Update OEDM manuscript for StrataSCAN 0.2.3

.github/workflows/oedm-performance-figures.yml

@@ -0,0 +1,45 @@
+name: Build OEDM performance figures
+
+on:
+  push:
+    branches:
+      - codex/oedm-v023-manuscript
+    paths:
+      - manuscript/oedm2026/scripts/make_manuscript_figures.py
+      - manuscript/oedm2026/tables/performance_tables_2026-08-06/**
+      - benchmarks/datasets.py
+      - src/stratascan/synthetic.py
+      - topology_noise_benchmark/generators.py
+      - .github/workflows/oedm-performance-figures.yml
+
+permissions:
+  contents: write
+
+jobs:
+  render:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: codex/oedm-v023-manuscript
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+          cache: pip
+      - name: Install package and plotting dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install -e .
+          python -m pip install matplotlib pandas scikit-learn
+      - name: Generate figures and tables
+        env:
+          PYTHONPATH: src:.
+        run: python manuscript/oedm2026/scripts/make_manuscript_figures.py
+      - name: Commit rendered artifacts
+        run: |
+          git config user.name github-actions[bot]
+          git config user.email 41898282+github-actions[bot]@users.noreply.github.com
+          git add manuscript/oedm2026/figures/performance_2026-08-06
+          git diff --cached --quiet && exit 0
+          git commit -m "Render reproducible OEDM performance figures"
+          git push

benchmarks/protocol.v0.2.3-biological.json

@@ -0,0 +1,69 @@
+{
+  "schema_version": 1,
+  "protocol_version": "0.2.3-biological",
+  "purpose": "released_stratascan_0_2_3_biological_revalidation_against_frozen_baselines",
+  "parameter_mode": "released_defaults_fixed",
+  "resources": {
+    "threads": 1,
+    "timeout_seconds": 1800,
+    "memory_limit_mb": 8192
+  },
+  "methods": ["StrataSCAN"],
+  "stochastic_methods": [],
+  "method_parameters": {
+    "version": "0.2.3",
+    "k": 32,
+    "shell_ranks": [4, 8, 16, 32],
+    "criterion": "semantic_icl",
+    "initial_component_bound": 8,
+    "hard_component_bound": 24,
+    "background_distribution": "gamma_components",
+    "component_roles": "largest_rate_gap",
+    "core_rank": 4,
+    "connectivity_rank": 4,
+    "border_rule": "strict_local_core_radius"
+  },
+  "comparison": {
+    "baseline_status": "frozen_reused",
+    "baseline_implementation_commit": "5071f3237a76cfd5c7c6c8179df446c0335127b7",
+    "baseline_results": [
+      "results/runs/publication_full_7synthetic_20260722/results.csv",
+      "results/runs/samusik_01/results.csv",
+      "results/runs/samusik_all_samples_02_10/results.csv"
+    ],
+    "quality_tolerance": 0.002,
+    "severe_regression": 0.02,
+    "runtime_ratio_limit": 1.25
+  },
+  "synthetic": {
+    "quality_seeds": [],
+    "scaling_seeds": [],
+    "quality_sizes": [],
+    "scaling_sizes": [],
+    "cases": []
+  },
+  "cytometry": {
+    "seeds": [42],
+    "datasets": [
+      {"id": "levine", "features": "data/processed/levine/features.csv", "feature_metadata": "data/processed/levine/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/levine/event_metadata.csv", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "mosmann", "features": "data/processed/mosmann/features.csv", "feature_metadata": "data/processed/mosmann/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 150.0}, "metadata": "data/processed/mosmann/event_metadata.csv", "label_column": "population_id", "background_labels": ["other"]},
+      {"id": "nilsson", "features": "data/processed/nilsson/features.csv", "feature_metadata": "data/processed/nilsson/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 150.0}, "metadata": "data/processed/nilsson/event_metadata.csv", "label_column": "population_id", "background_labels": ["other"]},
+      {"id": "samusik_01", "features": "data/processed/samusik_01/features.csv.gz", "feature_metadata": "data/processed/samusik_01/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_01/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "samusik_02", "features": "data/processed/samusik_02/features.csv.gz", "feature_metadata": "data/processed/samusik_02/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_02/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "samusik_03", "features": "data/processed/samusik_03/features.csv.gz", "feature_metadata": "data/processed/samusik_03/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_03/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "samusik_04", "features": "data/processed/samusik_04/features.csv.gz", "feature_metadata": "data/processed/samusik_04/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_04/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "samusik_05", "features": "data/processed/samusik_05/features.csv.gz", "feature_metadata": "data/processed/samusik_05/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_05/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "samusik_06", "features": "data/processed/samusik_06/features.csv.gz", "feature_metadata": "data/processed/samusik_06/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_06/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "samusik_07", "features": "data/processed/samusik_07/features.csv.gz", "feature_metadata": "data/processed/samusik_07/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_07/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "samusik_08", "features": "data/processed/samusik_08/features.csv.gz", "feature_metadata": "data/processed/samusik_08/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_08/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "samusik_09", "features": "data/processed/samusik_09/features.csv.gz", "feature_metadata": "data/processed/samusik_09/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_09/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]},
+      {"id": "samusik_10", "features": "data/processed/samusik_10/features.csv.gz", "feature_metadata": "data/processed/samusik_10/feature_metadata.csv", "feature_classes": ["type"], "transform": {"name": "arcsinh", "cofactor": 5.0}, "metadata": "data/processed/samusik_10/event_metadata.csv.gz", "label_column": "population_id", "background_labels": ["unassigned"]}
+    ]
+  },
+  "gaia": {
+    "seeds": [42],
+    "data_root": "data/raw/mctnc/MCTNC_open_data_release/MCTNC_open_data_release/data",
+    "profiles": ["unsupervised_field"],
+    "ruwe_max": 1.6
+  }
+}

