Add five-seed surface response and held-out calibration evidence

surface-contrast-grid/five-seed-addendum/README.md

@@ -0,0 +1,96 @@
+# Five-seed calibration addendum
+
+This addendum tests whether the original 4×4 contrast × pitch response surface
+was a seed-specific artifact and separates fixed-threshold recall from ranking
+quality. It does **not** change the production checkpoint or tune the model.
+
+![Five-seed response and held-out calibration](five_seed_response_addendum.png)
+
+## Design
+
+- Five independently generated phantom instances per cell: seeds 0–4.
+- Four papyrus levels × four winding pitches × five seeds = **80 inferences**.
+- Same pinned `surface_recto_059_redo` checkpoint, 192³ sliding windows, 50%
+  step, Gaussian blending, no TTA.
+- Seed 0 exactly reproduces all original recall, false-positive-rate, and AUC
+  values. It is also the **only** seed used to select global thresholds.
+- Global thresholds are frozen, then evaluated on held-out seeds 1–4.
+- Raw class-score histograms use 4,096 bins. Per-cell best-Youden results are
+  retained as descriptive same-cell upper bounds, not headline evidence.
+
+## What survived replication
+
+The AUC geometry response is highly stable. Mean ± sample SD across the five
+seeds is:
+
+| Pitch (µm) | Near-sheet AUC | Recall @ 0.5 | Recall @ 5% FPR |
+|---:|---:|---:|---:|
+| 260 | 0.7881 ± 0.0012 | 0.3427 ± 0.0025 | 0.1881 ± 0.0018 |
+| 200 | 0.8566 ± 0.0027 | 0.3106 ± 0.0027 | 0.3086 ± 0.0047 |
+| 150 | **0.9279 ± 0.0027** | **0.3836 ± 0.0028** | **0.6679 ± 0.0060** |
+| 110 | 0.9099 ± 0.0014 | 0.3410 ± 0.0008 | 0.4872 ± 0.0074 |
+
+Pitch 150 has higher mean AUC than pitch 110 in **all five seeds**. At both 1%
+and 5% controlled near-background FPR, the ordering is
+**150 > 110 > 200 > 260 in every seed**. The original intermediate-pitch peak
+is therefore not explained by the seed-0 instance.
+
+Contrast response also survives. Mean AUC rises monotonically across papyrus
+levels 35, 50, 65, and 90: **0.7984, 0.8735, 0.9015, 0.9092**.
+
+## What threshold calibration changes
+
+Recall at the fixed 0.5 threshold is not directly comparable across pitches:
+mean near-background FPR varies from **10.84% at pitch 260** to **1.74% at
+pitch 150**. The pitch-200 recall@0.5 dip is reproducible—it is the lowest in
+all five seeds—but fixed-threshold recall mixes detection with score
+calibration.
+
+The controlled-FPR readout agrees with AUC and gives the clearer geometry
+result. At 5% FPR, recall is **0.188, 0.309, 0.668, 0.487** for pitches 260,
+200, 150, and 110 µm.
+
+One pooled threshold selected on seed 0 transfers almost exactly:
+
+| Seed-0 target | Frozen threshold | Seed-0 recall / FPR | Held-out seeds 1–4 recall / FPR |
+|---:|---:|---:|---:|
+| 1% FPR | 0.8440 | 10.99% / 1.00% | 10.93% / 0.99% |
+| 5% FPR | 0.5725 | 30.39% / 5.00% | 30.38% / 4.86% |
+
+This shows stable calibration transfer within the generator. It does not imply
+that these thresholds transfer to real CT.
+
+## Interpretation
+
+The replicated result supports three narrow claims:
+
+1. Both controlled axes affect this checkpoint's response.
+2. The intermediate-pitch AUC peak is stable across five analytic instances.
+3. Recall@0.5 partly measures threshold placement; controlled-FPR recall is a
+   better companion to AUC on this instrument.
+
+It does **not** establish response independence, a universal pitch optimum, or
+real-scroll accuracy. Five seeds quantify variation inside this analytic
+phantom generator. They do not sample tearing, plastic deformation, papyrus
+texture, reconstruction artifacts, or annotation uncertainty.
+
+## Files and provenance
+
+- `seed-runs/seed_0.json` … `seed_4.json` — raw per-cell results and score
+  histograms.
+- `replicate_results_validated.json` — aggregate summaries and held-out
+  calibration.
+- `validate_replicates.py` — independent raw-to-summary validator; **1,930 / 1,930 checks pass**.
+- `aggregate_replicates.py` — deterministic post-processing from the raw seed
+  files.
+- `runtime/vesuvius-surface-contrast-grid.py` — exact completed Kaggle v8
+  source; SHA-256
+  `368e5d375e7e62221a49f9be8d91a6a7d66c1f9dcb7f6043ba985a6501cebc69`.
+- `five_seed_response_addendum.png` and `render_replicate_figure.py` — figure
+  and deterministic renderer.
+
+Validated aggregate SHA-256:
+`b265d6ecf7e8bfb8a03a616cceaf12f65d2535175778df149294ad39828fe553`.
+
+Checkpoint SHA-256:
+`f1990a02ac91889c1f989522ae0e45421a91cb666320448aaf579d42b081636f`.

surface-contrast-grid/five-seed-addendum/aggregate_replicates.py

@@ -0,0 +1,258 @@
+#!/usr/bin/env python3
+"""Aggregate seeded surface-grid runs and perform held-out calibration."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+from collections import defaultdict
+from pathlib import Path
+
+import numpy as np
+
+
+METRICS = (
+    "surface_recall_at_0_5",
+    "near_sheet_auc",
+    "recall_at_fpr_1pct",
+    "recall_at_fpr_5pct",
+    "best_youden_recall",
+    "best_youden_threshold",
+)
+
+
+def sha256_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(8 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def sample_summary(values: list[float]) -> dict:
+    array = np.asarray(values, dtype=np.float64)
+    return {
+        "n": int(array.size),
+        "mean": float(array.mean()),
+        "sample_sd": float(array.std(ddof=1)) if array.size > 1 else 0.0,
+        "min": float(array.min()),
+        "max": float(array.max()),
+        "values_by_seed": [float(value) for value in array],
+    }
+
+
+def rates_at_bin(positive: np.ndarray, negative: np.ndarray, index: int) -> tuple[float, float]:
+    return (
+        float(positive[index:].sum() / positive.sum()),
+        float(negative[index:].sum() / negative.sum()),
+    )
+
+
+def best_youden_index(positive: np.ndarray, negative: np.ndarray) -> int:
+    true_positive = np.cumsum(positive[::-1])[::-1] / positive.sum()
+    false_positive = np.cumsum(negative[::-1])[::-1] / negative.sum()
+    score = true_positive - false_positive
+    candidates = np.flatnonzero(
+        np.isclose(score, score.max(), rtol=0.0, atol=1e-15)
+    )
+    return int(candidates[-1])
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--runs", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    args = parser.parse_args()
+
+    paths = sorted(args.runs.glob("seed_*.json"), key=lambda path: int(path.stem.split("_")[-1]))
+    if len(paths) < 3:
+        raise RuntimeError(f"expected at least three seed runs, found {len(paths)}")
+
+    runs = []
+    for path in paths:
+        seed = int(path.stem.split("_")[-1])
+        result = json.loads(path.read_text(encoding="utf-8"))
+        if not result.get("canonical") or len(result.get("cells", [])) != 16:
+            raise RuntimeError(f"non-canonical run: {path}")
+        runs.append((seed, path, result))
+
+    bins = int(runs[0][2]["cells"][0]["metrics"]["histogram_bins"])
+    for _, path, result in runs:
+        for cell in result["cells"]:
+            if int(cell["metrics"]["histogram_bins"]) != bins:
+                raise RuntimeError(f"histogram bin mismatch in {path}")
+
+    # Seed 0 is calibration-only for a single global operating threshold.
+    calibration = runs[0][2]
+    calibration_positive = np.zeros(bins, dtype=np.int64)
+    calibration_negative = np.zeros(bins, dtype=np.int64)
+    for cell in calibration["cells"]:
+        calibration_positive += np.asarray(
+            cell["metrics"]["positive_score_histogram"], dtype=np.int64
+        )
+        calibration_negative += np.asarray(
+            cell["metrics"]["negative_score_histogram"], dtype=np.int64
+        )
+    threshold_index = best_youden_index(calibration_positive, calibration_negative)
+    threshold = threshold_index / bins
+    calibration_recall, calibration_fpr = rates_at_bin(
+        calibration_positive, calibration_negative, threshold_index
+    )
+
+    heldout_by_seed = []
+    heldout_positive = np.zeros(bins, dtype=np.int64)
+    heldout_negative = np.zeros(bins, dtype=np.int64)
+    for seed, _, result in runs[1:]:
+        positive = np.zeros(bins, dtype=np.int64)
+        negative = np.zeros(bins, dtype=np.int64)
+        for cell in result["cells"]:
+            positive += np.asarray(
+                cell["metrics"]["positive_score_histogram"], dtype=np.int64
+            )
+            negative += np.asarray(
+                cell["metrics"]["negative_score_histogram"], dtype=np.int64
+            )
+        recall, fpr = rates_at_bin(positive, negative, threshold_index)
+        heldout_by_seed.append({"seed": seed, "recall": recall, "fpr": fpr})
+        heldout_positive += positive
+        heldout_negative += negative
+    heldout_recall, heldout_fpr = rates_at_bin(
+        heldout_positive, heldout_negative, threshold_index
+    )
+
+    target_fpr_operating_points = {}
+    calibration_fpr_curve = (
+        np.cumsum(calibration_negative[::-1])[::-1]
+        / calibration_negative.sum()
+    )
+    for target in (0.01, 0.05):
+        candidates = np.flatnonzero(calibration_fpr_curve <= target)
+        index = int(candidates[0]) if candidates.size else bins - 1
+        calibration_target_recall, calibration_target_fpr = rates_at_bin(
+            calibration_positive, calibration_negative, index
+        )
+        target_heldout_by_seed = []
+        for seed, _, result in runs[1:]:
+            positive = np.zeros(bins, dtype=np.int64)
+            negative = np.zeros(bins, dtype=np.int64)
+            for cell in result["cells"]:
+                positive += np.asarray(
+                    cell["metrics"]["positive_score_histogram"], dtype=np.int64
+                )
+                negative += np.asarray(
+                    cell["metrics"]["negative_score_histogram"], dtype=np.int64
+                )
+            recall, fpr = rates_at_bin(positive, negative, index)
+            target_heldout_by_seed.append(
+                {"seed": seed, "recall": recall, "fpr": fpr}
+            )
+        target_heldout_recall, target_heldout_fpr = rates_at_bin(
+            heldout_positive, heldout_negative, index
+        )
+        target_fpr_operating_points[f"{int(target * 100)}pct"] = {
+            "target_fpr": target,
+            "threshold_selected_on_seed_0": index / bins,
+            "calibration_recall": calibration_target_recall,
+            "calibration_fpr": calibration_target_fpr,
+            "heldout_pooled_recall": target_heldout_recall,
+            "heldout_pooled_fpr": target_heldout_fpr,
+            "heldout_by_seed": target_heldout_by_seed,
+        }
+
+    grouped_cells: dict[tuple[int, float], dict[str, list[float]]] = defaultdict(
+        lambda: defaultdict(list)
+    )
+    pitch_seed_means: dict[str, dict[float, list[float]]] = {
+        metric: defaultdict(list) for metric in METRICS
+    }
+    papyrus_seed_means: dict[str, dict[int, list[float]]] = {
+        metric: defaultdict(list) for metric in METRICS
+    }
+    for _, _, result in runs:
+        per_pitch: dict[str, dict[float, list[float]]] = {
+            metric: defaultdict(list) for metric in METRICS
+        }
+        per_papyrus: dict[str, dict[int, list[float]]] = {
+            metric: defaultdict(list) for metric in METRICS
+        }
+        for cell in result["cells"]:
+            key = (int(cell["papyrus"]), float(cell["pitch_um"]))
+            for metric in METRICS:
+                value = float(cell["metrics"][metric])
+                grouped_cells[key][metric].append(value)
+                per_pitch[metric][key[1]].append(value)
+                per_papyrus[metric][key[0]].append(value)
+        for metric in METRICS:
+            for pitch, values in per_pitch[metric].items():
+                pitch_seed_means[metric][pitch].append(float(np.mean(values)))
+            for papyrus, values in per_papyrus[metric].items():
+                papyrus_seed_means[metric][papyrus].append(float(np.mean(values)))
+
+    output = {
+        "schema_version": "2.0",
+        "replicate_design": {
+            "seeds": [seed for seed, _, _ in runs],
+            "replicates_per_cell": len(runs),
+            "cells_per_seed": 16,
+            "total_inferences": 16 * len(runs),
+            "seed_0_role": "included in replicate summaries; calibration-only for the global threshold",
+            "heldout_threshold_seeds": [seed for seed, _, _ in runs[1:]],
+        },
+        "source_runs": [
+            {"seed": seed, "file": path.name, "sha256": sha256_file(path)}
+            for seed, path, _ in runs
+        ],
+        "global_threshold_calibration": {
+            "histogram_bins": bins,
+            "selected_on_seed": runs[0][0],
+            "threshold": threshold,
+            "selection_rule": "pooled maximum Youden J; highest threshold on ties",
+            "calibration_recall": calibration_recall,
+            "calibration_fpr": calibration_fpr,
+            "heldout_pooled_recall": heldout_recall,
+            "heldout_pooled_fpr": heldout_fpr,
+            "heldout_by_seed": heldout_by_seed,
+            "target_fpr_operating_points": target_fpr_operating_points,
+        },
+        "cell_replicate_summaries": [
+            {
+                "papyrus": papyrus,
+                "pitch_um": pitch,
+                "metrics": {
+                    metric: sample_summary(values)
+                    for metric, values in sorted(grouped_cells[(papyrus, pitch)].items())
+                },
+            }
+            for papyrus, pitch in sorted(grouped_cells)
+        ],
+        "pitch_summaries": {
+            metric: {
+                str(int(pitch) if float(pitch).is_integer() else pitch): sample_summary(values)
+                for pitch, values in sorted(mapping.items(), reverse=True)
+            }
+            for metric, mapping in pitch_seed_means.items()
+        },
+        "papyrus_summaries": {
+            metric: {
+                str(papyrus): sample_summary(values)
+                for papyrus, values in sorted(mapping.items())
+            }
+            for metric, mapping in papyrus_seed_means.items()
+        },
+        "limitations": [
+            "Five seeds estimate variation across this analytic phantom generator, not across real scrolls.",
+            "Per-cell best-Youden metrics are descriptive upper bounds chosen and measured on the same cell.",
+            "The global threshold is selected on seed 0 and evaluated on seeds 1-4 to avoid calibration leakage.",
+            "Histogram-derived thresholds have resolution 1/4096.",
+            "No confidence interval here supports claims of real-scroll accuracy.",
+        ],
+    }
+    args.output.write_text(json.dumps(output, indent=2) + "\n", encoding="utf-8")
+    print(f"wrote {args.output}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

surface-contrast-grid/five-seed-addendum/render_replicate_figure.py

@@ -0,0 +1,181 @@
+#!/usr/bin/env python3
+"""Render the five-seed calibration addendum as a deterministic PNG."""
