_matched": 0,
+  "rectangle_checks": [],
+  "scope_insertions": "Only balanced q/RGB/RGB/paint/Q wrappers; original non-colour operators preserved",
+  "preservation": {
+    "noncolor_operators": true,
+    "text_geometry": true,
+    "page_rectangles": true,
+    "image_payloads": true,
+    "source_unchanged": true
+  },
+  "preservation_pass": true,
+  "render": {
+    "pdf": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/method_palette_20261013/secA_scsemi_readout.pdf",
+    "png": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/method_palette_20261013/secA_scsemi_readout.png",
+    "dpi": 200,
+    "page_rect": [
+      0.0,
+      0.0,
+      714.8524780273438,
+      327.13275146484375
+    ],
+    "width": 1986,
+    "height": 909,
+    "png_sha256": "4ea8b33687f89e8545477e9c89d01e171026d3dadeac194a765b4b0c0ea96557",
+    "rendering": "native PDF rendering; no raster recolouring"
+  },
+  "original_png": "/home/sj_server_1/PBISC/immun-pbmc-fm/model_8_sparse_data_diffusion/presentations/2026-09-30--paper-summary--r00/figures/secA_scsemi_readout.png",
+  "original_png_sha256": "06468c2e50e6107184e5345dbfe3e9e79783410b3ff30f133f84a7e28d0d5b8b",
+  "semantic_reason": "PBISC gold; official scSemi blue and modified readout pale blue with original hatch.",
+  "policy_sha256": "a94e734ef7e35a2cb09f63acb81eb48b0c2edaccd3743caf3002cf8f8314d07d"
+}