intended_regime_benchmark/__init__.py

@@ -0,0 +1,4 @@
+"""Standalone synthetic benchmark for compact targets in diffuse background."""
+
+SCHEMA_VERSION = 1
+

intended_regime_benchmark/metrics.py

@@ -0,0 +1,34 @@
+from __future__ import annotations
+
+from typing import Any
+
+import numpy as np
+
+from redesigned_benchmark_pilot.metrics import evaluate as pilot_evaluate
+
+
+def evaluate(y_true: np.ndarray, y_pred: np.ndarray) -> dict[str, Any]:
+    metrics = pilot_evaluate(y_true, y_pred)
+    true_labels = np.unique(y_true[y_true >= 0])
+    pred_labels = np.unique(y_pred[y_pred >= 0])
+    substantial = np.zeros((true_labels.size, pred_labels.size), dtype=bool)
+    for row, true_label in enumerate(true_labels):
+        truth = y_true == true_label
+        threshold = max(1, int(np.ceil(np.sum(truth) * 0.05)))
+        for column, pred_label in enumerate(pred_labels):
+            substantial[row, column] = np.sum(truth & (y_pred == pred_label)) >= threshold
+    fragments = np.sum(substantial, axis=1) if substantial.size else np.zeros(true_labels.size)
+    metrics["extra_signal_fragments"] = int(np.sum(np.maximum(fragments - 1, 0)))
+    metrics["mean_target_fragments"] = float(np.mean(fragments)) if fragments.size else 0.0
+    metrics["merged_predicted_clusters"] = (
+        int(np.sum(np.sum(substantial, axis=0) >= 2)) if substantial.size else 0
+    )
+    metrics["spurious_background_majority_clusters"] = int(
+        sum(
+            np.sum((y_pred == label) & (y_true < 0))
+            > np.sum((y_pred == label) & (y_true >= 0))
+            for label in pred_labels
+        )
+    )
+    return metrics
+

manuscript/oedm2026/README.md

@@ -31,6 +31,12 @@ renders.
 