+
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+from PIL import Image, ImageDraw, ImageFont
+
+
+HERE = Path(__file__).resolve().parent
+DATA = HERE / "replicate_results_validated.json"
+OUTPUT = HERE / "five_seed_response_addendum.png"
+FONT = "/System/Library/Fonts/Avenir Next.ttc"
+
+WIDTH, HEIGHT = 1900, 1180
+BG = "#F4F0E7"
+PANEL = "#FFFCF6"
+INK = "#172725"
+MUTED = "#65716D"
+GRID = "#DDD8CD"
+TEAL = "#087F7A"
+RED = "#B44C3D"
+GOLD = "#C38B2D"
+BLUE = "#3D6FA5"
+
+
+def font(size: int, bold: bool = False) -> ImageFont.FreeTypeFont:
+    index = 1 if bold else 0
+    return ImageFont.truetype(FONT, size=size, index=index)
+
+
+def panel(draw: ImageDraw.ImageDraw, box: tuple[int, int, int, int]) -> None:
+    draw.rounded_rectangle(box, radius=24, fill=PANEL, outline="#E2DDD2", width=2)
+
+
+def line_chart(
+    draw: ImageDraw.ImageDraw,
+    box: tuple[int, int, int, int],
+    title: str,
+    subtitle: str,
+    series: list[dict],
+    y_min: float,
+    y_max: float,
+    y_ticks: list[float],
+) -> None:
+    x0, y0, x1, y1 = box
+    draw.text((x0 + 34, y0 + 26), title, fill=INK, font=font(29, True))
+    draw.text((x0 + 34, y0 + 66), subtitle, fill=MUTED, font=font(18))
+    plot = (x0 + 90, y0 + 125, x1 - 36, y1 - 72)
+    px0, py0, px1, py1 = plot
+    pitches = [260, 200, 150, 110]
+
+    def x_at(index: int) -> float:
+        return px0 + index * (px1 - px0) / 3
+
+    def y_at(value: float) -> float:
+        return py1 - (value - y_min) * (py1 - py0) / (y_max - y_min)
+
+    for tick in y_ticks:
+        y = y_at(tick)
+        draw.line((px0, y, px1, y), fill=GRID, width=2)
+        draw.text((px0 - 14, y), f"{tick:.2f}", anchor="rm", fill=MUTED, font=font(16))
+    for index, pitch in enumerate(pitches):
+        x = x_at(index)
+        draw.line((x, py1, x, py1 + 8), fill=MUTED, width=2)
+        draw.text((x, py1 + 19), str(pitch), anchor="ma", fill=INK, font=font(17))
+    draw.text(((px0 + px1) / 2, y1 - 18), "winding pitch (µm)", anchor="mm", fill=MUTED, font=font(17))
+
+    for item in series:
+        values = item["values"]
+        points = [(x_at(index), y_at(value)) for index, value in enumerate(values)]
+        draw.line(points, fill=item["color"], width=item.get("width", 5), joint="curve")
+        for x, y in points:
+            radius = item.get("radius", 6)
+            draw.ellipse((x - radius, y - radius, x + radius, y + radius), fill=item["color"])
+        label_x = item.get("label_x", px0 + 8)
+        label_y = item.get("label_y")
+        if label_y is not None:
+            draw.text((label_x, label_y), item["label"], fill=item["color"], font=font(17, True))
+
+
+def main() -> int:
+    data = json.loads(DATA.read_text(encoding="utf-8"))
+    image = Image.new("RGB", (WIDTH, HEIGHT), BG)
+    draw = ImageDraw.Draw(image)
+
+    draw.text((80, 58), "One grid became five", fill=INK, font=font(55, True))
+    draw.text(
+        (82, 126),
+        "80 production-model inferences separate stable geometry response from threshold placement.",
+        fill=MUTED,
+        font=font(24),
+    )
+
+    left = (70, 190, 930, 760)
+    right = (970, 190, 1830, 760)
+    bottom = (70, 800, 1830, 1100)
+    panel(draw, left)
+    panel(draw, right)
+    panel(draw, bottom)
+
+    pitches = ["260", "200", "150", "110"]
+    auc = data["pitch_summaries"]["near_sheet_auc"]
+    for seed in range(5):
+        values = [auc[pitch]["values_by_seed"][seed] for pitch in pitches]
+        points = []
+        px0, py0, px1, py1 = left[0] + 90, left[1] + 125, left[2] - 36, left[3] - 72
+        for index, value in enumerate(values):
+            x = px0 + index * (px1 - px0) / 3
+            y = py1 - (value - 0.76) * (py1 - py0) / (0.95 - 0.76)
+            points.append((x, y))
+        draw.line(points, fill="#B8D2CF", width=2)
+        for x, y in points:
+            draw.ellipse((x - 3, y - 3, x + 3, y + 3), fill="#8FBAB6")
+    line_chart(
+        draw,
+        left,
+        "AUC peak survives every seed",
+        "Thin traces: five instances · thick trace: mean",
+        [{"label": "mean AUC", "values": [auc[p]["mean"] for p in pitches], "color": TEAL}],
+        0.76,
+        0.95,
+        [0.78, 0.82, 0.86, 0.90, 0.94],
+    )
+
+    fixed = data["pitch_summaries"]["surface_recall_at_0_5"]
+    controlled = data["pitch_summaries"]["recall_at_fpr_5pct"]
+    line_chart(
+        draw,
+        right,
+        "Threshold choice changes the story",
+        "Mean across seeds; controlled rate aligns with AUC",
+        [
+            {"label": "recall @ 0.5", "values": [fixed[p]["mean"] for p in pitches], "color": RED},
+            {"label": "recall @ 5% FPR", "values": [controlled[p]["mean"] for p in pitches], "color": BLUE},
+        ],
+        0.15,
+        0.72,
+        [0.20, 0.30, 0.40, 0.50, 0.60, 0.70],
+    )
+    draw.line((right[0] + 565, right[1] + 37, right[0] + 605, right[1] + 37), fill=RED, width=5)
+    draw.text((right[0] + 618, right[1] + 37), "fixed 0.5", anchor="lm", fill=RED, font=font(16, True))
+    draw.line((right[0] + 565, right[1] + 68, right[0] + 605, right[1] + 68), fill=BLUE, width=5)
+    draw.text((right[0] + 618, right[1] + 68), "5% FPR", anchor="lm", fill=BLUE, font=font(16, True))
+
+    draw.text((bottom[0] + 34, bottom[1] + 25), "Seed-0 calibration transfers to four unseen instances", fill=INK, font=font(29, True))
+    draw.text((bottom[0] + 34, bottom[1] + 66), "One pooled threshold is chosen on seed 0, then frozen.", fill=MUTED, font=font(18))
+    operating = data["global_threshold_calibration"]["target_fpr_operating_points"]
+    headers = ["target", "threshold", "seed 0 recall / FPR", "held-out recall / FPR", "shift"]
+    xs = [bottom[0] + 45, bottom[0] + 280, bottom[0] + 580, bottom[0] + 1030, bottom[0] + 1490]
+    for x, header in zip(xs, headers):
+        draw.text((x, bottom[1] + 120), header.upper(), fill=MUTED, font=font(15, True))
+    for row, key in enumerate(("1pct", "5pct")):
+        record = operating[key]
+        y = bottom[1] + 169 + row * 58
+        shift = record["heldout_pooled_recall"] - record["calibration_recall"]
+        values = [
+            f"{record['target_fpr']:.0%} FPR",
+            f"{record['threshold_selected_on_seed_0']:.4f}",
+            f"{record['calibration_recall']:.1%} / {record['calibration_fpr']:.2%}",
+            f"{record['heldout_pooled_recall']:.1%} / {record['heldout_pooled_fpr']:.2%}",
+            f"{shift:+.2%} recall",
+        ]
+        for x, value in zip(xs, values):
+            draw.text((x, y), value, fill=INK if x != xs[-1] else TEAL, font=font(21, x == xs[0]))
+
+    draw.text(
+        (84, 1144),
+        "Five analytic seeds measure generator variation—not real-scroll accuracy. Error bars omitted because all seed traces are shown.",
+        fill=MUTED,
+        font=font(17),
+    )
+    image.save(OUTPUT, optimize=True)
+    print(OUTPUT)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

surface-contrast-grid/five-seed-addendum/replicate_results_validated.json

@@ -0,0 +1,2292 @@
+{
+  "schema_version": "2.0",
+  "replicate_design": {
+    "seeds": [
+      0,
+      1,
+      2,
+      3,
+      4
+    ],
+    "replicates_per_cell": 5,
+    "cells_per_seed": 16,
+    "total_inferences": 80,
+    "seed_0_role": "included in replicate summaries; calibration-only for the global threshold",
+    "heldout_threshold_seeds": [
+      1,
+      2,
+      3,
+      4
+    ]
+  },
+  "source_runs": [
+    {
+      "seed": 0,
+      "file": "seed_0.json",
+      "sha256": "c5a8abe3a54aafa5ae03c5faf48f98e97febb115c26104652e83275128046098"
+    },
+    {
+      "seed": 1,
+      "file": "seed_1.json",
+      "sha256": "1ad5597abc766c5e4c9541376264ab3c3e8de417c09231ca03e9df668a041ad9"
+    },
+    {
+      "seed": 2,
+      "file": "seed_2.json",
+      "sha256": "ab847533eab312ef754bd8e9a18adb40fd0f35c32bbd8162c51ecd353eff113a"
+    },
+    {
+      "seed": 3,
+      "file": "seed_3.json",
+      "sha256": "27708eb34343bd3a6c5467c87ff79c581a1515cfdb45f4553a93f1a860b2b821"
+    },
+    {
+      "seed": 4,
+      "file": "seed_4.json",
+      "sha256": "1c5468d277c697eff65ba907fc26c4cb74ecd6ae74d7c5a49a10d11579f50428"
+    }
+  ],
+  "global_threshold_calibration": {
+    "histogram_bins": 4096,
+    "selected_on_seed": 0,
+    "threshold": 0.02490234375,
+    "selection_rule": "pooled maximum Youden J; highest threshold on ties",
+    "calibration_recall": 0.8887617416205217,
+    "calibration_fpr": 0.31372956965062226,
+    "heldout_pooled_recall": 0.8916385828263844,
+    "heldout_pooled_fpr": 0.31075934084835016,
+    "heldout_by_seed": [
+      {
+        "seed": 1,
+        "recall": 0.8927469122890728,
+        "fpr": 0.31066218735207385
+      },
+      {
+        "seed": 2,
+        "recall": 0.8914329199492446,
+        "fpr": 0.3092296372740129
+      },
+      {
+        "seed": 3,
+        "recall": 0.8908086962695508,
+        "fpr": 0.3108871821333122
+      },
+      {
+        "seed": 4,
+        "recall": 0.8915658027976695,
+        "fpr": 0.31225835663400164
+      }
+    ],
+    "target_fpr_operating_points": {
+      "1pct": {
+        "target_fpr": 0.01,
+        "threshold_selected_on_seed_0": 0.843994140625,
+        "calibration_recall": 0.10985399978867832,
+        "calibration_fpr": 0.009999404620147654,
+        "heldout_pooled_recall": 0.10933108172555146,
+        "heldout_pooled_fpr": 0.009888946956800414,
+        "heldout_by_seed": [
+          {
+            "seed": 1,
+            "recall": 0.10874532295402844,
+            "fpr": 0.009700648890535372
+          },
+          {
+            "seed": 2,
+            "recall": 0.10983708688285225,
+            "fpr": 0.010281291253943474
+          },
+          {
+            "seed": 3,
+            "recall": 0.10932804128709245,
+            "fpr": 0.00952361526406934
+          },
+          {
+            "seed": 4,
+            "recall": 0.10941387577823272,
+            "fpr": 0.01005023241865347
+          }
+        ]
+      },
+      "5pct": {
+        "target_fpr": 0.05,
+        "threshold_selected_on_seed_0": 0.572509765625,
+        "calibration_recall": 0.3038878497445494,
+        "calibration_fpr": 0.04996582078625422,
+        "heldout_pooled_recall": 0.30380968731802194,
+        "heldout_pooled_fpr": 0.04862683171184203,
+        "heldout_by_seed": [
+          {
+            "seed": 1,
+            "recall": 0.3031352254352892,
+            "fpr": 0.04915698623128964
+          },
+          {
+            "seed": 2,
+            "recall": 0.30405089792589207,
+            "fpr": 0.049297892796344805
+          },
+          {
+            "seed": 3,
+            "recall": 0.30364713965111845,
+            "fpr": 0.04801837739144241
+          },
+          {
+            "seed": 4,
+            "recall": 0.30440548625978797,
+            "fpr": 0.048034070428291276
+          }
+        ]
+      }
+    }
+  },
+  "cell_replicate_summaries": [
+    {
+      "papyrus": 35,
+      "pitch_um": 110.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.7834497074103327,
+          "sample_sd": 0.007692383125101641,
+          "min": 0.7703761194144649,
+          "max": 0.7898880261049689,
+          "values_by_seed": [
+            0.7866374036939415,
+            0.7831466621030402,
+            0.7898880261049689,
+            0.7872003257352478,
+            0.7703761194144649
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.051708984375,
+          "sample_sd": 0.003322465983633594,
+          "min": 0.047607421875,
+          "max": 0.055908203125,
+          "values_by_seed": [
+            0.053955078125,
+            0.047607421875,
+            0.049560546875,
+            0.055908203125,
+            0.051513671875
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.880294330651757,
+          "sample_sd": 0.00572925176417904,
+          "min": 0.870837233685434,
+          "max": 0.8839901121813247,
+          "values_by_seed": [
+            0.8839901121813247,
+            0.8788349430708892,
+            0.8839416336092427,
+            0.8838677307118943,
+            0.870837233685434
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.2991776819437587,
+          "sample_sd": 0.018644405275267877,
+          "min": 0.28347691611019427,
+          "max": 0.3277845532461277,
+          "values_by_seed": [
+            0.28362521794601236,
+            0.3068336139931555,
+            0.29416810842330393,
+            0.3277845532461277,
+            0.28347691611019427
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.5350977806642288,
+          "sample_sd": 0.01835502137962596,
+          "min": 0.5026196786200671,
+          "max": 0.5460144302253083,
+          "values_by_seed": [
+            0.5402498873915558,
+            0.5408075215229949,
+            0.5460144302253083,
+            0.5457973855612179,
+            0.5026196786200671
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.3462258024042684,
+          "sample_sd": 0.006916537082698021,
+          "min": 0.3379635393807473,
+          "max": 0.3565966915951557,
+          "values_by_seed": [
+            0.34763680904444105,
+            0.3379635393807473,
+            0.34638285332732116,
+            0.3565966915951557,
+            0.3425491186736768
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 35,
+      "pitch_um": 150.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.7603861898121363,
+          "sample_sd": 0.019592644827094288,
+          "min": 0.730358939221849,
+          "max": 0.779483683999293,
+          "values_by_seed": [
+            0.730358939221849,
+            0.7711033338024358,
+            0.779483683999293,
+            0.7517102102763772,
+            0.7692747817607265
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.062353515625,
+          "sample_sd": 0.006824156282135741,
+          "min": 0.05615234375,
+          "max": 0.0703125,
+          "values_by_seed": [
+            0.0703125,
+            0.05615234375,
+            0.059326171875,
+            0.069091796875,
+            0.056884765625
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.8280318453575737,
+          "sample_sd": 0.009231196943603448,
+          "min": 0.8143528001832432,
+          "max": 0.8367763833352335,
+          "values_by_seed": [
+            0.8143528001832432,
+            0.8254740480532107,
+            0.8367763833352335,
+            0.8363968246401384,
+            0.8271591705760423
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.13569557238322547,
+          "sample_sd": 0.02017623042354421,
+          "min": 0.10134888289638094,
+          "max": 0.15192334391420068,
+          "values_by_seed": [
+            0.14209540976102783,
+            0.14770568604344267,
+            0.10134888289638094,
+            0.15192334391420068,
+            0.13540453930107518
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.30925894599777387,
+          "sample_sd": 0.026434405475942405,
+          "min": 0.27634472494246387,
+          "max": 0.3486817509179483,
+          "values_by_seed": [
+            0.2977748905072135,
+            0.3143299439493597,
+            0.27634472494246387,
+            0.3486817509179483,
+            0.30916341967188415
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.3177514543361891,
+          "sample_sd": 0.005706823328737219,
+          "min": 0.31028950257603605,
+          "max": 0.32510247994113,
+          "values_by_seed": [
+            0.3157559681446238,
+            0.31028950257603605,
+            0.32148352005034087,
+            0.32510247994113,
+            0.3161258009688151
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 35,
+      "pitch_um": 200.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.7182435656173964,
+          "sample_sd": 0.008950684387106365,
+          "min": 0.7072436701642336,
+          "max": 0.7306229565845499,
+          "values_by_seed": [
+            0.7228392069647201,
+            0.7133532447536496,
+            0.7171587496198297,
+            0.7306229565845499,
+            0.7072436701642336
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.053515625,
+          "sample_sd": 0.002356937184794369,
+          "min": 0.051513671875,
+          "max": 0.05712890625,
+          "values_by_seed": [
+            0.051513671875,
+            0.051513671875,
+            0.054443359375,
+            0.052978515625,
+            0.05712890625
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.7731977187919482,
+          "sample_sd": 0.008156678972861422,
+          "min": 0.7620244674318167,
+          "max": 0.784017572036195,
+          "values_by_seed": [
+            0.7620244674318167,
+            0.773087650690336,
+            0.7699584749613332,
+            0.784017572036195,
+            0.7769004288400603
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.16580339492092458,
+          "sample_sd": 0.015338520817859391,
+          "min": 0.14329118194951337,
+          "max": 0.18033426475060826,
+          "values_by_seed": [
+            0.14329118194951337,
+            0.17607041704683699,
+            0.15707401916058394,
+            0.18033426475060826,
+            0.1722470916970803
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.2974688735553528,
+          "sample_sd": 0.020508968715963048,
+          "min": 0.27224115153588807,
+          "max": 0.3197477094738443,
+          "values_by_seed": [
+            0.27224115153588807,
+            0.2972530318582725,
+            0.28254923205596105,
+            0.3197477094738443,
+            0.31555324285279807
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.2733109983272506,
+          "sample_sd": 0.009910030067169747,
+          "min": 0.26341264636557177,
+          "max": 0.2898556303223844,
+          "values_by_seed": [
+            0.2729658512013382,
+            0.27092243575121655,
+            0.26341264636557177,
+            0.2898556303223844,
+            0.2693984279957421
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 35,
+      "pitch_um": 260.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.7777409649896063,
+          "sample_sd": 0.008839939285266937,
+          "min": 0.7628285030428079,
+          "max": 0.7864929358388987,
+          "values_by_seed": [
+            0.7864929358388987,
+            0.7805683190743453,
+            0.7795668428080536,
+            0.7628285030428079,
+            0.779248224183926
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.04443359375,
+          "sample_sd": 0.0037306413061367494,
+          "min": 0.037841796875,
+          "max": 0.046875,
+          "values_by_seed": [
+            0.04541015625,
+            0.04638671875,
+            0.037841796875,
+            0.046875,
+            0.045654296875
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.7120678635262703,
+          "sample_sd": 0.005890269238238613,
+          "min": 0.7031513659292832,
+          "max": 0.7189490737203206,
+          "values_by_seed": [
+            0.7189490737203206,
+            0.713108974961786,
+            0.7031513659292832,
+            0.7102903089319346,
+            0.714839594088027
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.05609913287458688,
+          "sample_sd": 0.002314543175016506,
+          "min": 0.052245730269058946,
+          "max": 0.05793645199817325,
+          "values_by_seed": [
+            0.05637256558474724,
+            0.052245730269058946,
+            0.05793645199817325,
+            0.05607301580251824,
+            0.05786790071843672
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.1452640720811879,
+          "sample_sd": 0.0044699232898128115,
+          "min": 0.14096894820130132,
+          "max": 0.1522053236344537,
+          "values_by_seed": [
+            0.1470745982750181,
+            0.14323934727595558,
+            0.14283214301921077,
+            0.14096894820130132,
+            0.1522053236344537
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.25920759548179484,
+          "sample_sd": 0.008548302252103852,
+          "min": 0.24736270675210795,
+          "max": 0.26778133300377227,
+          "values_by_seed": [
+            0.26778133300377227,
+            0.2646991805708294,
+            0.24736270675210795,
+            0.2533088061360154,
+            0.26288595094624906
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 50,
+      "pitch_um": 110.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.9182471828660184,
+          "sample_sd": 0.010328300779795402,
+          "min": 0.9065890721977563,
+          "max": 0.9277342219585745,
+          "values_by_seed": [
+            0.9065890721977563,
+            0.9277342219585745,
+            0.9227650288407412,
+            0.9264975721282924,
+            0.9076500192047272
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.01689453125,
+          "sample_sd": 0.0022269051271467536,
+          "min": 0.0146484375,
+          "max": 0.020263671875,
+          "values_by_seed": [
+            0.015380859375,
+            0.017822265625,
+            0.016357421875,
+            0.0146484375,
+            0.020263671875
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.9329754938547316,
+          "sample_sd": 0.0051521235412859715,
+          "min": 0.9243245995564835,
+          "max": 0.9371863837117801,
+          "values_by_seed": [
+            0.9243245995564835,
+            0.9364270024540349,
+            0.9326930601274963,
+            0.9371863837117801,
+            0.934246423423863
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.27440541058937606,
+          "sample_sd": 0.02653198106408333,
+          "min": 0.23108718946148407,
+          "max": 0.298888289058305,
+          "values_by_seed": [
+            0.23108718946148407,
+            0.2864484810959626,
+            0.2685332828257821,
+            0.2870698105053463,
+            0.298888289058305
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.5593033803684904,
+          "sample_sd": 0.03066653944501061,
+          "min": 0.5159298284746254,
+          "max": 0.5862174475066784,
+          "values_by_seed": [
+            0.5159298284746254,
+            0.5739610338729084,
+            0.5385289130893556,
+            0.5862174475066784,
+            0.5818796788988841
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.3208570548168773,
+          "sample_sd": 0.014282211180310554,
+          "min": 0.29766509936223,
+          "max": 0.33282609458532453,
+          "values_by_seed": [
+            0.32557276476444524,
+            0.3308705293726568,
+            0.29766509936223,
+            0.3173507859997298,
+            0.33282609458532453
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 50,
+      "pitch_um": 150.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.8663341725055712,
+          "sample_sd": 0.011435777039670803,
+          "min": 0.8544306256642658,
+          "max": 0.8797631079793387,
+          "values_by_seed": [
+            0.8544680590955419,
+            0.8544306256642658,
+            0.8696596353415185,
+            0.8797631079793387,
+            0.8733494344471912
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.021484375,
+          "sample_sd": 0.0018911051495153403,
+          "min": 0.018798828125,
+          "max": 0.02392578125,
+          "values_by_seed": [
+            0.022216796875,
+            0.02392578125,
+            0.020751953125,
+            0.018798828125,
+            0.021728515625
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.923686059310732,
+          "sample_sd": 0.002586496027595273,
+          "min": 0.920181410083895,
+          "max": 0.9272754595628556,
+          "values_by_seed": [
+            0.920181410083895,
+            0.9229620735113581,
+            0.9272754595628556,
+            0.9233480216042601,
+            0.9246633317912911
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.35759332415029665,
+          "sample_sd": 0.016982047705787654,
+          "min": 0.33006833259651075,
+          "max": 0.3738116662473175,
+          "values_by_seed": [
+            0.3546921858422972,
+            0.36126554375849657,
+            0.36812889230686124,
+            0.33006833259651075,
+            0.3738116662473175
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.614111882366153,
+          "sample_sd": 0.012553599162070397,
+          "min": 0.5942171034769024,
+          "max": 0.6282732337224989,
+          "values_by_seed": [
+            0.6140452129512091,
+            0.6199445321839848,
+            0.6282732337224989,
+            0.5942171034769024,
+            0.6140793294961697
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.35666326922739766,
+          "sample_sd": 0.012744410850476215,
+          "min": 0.3361590835537878,
+          "max": 0.3687010551963052,
+          "values_by_seed": [
+            0.3537468732423464,
+            0.3600029946745021,
+            0.3647063394700469,
+            0.3361590835537878,
+            0.3687010551963052
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 50,
+      "pitch_um": 200.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.8533053432937956,
+          "sample_sd": 0.004915032477850661,
+          "min": 0.8479112017183699,
+          "max": 0.8590893970498783,
+          "values_by_seed": [
+            0.8550624429744526,
+            0.8590893970498783,
+            0.8560190465328467,
+            0.8484446281934307,
+            0.8479112017183699
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.013671875,
+          "sample_sd": 0.0010500885335745394,
+          "min": 0.0126953125,
+          "max": 0.015380859375,
+          "values_by_seed": [
+            0.01318359375,
+            0.01318359375,
+            0.015380859375,
+            0.0126953125,
+            0.013916015625
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.8515864816761971,
+          "sample_sd": 0.0065692013176396336,
+          "min": 0.8416606483848806,
+          "max": 0.857794173671184,
+          "values_by_seed": [
+            0.8517382158570956,
+            0.8571891985705125,
+            0.857794173671184,
+            0.8495501718973122,
+            0.8416606483848806
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.13919475174878346,
+          "sample_sd": 0.006831348052252065,
+          "min": 0.131709293263382,
+          "max": 0.14636842305352799,
+          "values_by_seed": [
+            0.14636842305352799,
+            0.1442656059914842,
+            0.14140529957420925,
+            0.131709293263382,
+            0.13222513686131387
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.32092352874087593,
+          "sample_sd": 0.011612871353801686,
+          "min": 0.30162308964416057,
+          "max": 0.3305471601277372,
+          "values_by_seed": [
+            0.3276992092457421,
+            0.3258522943278589,
+            0.3305471601277372,
+            0.3188958903588808,
+            0.30162308964416057
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.2942360002281022,
+          "sample_sd": 0.006507609123985453,
+          "min": 0.2875004752128954,
+          "max": 0.30457843864051093,
+          "values_by_seed": [
+            0.2904686549574209,
+            0.29551137659671534,
+            0.30457843864051093,
+            0.2875004752128954,
+            0.29312105573296837
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 50,
+      "pitch_um": 260.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.9027666717195075,
+          "sample_sd": 0.005880731873285525,
+          "min": 0.8947448202749582,
+          "max": 0.9106822686804651,
+          "values_by_seed": [
+            0.9033518679258217,
+            0.8947448202749582,
+            0.9048583064358066,
+            0.9106822686804651,
+            0.9001960952804857
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.035009765625,
+          "sample_sd": 0.0016157609480248337,
+          "min": 0.03271484375,
+          "max": 0.03662109375,
+          "values_by_seed": [
+            0.0341796875,
+            0.036376953125,
+            0.03662109375,
+            0.03271484375,
+            0.03515625
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.785642882037195,
+          "sample_sd": 0.004378811661534529,
+          "min": 0.779891566409677,
+          "max": 0.789344207150282,
+          "values_by_seed": [
+            0.779891566409677,
+            0.789344207150282,
+            0.7893058314602889,
+            0.782079546535479,
+            0.787593258630248
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.04674806246083646,
+          "sample_sd": 0.004345535416891844,
+          "min": 0.042419000869925394,
+          "max": 0.05279269224047789,
+          "values_by_seed": [
+            0.042419000869925394,
+            0.05279269224047789,
+            0.04866151212536967,
+            0.047170763168845664,
+            0.042696343899563685
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.188971354254669,
+          "sample_sd": 0.003683411745320408,
+          "min": 0.18461921695128322,
+          "max": 0.19365277907853562,
+          "values_by_seed": [
+            0.1859410014859214,
+            0.18461921695128322,
+            0.19086148682898194,
+            0.18978228692862267,
+            0.19365277907853562
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.3172652191082347,
+          "sample_sd": 0.004056291401508754,
+          "min": 0.3128226617013655,
+          "max": 0.32202639513923154,
+          "values_by_seed": [
+            0.3209947466547458,
+            0.3144331340196829,
+            0.32202639513923154,
+            0.3128226617013655,
+            0.31604915802614797
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 65,
+      "pitch_um": 110.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.9442509122846318,
+          "sample_sd": 0.00572126996161444,
+          "min": 0.9364309108136315,
+          "max": 0.9519480419590841,
+          "values_by_seed": [
+            0.9519480419590841,
+            0.9419264913227803,
+            0.9465079845035398,
+            0.9364309108136315,
+            0.9444411328241237
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.012939453125,
+          "sample_sd": 0.0009134905729428568,
+          "min": 0.01171875,
+          "max": 0.013671875,
+          "values_by_seed": [
+            0.01171875,
+            0.013671875,
+            0.01220703125,
+            0.013427734375,
+            0.013671875
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.925949308046342,
+          "sample_sd": 0.0023056506510992494,
+          "min": 0.9223103339969563,
+          "max": 0.9282440681804124,
+          "values_by_seed": [
+            0.9273111483624373,
+            0.9282440681804124,
+            0.9223103339969563,
+            0.925269197928377,
+            0.9266117917635269
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.19443509979968435,
+          "sample_sd": 0.017017348253264248,
+          "min": 0.17211353430916101,
+          "max": 0.21340489958499523,
+          "values_by_seed": [
+            0.21340489958499523,
+            0.17211353430916101,
+            0.20097037953251992,
+            0.2040789494557539,
+            0.18160773611599174
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.4762754078060126,
+          "sample_sd": 0.004872150946319605,
+          "min": 0.46772355960940615,
+          "max": 0.4792187545818537,
+          "values_by_seed": [
+            0.47681732242357405,
+            0.4792187545818537,
+            0.46772355960940615,
+            0.4789209491125205,
+            0.4786964533027083
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.3348013212083162,
+          "sample_sd": 0.0066956774844904735,
+          "min": 0.32532303120299466,
+          "max": 0.343106752805116,
+          "values_by_seed": [
+            0.32532303120299466,
+            0.33300588352080573,
+            0.33882330545322914,
+            0.3337476330594356,
+            0.343106752805116
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 65,
+      "pitch_um": 150.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.9676016495349489,
+          "sample_sd": 0.0024379365061452957,
+          "min": 0.964896681360473,
+          "max": 0.9708246679678337,
+          "values_by_seed": [
+            0.9708246679678337,
+            0.964896681360473,
+            0.9684398267069052,
+            0.9654323585004452,
+            0.968414713139087
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.008154296875,
+          "sample_sd": 0.0007033623188910124,
+          "min": 0.007568359375,
+          "max": 0.009033203125,
+          "values_by_seed": [
+            0.007568359375,
+            0.009033203125,
+            0.0078125,
+            0.0087890625,
+            0.007568359375
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.9798949634267238,
+          "sample_sd": 0.0013473365683890396,
+          "min": 0.9787235280235818,
+          "max": 0.9814689740301623,
+          "values_by_seed": [
+            0.981241648424781,
+            0.9792046929297363,
+            0.9788359737253582,
+            0.9787235280235818,
+            0.9814689740301623
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.5215664422082312,
+          "sample_sd": 0.009032990872568643,
+          "min": 0.5129386996763193,
+          "max": 0.5331868691104538,
+          "values_by_seed": [
+            0.5283150738741659,
+            0.5201538750945905,
+            0.5132376932856271,
+            0.5129386996763193,
+            0.5331868691104538
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.8812906288958606,
+          "sample_sd": 0.014892334985102118,
+          "min": 0.8674209739989281,
+          "max": 0.9003299354202234,
+          "values_by_seed": [
+            0.894033300590927,
+            0.8745096338963038,
+            0.8674209739989281,
+            0.870159300572921,
+            0.9003299354202234
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.412491772937335,
+          "sample_sd": 0.0034803416095392082,
+          "min": 0.40682487266710354,
+          "max": 0.41517466012576687,
+          "values_by_seed": [
+            0.40682487266710354,
+            0.4141265240500319,
+            0.4114995500880633,
+            0.41517466012576687,
+            0.41483325775570945
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 65,
+      "pitch_um": 200.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.9534663454227494,
+          "sample_sd": 0.002098682965952258,
+          "min": 0.9513576832420925,
+          "max": 0.9565035260796837,
+          "values_by_seed": [
+            0.9517896517639902,
+            0.9531863024635037,
+            0.9513576832420925,
+            0.9565035260796837,
+            0.9544945635644769
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.012646484375,
+          "sample_sd": 0.0004011639825477291,
+          "min": 0.01220703125,
+          "max": 0.012939453125,
+          "values_by_seed": [
+            0.012939453125,
+            0.012939453125,
+            0.012939453125,
+            0.01220703125,
+            0.01220703125
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.8922975958186949,
+          "sample_sd": 0.0025878785758021903,
+          "min": 0.8880090956020595,
+          "max": 0.8949404114736833,
+          "values_by_seed": [
+            0.8933130298214454,
+            0.8949404114736833,
+            0.8929152432466,
+            0.8923101989496864,
+            0.8880090956020595
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.08816544061739659,
+          "sample_sd": 0.005585739408540686,
+          "min": 0.07999329949817519,
+          "max": 0.09336103824513382,
+          "values_by_seed": [
+            0.09336103824513382,
+            0.09035175258515815,
+            0.07999329949817519,
+            0.08495974946776155,
+            0.09216136329075426
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.2839723996350365,
+          "sample_sd": 0.011399589006420092,
+          "min": 0.2732027448296837,
+          "max": 0.30213085462287104,
+          "values_by_seed": [
+            0.28680119943734794,
+            0.30213085462287104,
+            0.27622866294099757,
+            0.28149853634428224,
+            0.2732027448296837
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.3171210652372262,
+          "sample_sd": 0.005614390108867826,
+          "min": 0.31271931075121656,
+          "max": 0.32494440009124087,
+          "values_by_seed": [
+            0.31377071928223843,
+            0.32494440009124087,
+            0.31294551208941607,
+            0.32122538397201944,
+            0.31271931075121656
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 65,
+      "pitch_um": 260.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.9313663139011374,
+          "sample_sd": 0.0037643988266445913,
+          "min": 0.9271777824336863,
+          "max": 0.9373769576942053,
+          "values_by_seed": [
+            0.9373769576942053,
+            0.9303929340044278,
+            0.9271777824336863,
+            0.9318648554967989,
+            0.9300190398765691
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.036669921875,
+          "sample_sd": 0.0017723331457892124,
+          "min": 0.03466796875,
+          "max": 0.039306640625,
+          "values_by_seed": [
+            0.03466796875,
+            0.03564453125,
+            0.039306640625,
+            0.037353515625,
+            0.036376953125
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.8079613747540144,
+          "sample_sd": 0.0038761165886561366,
+          "min": 0.8014965624532725,
+          "max": 0.8108617375611977,
+          "values_by_seed": [
+            0.8072368910198662,
+            0.8014965624532725,
+            0.8098152678475685,
+            0.8103964148881666,
+            0.8108617375611977
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.033527513683704394,
+          "sample_sd": 0.0014273141208396374,
+          "min": 0.03132359004017491,
+          "max": 0.034661844261147545,
+          "values_by_seed": [
+            0.034661844261147545,
+            0.034177157748080805,
+            0.03287299378886785,
+            0.03460198258025086,
+            0.03132359004017491
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.18338794907122669,
+          "sample_sd": 0.00368246093505422,
+          "min": 0.1781782931889959,
+          "max": 0.18815274576840674,
+          "values_by_seed": [
+            0.18815274576840674,
+            0.1781782931889959,
+            0.18186220034159778,
+            0.18462694103914085,
+            0.1841195650179923
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.34417082399603754,
+          "sample_sd": 0.005320906510067709,
+          "min": 0.3357726163706249,
+          "max": 0.34985676644578995,
+          "values_by_seed": [
+            0.34733774829321795,
+            0.3357726163706249,
+            0.34985676644578995,
+            0.34430097487643874,
+            0.3435860139941162
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 90,
+      "pitch_um": 110.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.92586009055786,
+          "sample_sd": 0.003008403884052391,
+          "min": 0.9207171655649387,
+          "max": 0.9285002478106852,
+          "values_by_seed": [
+            0.9273732019303755,
+            0.9285002478106852,
+            0.9207171655649387,
+            0.9263185042714294,
+            0.9263913332118716
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.025732421875,
+          "sample_sd": 0.0007445288418934036,
+          "min": 0.025146484375,
+          "max": 0.02685546875,
+          "values_by_seed": [
+            0.025390625,
+            0.026123046875,
+            0.02685546875,
+            0.025146484375,
+            0.025146484375
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.9005692332641138,
+          "sample_sd": 0.00273338662463309,
+          "min": 0.8977563566707059,
+          "max": 0.9036636467066524,
+          "values_by_seed": [
+            0.8978931208673863,
+            0.8977563566707059,
+            0.9028549526823652,
+            0.9006780893934592,
+            0.9036636467066524
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.06964581098991116,
+          "sample_sd": 0.009862125895304875,
+          "min": 0.05547618349431814,
+          "max": 0.08244187986157213,
+          "values_by_seed": [
+            0.0671766455134248,
+            0.05547618349431814,
+            0.08244187986157213,
+            0.07391752879627463,
+            0.06921681728396605
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.3780779601358897,
+          "sample_sd": 0.015886184421319137,
+          "min": 0.3554612331309666,
+          "max": 0.39863509431948685,
+          "values_by_seed": [
+            0.37272962388058556,
+            0.3554612331309666,
+            0.39863509431948685,
+            0.378673234571201,
+            0.3848906147772083
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.3621248167859411,
+          "sample_sd": 0.009995691398473555,
+          "min": 0.3491525162999822,
+          "max": 0.3769171677762464,
+          "values_by_seed": [
+            0.36316860212179786,
+            0.3626126505071346,
+            0.3769171677762464,
+            0.35877314722454434,
+            0.3491525162999822
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 90,
+      "pitch_um": 150.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.9677863527186383,
+          "sample_sd": 0.0022024557786521643,
+          "min": 0.964162227962015,
+          "max": 0.969668259245939,
+          "values_by_seed": [
+            0.964162227962015,
+            0.9675103877771791,
+            0.9682410504484193,
+            0.9693498381596398,
+            0.969668259245939
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.01455078125,
+          "sample_sd": 0.0004085254035810916,
+          "min": 0.01416015625,
+          "max": 0.01513671875,
+          "values_by_seed": [
+            0.0146484375,
+            0.01416015625,
+            0.01513671875,
+            0.01416015625,
+            0.0146484375
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.9800285884040416,
+          "sample_sd": 0.0006520135967188564,
+          "min": 0.979308049128558,
+          "max": 0.9809261311721641,
+          "values_by_seed": [
+            0.979308049128558,
+            0.9796954790067287,
+            0.9797512678516079,
+            0.9804620148611497,
+            0.9809261311721641
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.5688362893898019,
+          "sample_sd": 0.010426322719884189,
+          "min": 0.5593841110720982,
+          "max": 0.5809270507715788,
+          "values_by_seed": [
+            0.5791340367975368,
+            0.5593841110720982,
+            0.5601789791855907,
+            0.5645572691222049,
+            0.5809270507715788
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.8670746910675777,
+          "sample_sd": 0.007177386285870165,
+          "min": 0.8580192597373121,
+          "max": 0.8749351430265071,
+          "values_by_seed": [
+            0.8580192597373121,
+            0.8647774156053815,
+            0.8637761897789674,
+            0.8738654471897207,
+            0.8749351430265071
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.4475626334750592,
+          "sample_sd": 0.005884033560130355,
+          "min": 0.44061067667797726,
+          "max": 0.454524777851539,
+          "values_by_seed": [
+            0.454524777851539,
+            0.44061067667797726,
+            0.4426372942168192,
+            0.44838498436088103,
+            0.4516554342680793
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 90,
+      "pitch_um": 200.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.9758580919251825,
+          "sample_sd": 0.0009063152556215596,
+          "min": 0.9744017069647202,
+          "max": 0.976681540830292,
+          "values_by_seed": [
+            0.9744017069647202,
+            0.975650804060219,
+            0.9764909804592458,
+            0.9760654273114355,
+            0.976681540830292
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.015673828125,
+          "sample_sd": 0.0005828777532379552,
+          "min": 0.014892578125,
+          "max": 0.016357421875,
+          "values_by_seed": [
+            0.016357421875,
+            0.015625,
+            0.014892578125,
+            0.01611328125,
+            0.015380859375
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.9093025183613597,
+          "sample_sd": 0.0017662508083241544,
+          "min": 0.9064512890766016,
+          "max": 0.9111407880002165,
+          "values_by_seed": [
+            0.9091175695149515,
+            0.9095712260946959,
+            0.9111407880002165,
+            0.9102317191203324,
+            0.9064512890766016
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.08319362074209244,
+          "sample_sd": 0.0015858291324871814,
+          "min": 0.08109579341545012,
+          "max": 0.08508995780109489,
+          "values_by_seed": [
+            0.08508995780109489,
+            0.08109579341545012,
+            0.0841485610553528,
+            0.08215860705596106,
+            0.0834751843826034
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.3321389332420925,
+          "sample_sd": 0.0052442761214303545,
+          "min": 0.3268704379562044,
+          "max": 0.3409101277372263,
+          "values_by_seed": [
+            0.3303397296989051,
+            0.3307256025699513,
+            0.3409101277372263,
+            0.33184876824817516,
+            0.3268704379562044
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.3576241731295621,
+          "sample_sd": 0.001962464937252688,
+          "min": 0.3542132375304136,
+          "max": 0.35890857854318736,
+          "values_by_seed": [
+            0.3585371996654501,
+            0.3587557975973236,
+            0.35890857854318736,
+            0.35770605231143554,
+            0.3542132375304136
+          ]
+        }
+      }
+    },
+    {
+      "papyrus": 90,
+      "pitch_um": 260.0,
+      "metrics": {
+        "best_youden_recall": {
+          "n": 5,
+          "mean": 0.9458832059985266,
+          "sample_sd": 0.002911722003875682,
+          "min": 0.9435277454063401,
+          "max": 0.950525527627614,
+          "values_by_seed": [
+            0.94695893005935,
+            0.950525527627614,
+            0.9435277454063401,
+            0.944237154600515,
+            0.944166672298814
+          ]
+        },
+        "best_youden_threshold": {
+          "n": 5,
+          "mean": 0.04814453125,
+          "sample_sd": 0.0027642929403746412,
+          "min": 0.043701171875,
+          "max": 0.051025390625,
+          "values_by_seed": [
+            0.047607421875,
+            0.043701171875,
+            0.04931640625,
+            0.049072265625,
+            0.051025390625
+          ]
+        },
+        "near_sheet_auc": {
+          "n": 5,
+          "mean": 0.8467167568400467,
+          "sample_sd": 0.0011567145653170493,
+          "min": 0.8450371870144071,
+          "max": 0.8479436972595346,
+          "values_by_seed": [
+            0.8450371870144071,
+            0.8479436972595346,
+            0.8466357117094525,
+            0.8463215351301117,
+            0.8476456530867273
+          ]
+        },
+        "recall_at_fpr_1pct": {
+          "n": 5,
+          "mean": 0.030588739631618938,
+          "sample_sd": 0.000713486007457848,
+          "min": 0.029436016069964786,
+          "max": 0.031330831372541446,
+          "values_by_seed": [
+            0.03087993774385187,
+            0.031330831372541446,
+            0.03046838868768713,
+            0.030828524284049467,
+            0.029436016069964786
+          ]
+        },
+        "recall_at_fpr_5pct": {
+          "n": 5,
+          "mean": 0.2346922095815379,
+          "sample_sd": 0.0037242811766484475,
+          "min": 0.22944571945533596,
+          "max": 0.23964827400429267,
+          "values_by_seed": [
+            0.22944571945533596,
+            0.23964827400429267,
+            0.23383251860298285,
+            0.23626536490039307,
+            0.2342691709446849
+          ]
+        },
+        "surface_recall_at_0_5": {
+          "n": 5,
+          "mean": 0.4499706001905919,
+          "sample_sd": 0.0036286706339048076,
+          "min": 0.4458119030124908,
+          "max": 0.45415126274353806,
+          "values_by_seed": [
+            0.4469058269553287,
+            0.4458119030124908,
+            0.4528837882016489,
+            0.45010022003995287,
+            0.45415126274353806
+          ]
+        }
+      }
+    }
+  ],
+  "pitch_summaries": {
+    "surface_recall_at_0_5": {
+      "260": {
+        "n": 5,
+        "mean": 0.3426535596941648,
+        "sample_sd": 0.0024764162033103316,
+        "min": 0.34013316568844315,
+        "max": 0.3457549137267662,
+        "values_by_seed": [
+          0.3457549137267662,
+          0.340179208493407,
+          0.34303241413469454,
+          0.34013316568844315,
+          0.34416809642751284
+        ]
+      },
+      "200": {
+        "n": 5,
+        "mean": 0.31057305923053524,
+        "sample_sd": 0.002713165156488845,
+        "min": 0.30736300800258515,
+        "max": 0.3140718854546837,
+        "values_by_seed": [
+          0.30893560627661193,
+          0.3125335025091241,
+          0.30996129390967153,
+          0.3140718854546837,
+          0.30736300800258515
+        ]
+      },
+      "150": {
+        "n": 5,
+        "mean": 0.3836172824939953,
+        "sample_sd": 0.0028322316846142006,
+        "min": 0.3812053019953915,
+        "max": 0.3878288870472273,
+        "values_by_seed": [
+          0.38271312297640314,
+          0.38125742449463684,
+          0.3850816759563176,
+          0.3812053019953915,
+          0.3878288870472273
+        ]
+      },
+      "110": {
+        "n": 5,
+        "mean": 0.34100224880385077,
+        "sample_sd": 0.0008151687766335423,
+        "min": 0.3399471064797567,
+        "max": 0.3419086205910249,
+        "values_by_seed": [
+          0.34042530178341973,
+          0.3411131506953361,
+          0.3399471064797567,
+          0.34161706446971635,
+          0.3419086205910249
+        ]
+      }
+    },
+    "near_sheet_auc": {
+      "260": {
+        "n": 5,
+        "mean": 0.7880972192893816,
+        "sample_sd": 0.0012374820041921167,
+        "min": 0.7872270442366482,
+        "max": 0.7902350608415499,
+        "values_by_seed": [
+          0.7877786795410677,
+          0.7879733604562188,
+          0.7872270442366482,
+          0.7872719513714228,
+          0.7902350608415499
+        ]
+      },
+      "200": {
+        "n": 5,
+        "mean": 0.8565960786620499,
+        "sample_sd": 0.0027302106047320304,
+        "min": 0.8532553654759004,
+        "max": 0.8590274155008815,
+        "values_by_seed": [
+          0.8540483206563273,
+          0.8586971217073069,
+          0.8579521699698334,
+          0.8590274155008815,
+          0.8532553654759004
+        ]
+      },
+      "150": {
+        "n": 5,
+        "mean": 0.9279103641247678,
+        "sample_sd": 0.002719718968150337,
+        "min": 0.9237709769551192,
+        "max": 0.9306597711187639,
+        "values_by_seed": [
+          0.9237709769551192,
+          0.9268340733752585,
+          0.9306597711187639,
+          0.9297325972822825,
+          0.9285544018924149
+        ]
+      },
+      "110": {
+        "n": 5,
+        "mean": 0.909947091454236,
+        "sample_sd": 0.001353068222063455,
+        "min": 0.9083797452419079,
+        "max": 0.9117503504363775,
+        "values_by_seed": [
+          0.9083797452419079,
+          0.9103155925940105,
+          0.9104499951040151,
+          0.9117503504363775,
+          0.9088397738948691
+        ]
+      }
+    },
+    "recall_at_fpr_1pct": {
+      "260": {
+        "n": 5,
+        "mean": 0.041740862162686666,
+        "sample_sd": 0.0009948543229872336,
+        "min": 0.040330962682035024,
+        "max": 0.042636602907539775,
+        "values_by_seed": [
+          0.04108333711491801,
+          0.042636602907539775,
+          0.04248483665002448,
+          0.04216857145891606,
+          0.040330962682035024
+        ]
+      },
+      "200": {
+        "n": 5,
+        "mean": 0.11908930200729928,
+        "sample_sd": 0.0028410997783414427,
+        "min": 0.11565529482208028,
+        "max": 0.12294589225973238,
+        "values_by_seed": [
+          0.11702765026231751,
+          0.12294589225973238,
+          0.11565529482208028,
+          0.11979047863442822,
+          0.12002719405793796
+        ]
+      },
+      "150": {
+        "n": 5,
+        "mean": 0.39592290703288885,
+        "sample_sd": 0.00816505757626073,
+        "min": 0.385723611918615,
+        "max": 0.40583253135760633,
+        "values_by_seed": [
+          0.40105917656875695,
+          0.397127303992157,
+          0.385723611918615,
+          0.3898719113273089,
+          0.40583253135760633
+        ]
+      },
+      "110": {
+        "n": 5,
+        "mean": 0.2094160008306826,
+        "sample_sd": 0.009026178450761911,
+        "min": 0.19882348812647913,
+        "max": 0.22321271050087566,
+        "values_by_seed": [
+          0.19882348812647913,
+          0.2052179532231493,
+          0.21152841266079453,
+          0.22321271050087566,
+          0.20829743964211428
+        ]
+      }
+    },
+    "recall_at_fpr_5pct": {
+      "260": {
+        "n": 5,
+        "mean": 0.18807889624715537,
+        "sample_sd": 0.0017599416457063125,
+        "min": 0.18642128285513185,
+        "max": 0.19106170966891664,
+        "values_by_seed": [
+          0.18765351624617058,
+          0.18642128285513185,
+          0.18734708719819332,
+          0.18791088526736446,
+          0.19106170966891664
+        ]
+      },
+      "200": {
+        "n": 5,
+        "mean": 0.3086259337933394,
+        "sample_sd": 0.0046531561244446355,
+        "min": 0.30427032247947083,
+        "max": 0.31399044584473845,
+        "values_by_seed": [
+          0.30427032247947083,
+          0.31399044584473845,
+          0.3075587957154805,
+          0.3129977261062956,
+          0.30431237882071166
+        ]
+      },
+      "150": {
+        "n": 5,
+        "mean": 0.6679340370818413,
+        "sample_sd": 0.005998525935781186,
+        "min": 0.6589537806107145,
+        "max": 0.6746269569036961,
+        "values_by_seed": [
+          0.6659681659466654,
+          0.6683903814087574,
+          0.6589537806107145,
+          0.6717309005393731,
+          0.6746269569036961
+        ]
+      },
+      "110": {
+        "n": 5,
+        "mean": 0.4871886322436553,
+        "sample_sd": 0.0074225439234899175,
+        "min": 0.4764316655425852,
+        "max": 0.49740225418790446,
+        "values_by_seed": [
+          0.4764316655425852,
+          0.4873621357771809,
+          0.48772549931088927,
+          0.49740225418790446,
+          0.4870216063997169
+        ]
+      }
+    },
+    "best_youden_recall": {
+      "260": {
+        "n": 5,
+        "mean": 0.8894392891521944,
+        "sample_sd": 0.002379345415467133,
+        "min": 0.8874031954551467,
+        "max": 0.8935451728795689,
+        "values_by_seed": [
+          0.8935451728795689,
+          0.8890579002453363,
+          0.8887826692709716,
+          0.8874031954551467,
+          0.8884075079099487
+        ]
+      },
+      "200": {
+        "n": 5,
+        "mean": 0.875218336564781,
+        "sample_sd": 0.00229769469121162,
+        "min": 0.8715827440693431,
+        "max": 0.8779091345422749,
+        "values_by_seed": [
+          0.8760232521669709,
+          0.8753199370818125,
+          0.8752566149635037,
+          0.8779091345422749,
+          0.8715827440693431
+        ]
+      },
+      "150": {
+        "n": 5,
+        "mean": 0.8905270911428238,
+        "sample_sd": 0.0065332618957594174,
+        "min": 0.87995347356181,
+        "max": 0.8964560491240341,
+        "values_by_seed": [
+          0.87995347356181,
+          0.8894852571510884,
+          0.8964560491240341,
+          0.8915638787289502,
+          0.8951767971482358
+        ]
+      },
+      "110": {
+        "n": 5,
+        "mean": 0.8929519732797108,
+        "sample_sd": 0.003316898510008606,
+        "min": 0.8872146511637968,
+        "max": 0.8953269057987701,
+        "values_by_seed": [
+          0.8931369299452894,
+          0.8953269057987701,
+          0.8949695512535472,
+          0.8941118282371503,
+          0.8872146511637968
+        ]
+      }
+    },
+    "best_youden_threshold": {
+      "260": {
+        "n": 5,
+        "mean": 0.041064453125,
+        "sample_sd": 0.0006897242718025708,
+        "min": 0.04046630859375,
+        "max": 0.04205322265625,
+        "values_by_seed": [
+          0.04046630859375,
+          0.04052734375,
+          0.040771484375,
+          0.04150390625,
+          0.04205322265625
+        ]
+      },
+      "200": {
+        "n": 5,
+        "mean": 0.023876953125,
+        "sample_sd": 0.000612484067592876,
+        "min": 0.0233154296875,
+        "max": 0.024658203125,
+        "values_by_seed": [
+          0.02349853515625,
+          0.0233154296875,
+          0.0244140625,
+          0.02349853515625,
+          0.024658203125
+        ]
+      },
+      "150": {
+        "n": 5,
+        "mean": 0.0266357421875,
+        "sample_sd": 0.0014866790958634798,
+        "min": 0.02520751953125,
+        "max": 0.0286865234375,
+        "values_by_seed": [
+          0.0286865234375,
+          0.02581787109375,
+          0.0257568359375,
+          0.0277099609375,
+          0.02520751953125
+        ]
+      },
+      "110": {
+        "n": 5,
+        "mean": 0.02681884765625,
+        "sample_sd": 0.0006203406224253863,
+        "min": 0.0262451171875,
+        "max": 0.02764892578125,
+        "values_by_seed": [
+          0.026611328125,
+          0.02630615234375,
+          0.0262451171875,
+          0.02728271484375,
+          0.02764892578125
+        ]
+      }
+    }
+  },
+  "papyrus_summaries": {
+    "surface_recall_at_0_5": {
+      "35": {
+        "n": 5,
+        "mean": 0.29912396263737573,
+        "sample_sd": 0.004629861871607253,
+        "min": 0.29466043162383543,
+        "max": 0.3062159019986714,
+        "values_by_seed": [
+          0.30103499034854386,
+          0.29596866456970733,
+          0.29466043162383543,
+          0.3062159019986714,
+          0.29773982464612075
+        ]
+      },
+      "50": {
+        "n": 5,
+        "mean": 0.322255385845153,
+        "sample_sd": 0.005376942141434111,
+        "min": 0.3134582516169446,
+        "max": 0.32767434088518654,
+        "values_by_seed": [
+          0.3226957599047396,
+          0.3252045086658893,
+          0.3222440681530048,
+          0.3134582516169446,
+          0.32767434088518654
+        ]
+      },
+      "65": {
+        "n": 5,
+        "mean": 0.35214624584472876,
+        "sample_sd": 0.0022449337189025774,
+        "min": 0.3483140928613887,
+        "max": 0.35361216300841514,
+        "values_by_seed": [
+          0.3483140928613887,
+          0.3519623560081759,
+          0.3532812835191246,
+          0.35361216300841514,
+          0.35356133382653954
+        ]
+      },
+      "90": {
+        "n": 5,
+        "mean": 0.40432055589528854,
+        "sample_sd": 0.0024789995342160195,
+        "min": 0.4019477569487316,
+        "max": 0.4078367071844754,
+        "values_by_seed": [
+          0.4057841016485289,
+          0.4019477569487316,
+          0.4078367071844754,
+          0.40374110098420346,
+          0.4022931127105033
+        ]
+      }
+    },
+    "near_sheet_auc": {
+      "35": {
+        "n": 5,
+        "mean": 0.7983979395818872,
+        "sample_sd": 0.0032317196643874215,
+        "min": 0.7948291133791763,
+        "max": 0.8036431090800407,
+        "values_by_seed": [
+          0.7948291133791763,
+          0.7976264041940555,
+          0.7984569644587731,
+          0.8036431090800407,
+          0.7974341067973909
+        ]
+      },
+      "50": {
+        "n": 5,
+        "mean": 0.8734727292197139,
+        "sample_sd": 0.0032341817596765357,
+        "min": 0.8690339479767878,
+        "max": 0.8767671312054561,
+        "values_by_seed": [
+          0.8690339479767878,
+          0.8764806204215468,
+          0.8767671312054561,
+          0.8730410309372079,
+          0.8720409155575707
+        ]
+      },
+      "65": {
+        "n": 5,
+        "mean": 0.9015258105114438,
+        "sample_sd": 0.0005582645188180607,
+        "min": 0.9009692047041208,
+        "max": 0.9022756794071325,
+        "values_by_seed": [
+          0.9022756794071325,
+          0.9009714337592761,
+          0.9009692047041208,
+          0.9016748349474528,
+          0.9017378997392367
+        ]
+      },
+      "90": {
+        "n": 5,
+        "mean": 0.9091542742173905,
+        "sample_sd": 0.0008841055415507703,
+        "min": 0.9078389816313257,
+        "max": 0.9100956800609106,
+        "values_by_seed": [
+          0.9078389816313257,
+          0.9087416897579162,
+          0.9100956800609106,
+          0.9094233396262632,
+          0.9096716800105363
+        ]
+      }
+    },
+    "recall_at_fpr_1pct": {
+      "35": {
+        "n": 5,
+        "mean": 0.1641939455306239,
+        "sample_sd": 0.010742845253915527,
+        "min": 0.1526318656196105,
+        "max": 0.17902879442836372,
+        "values_by_seed": [
+          0.15634609381032522,
+          0.17071386183812354,
+          0.1526318656196105,
+          0.17902879442836372,
+          0.1622491119566966
+        ]
+      },
+      "50": {
+        "n": 5,
+        "mean": 0.20448538723732318,
+        "sample_sd": 0.007945131535713587,
+        "min": 0.19364169980680868,
+        "max": 0.21190535901662502,
+        "values_by_seed": [
+          0.19364169980680868,
+          0.2111930807716053,
+          0.20668224670805554,
+          0.19900454988352118,
+          0.21190535901662502
+        ]
+      },
+      "65": {
+        "n": 5,
+        "mean": 0.20942362407725418,
+        "sample_sd": 0.004965827105444433,
+        "min": 0.20419907993424763,
+        "max": 0.21743571399136064,
+        "values_by_seed": [
+          0.21743571399136064,
+          0.20419907993424763,
+          0.2067685915262975,
+          0.2091448452950214,
+          0.2095698896393437
+        ]
+      },
+      "90": {
+        "n": 5,
+        "mean": 0.18806611518835611,
+        "sample_sd": 0.0036785179576858144,
+        "min": 0.18182172983860198,
+        "max": 0.19076376712702825,
+        "values_by_seed": [
+          0.1905701444639771,
+          0.18182172983860198,
+          0.1893094521975507,
+          0.18786548231462255,
+          0.19076376712702825
+        ]
+      }
+    },
+    "recall_at_fpr_5pct": {
+      "35": {
+        "n": 5,
+        "mean": 0.3217724180746358,
+        "sample_sd": 0.010607866611533017,
+        "min": 0.31193513256073596,
+        "max": 0.33879894853857795,
+        "values_by_seed": [
+          0.31433513192741885,
+          0.3239074611516457,
+          0.31193513256073596,
+          0.33879894853857795,
+          0.3198854161948007
+        ]
+      },
+      "50": {
+        "n": 5,
+        "mean": 0.42082753643254706,
+        "sample_sd": 0.005782500359297558,
+        "min": 0.4109038130393745,
+        "max": 0.42609426933400885,
+        "values_by_seed": [
+          0.4109038130393745,
+          0.42609426933400885,
+          0.4220526984421434,
+          0.42227818206777107,
+          0.4228087192794375
+        ]
+      },
+      "65": {
+        "n": 5,
+        "mean": 0.4562315963520341,
+        "sample_sd": 0.005226421423965909,
+        "min": 0.4483088492227324,
+        "max": 0.4614511420550639,
+        "values_by_seed": [
+          0.4614511420550639,
+          0.45850938407250613,
+          0.4483088492227324,
+          0.45380143176721616,
+          0.4590871746426519
+        ]
+      },
+      "90": {
+        "n": 5,
+        "mean": 0.4529959485067745,
+        "sample_sd": 0.005163221578624922,
+        "min": 0.4476335831930347,
+        "max": 0.4592884826096658,
+        "values_by_seed": [
+          0.4476335831930347,
+          0.447653131327648,
+          0.4592884826096658,
+          0.4551632037273725,
+          0.4552413416761512
+        ]
+      }
+    },
+    "best_youden_recall": {
+      "35": {
+        "n": 5,
+        "mean": 0.7599551069573679,
+        "sample_sd": 0.0043028948946480065,
+        "min": 0.7565356988808377,
+        "max": 0.7665243256330363,
+        "values_by_seed": [
+          0.7565821214298524,
+          0.7620428899333677,
+          0.7665243256330363,
+          0.7580904989097457,
+          0.7565356988808377
+        ]
+      },
+      "50": {
+        "n": 5,
+        "mean": 0.8851633425962231,
+        "sample_sd": 0.0046358492838186145,
+        "min": 0.879867860548393,
+        "max": 0.8913468942453817,
+        "values_by_seed": [
+          0.879867860548393,
+          0.8839997662369192,
+          0.8883255042877282,
+          0.8913468942453817,
+          0.8822766876626935
+        ]
+      },
+      "65": {
+        "n": 5,
+        "mean": 0.9491713052858669,
+        "sample_sd": 0.0022518409969842427,
+        "min": 0.9475579127226398,
+        "max": 0.9529848298462783,
+        "values_by_seed": [
+          0.9529848298462783,
+          0.9476006022877962,
+          0.9483708192215559,
+          0.9475579127226398,
+          0.9493423623510642
+        ]
+      },
+      "90": {
+        "n": 5,
+        "mean": 0.9538469353000518,
+        "sample_sd": 0.0012259369007595912,
+        "min": 0.952244235469736,
+        "max": 0.9555467418189243,
+        "values_by_seed": [
+          0.9532240167291152,
+          0.9555467418189243,
+          0.952244235469736,
+          0.953992731085755,
+          0.9542269513967292
+        ]
+      }
+    },
+    "best_youden_threshold": {
+      "35": {
+        "n": 5,
+        "mean": 0.0530029296875,
+        "sample_sd": 0.002722947902738638,
+        "min": 0.05029296875,
+        "max": 0.05621337890625,
+        "values_by_seed": [
+          0.0552978515625,
+          0.0504150390625,
+          0.05029296875,
+          0.05621337890625,
+          0.05279541015625
+        ]
+      },
+      "50": {
+        "n": 5,
+        "mean": 0.02176513671875,
+        "sample_sd": 0.0013109067147133243,
+        "min": 0.01971435546875,
+        "max": 0.0228271484375,
+        "values_by_seed": [
+          0.021240234375,
+          0.0228271484375,
+          0.02227783203125,
+          0.01971435546875,
+          0.02276611328125
+        ]
+      },
+      "65": {
+        "n": 5,
+        "mean": 0.0176025390625,
+        "sample_sd": 0.0005418052007122105,
+        "min": 0.0167236328125,
+        "max": 0.01806640625,
+        "values_by_seed": [
+          0.0167236328125,
+          0.017822265625,
+          0.01806640625,
+          0.0179443359375,
+          0.0174560546875
+        ]
+      },
+      "90": {
+        "n": 5,
+        "mean": 0.026025390625,
+        "sample_sd": 0.0006749837061010377,
+        "min": 0.02490234375,
+        "max": 0.02655029296875,
+        "values_by_seed": [
+          0.0260009765625,
+          0.02490234375,
+          0.02655029296875,
+          0.026123046875,
+          0.02655029296875
+        ]
+      }
+    }
+  },
+  "limitations": [
+    "Five seeds estimate variation across this analytic phantom generator, not across real scrolls.",
+    "Per-cell best-Youden metrics are descriptive upper bounds chosen and measured on the same cell.",
+    "The global threshold is selected on seed 0 and evaluated on seeds 1-4 to avoid calibration leakage.",
+    "Histogram-derived thresholds have resolution 1/4096.",
+    "No confidence interval here supports claims of real-scroll accuracy."
+  ]
+}

surface-contrast-grid/five-seed-addendum/runtime/kernel-metadata.json

@@ -0,0 +1,23 @@
+{
+  "id": "aviadcohen1/vesuvius-surface-contrast-grid",
+  "id_no": 129501090,
+  "title": "Vesuvius Surface Contrast Grid",
+  "code_file": "vesuvius-surface-contrast-grid.py",
+  "language": "python",
+  "kernel_type": "script",
+  "is_private": true,
+  "enable_gpu": true,
+  "enable_tpu": false,
+  "enable_internet": true,
+  "keywords": [
+    "gpu"
+  ],
+  "dataset_sources": [
+    "aviadcohen1/vesuvius-surface-contrast-harness"
+  ],
+  "kernel_sources": [],
+  "competition_sources": [],
+  "model_sources": [],
+  "docker_image": "gcr.io/kaggle-private-byod/python@sha256:37c64f7dd9c54116ecd1bcc88817c5469b88387388fade02bfa8bf3fc647d461",
+  "machine_shape": "Gpu"
+}

surface-contrast-grid/five-seed-addendum/runtime/vesuvius-surface-contrast-grid.py

@@ -0,0 +1,397 @@
+#!/usr/bin/env python3
+"""Run five seeded Vesuvius contrast/geometry grids on a Kaggle GPU."""
+
+from __future__ import annotations
+
+import os
+import json
+import shutil
+import subprocess
+import sys
+from pathlib import Path
+
+
+GRID_COMMIT = "130d83f"
+VILLA_COMMIT = "4d5c9e60"
+HARNESS_SHA256 = "54cbc7521ee34f559724427f34ca723b6ecec004954c389b57b6a82cbe1be881"
+
+
+def run(command: list[str], **kwargs) -> None:
+    print("+", " ".join(map(str, command)), flush=True)
+    subprocess.run(command, check=True, **kwargs)
+
+
+def sha256(path: Path) -> str:
+    import hashlib
+
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(8 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def aggregate_replicates(runs_dir: Path, output: Path) -> None:
+    """Aggregate five seeds and calibrate one threshold on seed 0 only."""
+    import numpy as np
+
+    paths = sorted(
+        runs_dir.glob("seed_*.json"),
+        key=lambda path: int(path.stem.split("_")[-1]),
+    )
+    if len(paths) != 5:
+        raise RuntimeError(f"expected five seed runs, found {len(paths)}")
+    runs = []
+    for path in paths:
+        seed = int(path.stem.split("_")[-1])
+        result = json.loads(path.read_text(encoding="utf-8"))
+        if not result.get("canonical") or len(result.get("cells", [])) != 16:
+            raise RuntimeError(f"non-canonical seed run: {path}")
+        runs.append((seed, path, result))
+
+    bins = int(runs[0][2]["cells"][0]["metrics"]["histogram_bins"])
+
+    def pooled_histogram(result: dict) -> tuple[np.ndarray, np.ndarray]:
+        positive = np.zeros(bins, dtype=np.int64)
+        negative = np.zeros(bins, dtype=np.int64)
+        for cell in result["cells"]:
+            metrics = cell["metrics"]
+            if int(metrics["histogram_bins"]) != bins:
+                raise RuntimeError("histogram bin mismatch")
+            positive += np.asarray(metrics["positive_score_histogram"], dtype=np.int64)
+            negative += np.asarray(metrics["negative_score_histogram"], dtype=np.int64)
+        return positive, negative
+
+    def rates(positive: np.ndarray, negative: np.ndarray, index: int) -> tuple[float, float]:
+        return (
+            float(positive[index:].sum() / positive.sum()),
+            float(negative[index:].sum() / negative.sum()),
+        )
+
+    calibration_positive, calibration_negative = pooled_histogram(runs[0][2])
+    calibration_tpr = np.cumsum(calibration_positive[::-1])[::-1] / calibration_positive.sum()
+    calibration_fpr_curve = np.cumsum(calibration_negative[::-1])[::-1] / calibration_negative.sum()
+    youden = calibration_tpr - calibration_fpr_curve
+    candidates = np.flatnonzero(np.isclose(youden, youden.max(), rtol=0.0, atol=1e-15))
+    threshold_index = int(candidates[-1])
+    threshold = threshold_index / bins
+    calibration_recall, calibration_fpr = rates(
+        calibration_positive, calibration_negative, threshold_index
+    )
+
+    heldout_positive = np.zeros(bins, dtype=np.int64)
+    heldout_negative = np.zeros(bins, dtype=np.int64)
+    heldout_by_seed = []
+    for seed, _, result in runs[1:]:
+        positive, negative = pooled_histogram(result)
+        recall, fpr = rates(positive, negative, threshold_index)
+        heldout_by_seed.append({"seed": seed, "recall": recall, "fpr": fpr})
+        heldout_positive += positive
+        heldout_negative += negative
+    heldout_recall, heldout_fpr = rates(
+        heldout_positive, heldout_negative, threshold_index
+    )
+
+    def summary(values: list[float]) -> dict:
+        array = np.asarray(values, dtype=np.float64)
+        return {
+            "n": int(array.size),
+            "mean": float(array.mean()),
+            "sample_sd": float(array.std(ddof=1)) if array.size > 1 else 0.0,
+            "min": float(array.min()),
+            "max": float(array.max()),
+            "values_by_seed": [float(value) for value in array],
+        }
+
+    metric_names = (
+        "surface_recall_at_0_5",
+        "near_sheet_auc",
+        "best_youden_recall",
+        "best_youden_threshold",
+    )
+    cells = {}
+    pitch_seed_means = {metric: {} for metric in metric_names}
+    papyrus_seed_means = {metric: {} for metric in metric_names}
+    for seed_index, (_, _, result) in enumerate(runs):
+        per_pitch = {metric: {} for metric in metric_names}
+        per_papyrus = {metric: {} for metric in metric_names}
+        for cell in result["cells"]:
+            papyrus = int(cell["papyrus"])
+            pitch = float(cell["pitch_um"])
+            key = (papyrus, pitch)
+            cells.setdefault(key, {metric: [] for metric in metric_names})
+            for metric in metric_names:
+                value = float(cell["metrics"][metric])
+                cells[key][metric].append(value)
+                per_pitch[metric].setdefault(pitch, []).append(value)
+                per_papyrus[metric].setdefault(papyrus, []).append(value)
+        for metric in metric_names:
+            for pitch, values in per_pitch[metric].items():
+                pitch_seed_means[metric].setdefault(pitch, []).append(float(np.mean(values)))
+            for papyrus, values in per_papyrus[metric].items():
+                papyrus_seed_means[metric].setdefault(papyrus, []).append(float(np.mean(values)))
+
+    aggregate = {
+        "schema_version": "2.0",
+        "replicate_design": {
+            "seeds": [seed for seed, _, _ in runs],
+            "replicates_per_cell": 5,
+            "cells_per_seed": 16,
+            "total_inferences": 80,
+            "seed_0_role": "replicate summary and global-threshold calibration only",
+            "heldout_threshold_seeds": [1, 2, 3, 4],
+        },
+        "source_runs": [
+            {"seed": seed, "file": path.name, "sha256": sha256(path)}
+            for seed, path, _ in runs
+        ],
+        "global_threshold_calibration": {
+            "histogram_bins": bins,
+            "selected_on_seed": 0,
+            "threshold": threshold,
+            "selection_rule": "pooled maximum Youden J; highest threshold on ties",
+            "calibration_recall": calibration_recall,
+            "calibration_fpr": calibration_fpr,
+            "heldout_pooled_recall": heldout_recall,
+            "heldout_pooled_fpr": heldout_fpr,
+            "heldout_by_seed": heldout_by_seed,
+        },
+        "cell_replicate_summaries": [
+            {
+                "papyrus": papyrus,
+                "pitch_um": pitch,
+                "metrics": {metric: summary(values) for metric, values in metrics.items()},
+            }
+            for (papyrus, pitch), metrics in sorted(cells.items())
+        ],
+        "pitch_summaries": {
+            metric: {
+                str(int(pitch)): summary(values)
+                for pitch, values in sorted(mapping.items(), reverse=True)
+            }
+            for metric, mapping in pitch_seed_means.items()
+        },
+        "papyrus_summaries": {
+            metric: {
+                str(papyrus): summary(values)
+                for papyrus, values in sorted(mapping.items())
+            }
+            for metric, mapping in papyrus_seed_means.items()
+        },
+        "limitations": [
+            "Five seeds estimate variation across this analytic generator, not real scrolls.",
+            "Per-cell best-Youden metrics are same-cell descriptive upper bounds.",
+            "The global threshold is selected on seed 0 and evaluated on seeds 1-4.",
+            "Histogram-derived thresholds have resolution 1/4096.",
+        ],
+    }
+    output.write_text(json.dumps(aggregate, indent=2) + "\n", encoding="utf-8")
+
+
+def main() -> int:
+    working = Path("/kaggle/working/vesuvius-contrast")
+    working.mkdir(parents=True, exist_ok=True)
+    harness_candidates = sorted(
+        Path("/kaggle/input").glob("**/run_surface_grid.py")
+    )
+    if len(harness_candidates) != 1:
+        raise RuntimeError(
+            "expected exactly one mounted run_surface_grid.py; "
+            f"found {len(harness_candidates)}: {harness_candidates}"
+        )
+    harness = harness_candidates[0]
+    print(f"HARNESS_PATH={harness}", flush=True)
+    actual_harness = sha256(harness)
+    if actual_harness != HARNESS_SHA256:
+        raise RuntimeError(
+            f"harness SHA-256 mismatch: {actual_harness}; expected {HARNESS_SHA256}"
+        )
+
+    run(
+        [
+            sys.executable,
+            "-m",
+            "pip",
+            "install",
+            "--quiet",
+            "--no-cache-dir",
+            "torch==2.5.1",
+            "--index-url",
+            "https://download.pytorch.org/whl/cu121",
+        ]
+    )
+    run(
+        [
+            sys.executable,
+            "-c",
+            (
+                "import torch; "
+                "print('TORCH_VERSION=' + torch.__version__); "
+                "print('CUDA_ARCH_LIST=' + ','.join(torch.cuda.get_arch_list())); "
+                "assert 'sm_60' in torch.cuda.get_arch_list()"
+            ),
+        ]
+    )
+    run(
+        [
+            sys.executable,
+            "-m",
+            "pip",
+            "install",
+            "--quiet",
+            "huggingface-hub==1.11.0",
+            "dynamic-network-architectures==0.3.1",
+        ]
+    )
+    run(
+        [
+            "git",
+            "clone",
+            "--filter=blob:none",
+            "https://github.com/Diego-dcv/vesuvius-topological-grid.git",
+            str(working / "topological-grid"),
+        ]
+    )
+    run(["git", "-C", str(working / "topological-grid"), "checkout", GRID_COMMIT])
+    run(
+        [
+            "git",
+            "clone",
+            "--filter=blob:none",
+            "https://github.com/ScrollPrize/villa.git",
+            str(working / "villa"),
+        ]
+    )
+    run(["git", "-C", str(working / "villa"), "checkout", VILLA_COMMIT])
+
+    from huggingface_hub import hf_hub_download
+
+    checkpoint = Path(
+        hf_hub_download(
+            "scrollprize/surface_recto_059_redo",
+            "Model_epoch499.pth",
+            cache_dir=working / "hf-cache",
+        )
+    )
+    run(
+        [
+            sys.executable,
+            str(working / "topological-grid/scripts/contrast_phantom.py"),
+            "test",
+        ],
+        cwd=working / "topological-grid",
+    )
+    environment = dict(os.environ)
+    environment["PYTHONPATH"] = (
+        str(working / "villa/segmentation/models/multi-task-3d-unet")
+        + os.pathsep
+        + environment.get("PYTHONPATH", "")
+    )
+    metric_compat = r'''
+import importlib.util
+import sys
+from pathlib import Path
+import numpy as np
+
+harness_path = Path(sys.argv.pop(1))
+spec = importlib.util.spec_from_file_location("frozen_surface_grid", harness_path)
+if spec is None or spec.loader is None:
+    raise RuntimeError(f"unable to import frozen harness: {harness_path}")
+module = importlib.util.module_from_spec(spec)
+spec.loader.exec_module(module)
+
+original_safe_globals = module._checkpoint_safe_globals
+def torch25_safe_globals():
+    return [entry[0] if isinstance(entry, tuple) else entry
+            for entry in original_safe_globals()]
+module._checkpoint_safe_globals = torch25_safe_globals
+
+original_metrics = module.metrics
+HISTOGRAM_BINS = 4096
+def extended_metrics(probability, truth):
+    result = original_metrics(probability, truth)
+    surface = truth.astype(bool)
+    distance = module.ndimage.distance_transform_edt(~surface)
+    background = (~surface) & (distance >= 2.0) & (distance <= 10.0)
+    positive_hist, _ = np.histogram(
+        probability[surface], bins=HISTOGRAM_BINS, range=(0.0, 1.0))
+    negative_hist, _ = np.histogram(
+        probability[background], bins=HISTOGRAM_BINS, range=(0.0, 1.0))
+    positive_hist = positive_hist.astype(np.int64)
+    negative_hist = negative_hist.astype(np.int64)
+    tpr = np.cumsum(positive_hist[::-1])[::-1] / positive_hist.sum()
+    fpr = np.cumsum(negative_hist[::-1])[::-1] / negative_hist.sum()
+    thresholds = np.arange(HISTOGRAM_BINS, dtype=np.float64) / HISTOGRAM_BINS
+    youden = tpr - fpr
+    candidates = np.flatnonzero(
+        np.isclose(youden, youden.max(), rtol=0.0, atol=1e-15))
+    index = int(candidates[-1])
+    result.update({
+        "histogram_bins": HISTOGRAM_BINS,
+        "positive_score_histogram": positive_hist.tolist(),
+        "negative_score_histogram": negative_hist.tolist(),
+        "best_youden_threshold": float(thresholds[index]),
+        "best_youden_recall": float(tpr[index]),
+        "best_youden_fpr": float(fpr[index]),
+        "best_youden_j": float(youden[index]),
+    })
+    for target in (0.01, 0.05):
+        valid = np.flatnonzero(fpr <= target)
+        target_index = int(valid[0]) if valid.size else HISTOGRAM_BINS - 1
+        label = f"fpr_{int(target * 100)}pct"
+        result[f"threshold_at_{label}"] = float(thresholds[target_index])
+        result[f"recall_at_{label}"] = float(tpr[target_index])
+        result[f"actual_{label}"] = float(fpr[target_index])
+    return result
+
+module.metrics = extended_metrics
+print("TORCH_SAFE_GLOBALS_ADAPTER=tuple_to_callable", flush=True)
+print("CALIBRATION_HISTOGRAM_BINS=4096", flush=True)
+raise SystemExit(module.main())
+'''
+    seed_runs = Path("/kaggle/working/seed-runs")
+    seed_runs.mkdir(parents=True, exist_ok=True)
+    for seed in range(5):
+        phantoms = working / f"phantoms-seed-{seed}"
+        run(
+            [
+                sys.executable,
+                str(working / "topological-grid/scripts/contrast_phantom.py"),
+                "grid",
+                "--out",
+                str(phantoms),
+                "--seed",
+                str(seed),
+            ],
+            cwd=working / "topological-grid",
+        )
+        result = seed_runs / f"seed_{seed}.json"
+        run(
+            [
+                sys.executable,
+                "-c",
+                metric_compat,
+                str(harness),
+                "--checkpoint",
+                str(checkpoint),
+                "--phantoms",
+                str(phantoms),
+                "--output",
+                str(result),
+                "--device",
+                "cuda",
+            ],
+            env=environment,
+        )
+        print(f"SEED_{seed}_RESULT_SHA256={sha256(result)}", flush=True)
+        shutil.rmtree(phantoms)
+
+    aggregate = Path("/kaggle/working/replicate_results.json")
+    aggregate_replicates(seed_runs, aggregate)
+    print(f"REPLICATE_RESULT_SHA256={sha256(aggregate)}", flush=True)
+    print("FIVE_SEED_RESULT_READY", flush=True)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

surface-contrast-grid/five-seed-addendum/validate_replicates.py

@@ -0,0 +1,226 @@
+#!/usr/bin/env python3
+"""Independent validation for the five-seed surface response experiment."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import math
+from pathlib import Path
+
+import numpy as np
+
+
+EXPECTED_CHECKPOINT = "f1990a02ac91889c1f989522ae0e45421a91cb666320448aaf579d42b081636f"
+EXPECTED_PARAMETERS = 102_349_770
+EXPECTED_GRID = {(papyrus, pitch) for papyrus in (35, 50, 65, 90) for pitch in (260.0, 200.0, 150.0, 110.0)}
+SUMMARY_METRICS = (
+    "surface_recall_at_0_5",
+    "near_sheet_auc",
+    "recall_at_fpr_1pct",
+    "recall_at_fpr_5pct",
+    "best_youden_recall",
+    "best_youden_threshold",
+)
+
+
+def sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(8 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+class Checks:
+    def __init__(self) -> None:
+        self.count = 0
+
+    def require(self, condition: bool, label: str) -> None:
+        if not condition:
+            raise AssertionError(label)
+        self.count += 1
+        print(f"PASS {self.count:03d}: {label}")
+
+    def close(self, actual: float, expected: float, label: str, tolerance: float = 1e-12) -> None:
+        self.require(math.isclose(actual, expected, rel_tol=0.0, abs_tol=tolerance), label)
+
+
+def histogram_rates(positive: np.ndarray, negative: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+    return (
+        np.cumsum(positive[::-1])[::-1] / positive.sum(),
+        np.cumsum(negative[::-1])[::-1] / negative.sum(),
+    )
+
+
+def summary(values: list[float]) -> dict:
+    array = np.asarray(values, dtype=np.float64)
+    return {
+        "n": int(array.size),
+        "mean": float(array.mean()),
+        "sample_sd": float(array.std(ddof=1)),
+        "min": float(array.min()),
+        "max": float(array.max()),
+        "values_by_seed": [float(value) for value in array],
+    }
+
+
+def compare_summary(checks: Checks, actual: dict, expected: dict, label: str) -> None:
+    checks.require(actual["n"] == expected["n"], f"{label}: n")
+    for key in ("mean", "sample_sd", "min", "max"):
+        checks.close(float(actual[key]), float(expected[key]), f"{label}: {key}")
+    checks.require(
+        np.allclose(actual["values_by_seed"], expected["values_by_seed"], rtol=0.0, atol=1e-12),
+        f"{label}: values by seed",
+    )
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--runs", type=Path, required=True)
+    parser.add_argument("--aggregate", type=Path, required=True)
+    parser.add_argument("--seed0-baseline", type=Path, required=True)
+    args = parser.parse_args()
+    checks = Checks()
+
+    paths = sorted(args.runs.glob("seed_*.json"), key=lambda path: int(path.stem.split("_")[-1]))
+    checks.require(len(paths) == 5, "five raw seed files exist")
+    checks.require([int(path.stem.split("_")[-1]) for path in paths] == list(range(5)), "seeds are exactly 0 through 4")
+
+    runs = [json.loads(path.read_text(encoding="utf-8")) for path in paths]
+    aggregate = json.loads(args.aggregate.read_text(encoding="utf-8"))
+    baseline = json.loads(args.seed0_baseline.read_text(encoding="utf-8"))
+    bins = int(runs[0]["cells"][0]["metrics"]["histogram_bins"])
+    checks.require(bins == 4096, "histogram resolution is 4096 bins")
+
+    source_hashes = {int(item["seed"]): item["sha256"] for item in aggregate["source_runs"]}
+    for seed, (path, run) in enumerate(zip(paths, runs)):
+        checks.require(run["canonical"] is True, f"seed {seed}: canonical run")
+        checks.require(len(run["cells"]) == 16, f"seed {seed}: sixteen cells")
+        checks.require(run["checkpoint"]["sha256"] == EXPECTED_CHECKPOINT, f"seed {seed}: checkpoint hash")
+        checks.require(run["checkpoint"]["parameter_count"] == EXPECTED_PARAMETERS, f"seed {seed}: strict parameter count")
+        checks.require(run["inference"]["patch_size"] == 192, f"seed {seed}: 192-cube patches")
+        checks.close(float(run["inference"]["step_fraction"]), 0.5, f"seed {seed}: 50% step")
+        checks.require({(int(cell["papyrus"]), float(cell["pitch_um"])) for cell in run["cells"]} == EXPECTED_GRID, f"seed {seed}: complete factorial grid")
+        checks.require(source_hashes[seed] == sha256(path), f"seed {seed}: aggregate pins raw file hash")
+
+        for cell in run["cells"]:
+            metric = cell["metrics"]
+            positive = np.asarray(metric["positive_score_histogram"], dtype=np.int64)
+            negative = np.asarray(metric["negative_score_histogram"], dtype=np.int64)
+            tag = f"seed {seed} pap{cell['papyrus']} pitch{cell['pitch_um']:g}"
+            checks.require(positive.size == bins and negative.size == bins, f"{tag}: histogram lengths")
+            checks.require(int(positive.sum()) == int(metric["surface_voxels"]), f"{tag}: positive histogram count")
+            checks.require(int(negative.sum()) == int(metric["near_background_voxels"]), f"{tag}: negative histogram count")
+            tpr, fpr = histogram_rates(positive, negative)
+            youden = tpr - fpr
+            candidates = np.flatnonzero(np.isclose(youden, youden.max(), rtol=0.0, atol=1e-15))
+            index = int(candidates[-1])
+            checks.close(float(metric["best_youden_threshold"]), index / bins, f"{tag}: best threshold")
+            checks.close(float(metric["best_youden_recall"]), float(tpr[index]), f"{tag}: best recall")
+            checks.close(float(metric["best_youden_fpr"]), float(fpr[index]), f"{tag}: best FPR")
+            for target in (0.01, 0.05):
+                valid = np.flatnonzero(fpr <= target)
+                target_index = int(valid[0]) if valid.size else bins - 1
+                label = f"fpr_{int(target * 100)}pct"
+                checks.close(float(metric[f"threshold_at_{label}"]), target_index / bins, f"{tag}: {label} threshold")
+                checks.close(float(metric[f"recall_at_{label}"]), float(tpr[target_index]), f"{tag}: {label} recall")
+                checks.close(float(metric[f"actual_{label}"]), float(fpr[target_index]), f"{tag}: {label} actual FPR")
+
+    baseline_lookup = {(int(cell["papyrus"]), float(cell["pitch_um"])): cell for cell in baseline["cells"]}
+    seed0_lookup = {(int(cell["papyrus"]), float(cell["pitch_um"])): cell for cell in runs[0]["cells"]}
+    checks.require(set(baseline_lookup) == set(seed0_lookup), "seed 0 matches original grid keys")
+    for key in sorted(seed0_lookup):
+        for metric in ("surface_recall_at_0_5", "near_background_fpr_at_0_5", "near_sheet_auc"):
+            checks.close(float(seed0_lookup[key]["metrics"][metric]), float(baseline_lookup[key]["metrics"][metric]), f"seed 0 reproduces original {key} {metric}")
+
+    grouped = {key: {metric: [] for metric in SUMMARY_METRICS} for key in EXPECTED_GRID}
+    pitch_values = {metric: {pitch: [] for pitch in (260.0, 200.0, 150.0, 110.0)} for metric in SUMMARY_METRICS}
+    papyrus_values = {metric: {papyrus: [] for papyrus in (35, 50, 65, 90)} for metric in SUMMARY_METRICS}
+    for run in runs:
+        by_pitch = {metric: {pitch: [] for pitch in pitch_values[metric]} for metric in SUMMARY_METRICS}
+        by_papyrus = {metric: {papyrus: [] for papyrus in papyrus_values[metric]} for metric in SUMMARY_METRICS}
+        for cell in run["cells"]:
+            key = (int(cell["papyrus"]), float(cell["pitch_um"]))
+            for metric in SUMMARY_METRICS:
+                value = float(cell["metrics"][metric])
+                grouped[key][metric].append(value)
+                by_pitch[metric][key[1]].append(value)
+                by_papyrus[metric][key[0]].append(value)
+        for metric in SUMMARY_METRICS:
+            for pitch, values in by_pitch[metric].items():
+                pitch_values[metric][pitch].append(float(np.mean(values)))
+            for papyrus, values in by_papyrus[metric].items():
+                papyrus_values[metric][papyrus].append(float(np.mean(values)))
+
+    aggregate_cells = {(int(cell["papyrus"]), float(cell["pitch_um"])): cell for cell in aggregate["cell_replicate_summaries"]}
+    for key in sorted(EXPECTED_GRID):
+        for metric in SUMMARY_METRICS:
+            compare_summary(checks, aggregate_cells[key]["metrics"][metric], summary(grouped[key][metric]), f"cell {key} {metric}")
+    for metric in SUMMARY_METRICS:
+        for pitch, values in pitch_values[metric].items():
+            compare_summary(checks, aggregate["pitch_summaries"][metric][str(int(pitch))], summary(values), f"pitch {pitch:g} {metric}")
+        for papyrus, values in papyrus_values[metric].items():
+            compare_summary(checks, aggregate["papyrus_summaries"][metric][str(papyrus)], summary(values), f"papyrus {papyrus} {metric}")
+
+    def pool(run: dict) -> tuple[np.ndarray, np.ndarray]:
+        positive = np.zeros(bins, dtype=np.int64)
+        negative = np.zeros(bins, dtype=np.int64)
+        for cell in run["cells"]:
+            positive += np.asarray(cell["metrics"]["positive_score_histogram"], dtype=np.int64)
+            negative += np.asarray(cell["metrics"]["negative_score_histogram"], dtype=np.int64)
+        return positive, negative
+
+    calibration_positive, calibration_negative = pool(runs[0])
+    heldout_positive = np.zeros(bins, dtype=np.int64)
+    heldout_negative = np.zeros(bins, dtype=np.int64)
+    for run in runs[1:]:
+        positive, negative = pool(run)
+        heldout_positive += positive
+        heldout_negative += negative
+    _, calibration_fpr_curve = histogram_rates(calibration_positive, calibration_negative)
+    operating = aggregate["global_threshold_calibration"]["target_fpr_operating_points"]
+    for target in (0.01, 0.05):
+        index = int(np.flatnonzero(calibration_fpr_curve <= target)[0])
+        record = operating[f"{int(target * 100)}pct"]
+        checks.close(float(record["threshold_selected_on_seed_0"]), index / bins, f"global {target:g} threshold")
+        calibration_tpr, calibration_fpr = histogram_rates(calibration_positive, calibration_negative)
+        heldout_tpr, heldout_fpr = histogram_rates(heldout_positive, heldout_negative)
+        checks.close(float(record["calibration_recall"]), float(calibration_tpr[index]), f"global {target:g} calibration recall")
+        checks.close(float(record["calibration_fpr"]), float(calibration_fpr[index]), f"global {target:g} calibration FPR")
+        checks.close(float(record["heldout_pooled_recall"]), float(heldout_tpr[index]), f"global {target:g} heldout recall")
+        checks.close(float(record["heldout_pooled_fpr"]), float(heldout_fpr[index]), f"global {target:g} heldout FPR")
+
+    auc_pitch = aggregate["pitch_summaries"]["near_sheet_auc"]
+    checks.require(
+        all(a > b for a, b in zip(auc_pitch["150"]["values_by_seed"], auc_pitch["110"]["values_by_seed"])),
+        "pitch 150 AUC exceeds pitch 110 in all five seeds",
+    )
+    fixed_recall = aggregate["pitch_summaries"]["surface_recall_at_0_5"]
+    checks.require(
+        all(
+            fixed_recall["200"]["values_by_seed"][index]
+            < min(fixed_recall[pitch]["values_by_seed"][index] for pitch in ("260", "150", "110"))
+            for index in range(5)
+        ),
+        "pitch 200 recall@0.5 is lowest in all five seeds",
+    )
+    for metric in ("recall_at_fpr_1pct", "recall_at_fpr_5pct"):
+        controlled = aggregate["pitch_summaries"][metric]
+        checks.require(
+            all(
+                controlled["150"]["values_by_seed"][index]
+                > controlled["110"]["values_by_seed"][index]
+                > controlled["200"]["values_by_seed"][index]
+                > controlled["260"]["values_by_seed"][index]
+                for index in range(5)
+            ),
+            f"{metric} ordering 150 > 110 > 200 > 260 holds in all seeds",
+        )
+    print(f"ALL CHECKS PASSED: {checks.count}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())