 The evidence is deliberately separated by evaluation stage:
 
+- the new 0.2.3 release robustness grid: seven families, five sizes, and five
+  seeds (175 StrataSCAN jobs), with all jobs completed, a 22.6x median paired
+  speedup against the conservative solver, and retained negative cells;
+- completed serial 0.2.3 multidensity scaling measurements through 2,000,000
+  points; the remaining five-million-point campaign cells are explicitly
+  incomplete and are not reported as endpoints;
 - the original 315-job synthetic reference matrix: seven families, five
   pre-freeze seeds, nine methods, `n = 5,000`, with quality, segmentation,
   runtime, and process-memory measurements;
@@ -44,11 +50,13 @@ The evidence is deliberately separated by evaluation stage:
 - the prespecified 24-field Gaia DR3 pilot;
 - conventional graph/density baselines where protocol coverage permits.
 
-The scalability evidence is deliberately separated by algorithm generation.
-The evaluated fixed-range snapshot has direct monitored runs through the
+The scalability evidence is deliberately separated by algorithm generation and
+protocol. The 0.2.3 release has completed direct serial multidensity runs
+through 2,000,000 points; its other scaling cells remain in progress. The
+historical fixed-range snapshot has direct monitored runs through the
 396,460-cell Mosmann dataset. The one-million-point dev9 smoke test and the
 five-million-point 0.1.2 matrix are retained as lineage evidence, not attributed
-to the submitted solver. `tables/scalability_evidence_boundaries.csv` records
+to the submitted solver. The historical scalability evidence table records
 this distinction explicitly.
 
 The incomplete 359-field Gaia sweep is not included in any result, table, test,
@@ -82,5 +90,5 @@ artifact. The manuscript distinguishes strong biological recovery evidence from
 segmentation/background limitations and the mixed Gaia pilot rather than making
 a universal-superiority claim.
 
-The checked PDF is nine US-Letter pages including references. Its SHA-256
-digest is `fa85c74693abfe82b92f1bacf73ff81d983e0449c3da5a898d382a3bf4e7c19f`.
+The checked 0.2.3 PDF is ten US-Letter pages including references. Its SHA-256
+digest is `95fd4970b9925677081403ceecd0e26cea4fbcb5afed9077bb7c68b370322a15`.

manuscript/oedm2026/REVIEW_AND_REVISION.md

@@ -158,3 +158,39 @@ the locked revision still has only two unseen seeds per family.
 - Offline source and wheel builds completed for version 0.2.2.
 - The exact final PDF is nine US-Letter pages with result floats before the
   discussion and references.
+
+## Round 5 - 0.2.3 release update and review
+
+The workshop manuscript was updated for the compiled constrained-Gamma solver
+and guarded local confirmation introduced in 0.2.3. The review found that the
+earlier draft would have materially overstated provenance if its fixed 1--8
+component description and historical real-data outputs were silently treated as
+the released 0.2.3 method.
+
+Corrections in this round:
+
+- The method and protocol sections now describe the released adaptive 8--24
+  component search, one split-warm ordinary fit, and the stricter local
+  confirmation conditions.
+- A new results subsection reports the completed 175-cell seven-family
+  robustness grid, including all recovered conservative timeouts, paired speed
+  and pairwise-F1 summaries, and the four retained material regressions.
+- A new serial scaling table reports only the completed 0.5M, 1M, and 2M
+  multidensity jobs. The unfinished 5M and remaining-family cells, and all
+  unfinished campaign phases, are explicitly excluded from conclusions.
+- The biological, Gaia, and older locked-synthetic analyses are now labelled
+  historical rather than evidence for 0.2.3, because they were not rerun under
+  the release profile.
+
+Remaining release-specific limitations are visible in the paper: quality can
+still change because local optima differ from the conservative solver; four
+non-imbalanced cells lose more than 0.05 pairwise F1; and the completed scaling
+evidence currently covers a single family and seed. The review recommendation
+therefore remains scoped: the new data support a speed-and-robustness release,
+not universal quality or real-data superiority.
+
+### Round 5 verification
+
+- Tectonic compilation completed with no unresolved references.
+- The final ten-page PDF was rasterized at 150 dpi and every page was visually
+  inspected; no clipping, overlap, or broken table/figure layout was found.

manuscript/oedm2026/artifact_manifest.json

@@ -4,18 +4,18 @@
   "evaluated_method": {
     "display_name": "MDL-StrataSCAN",
     "historical_result_id": "StrataSCAN-Optimization",
-    "snapshot": "0.2.0-dev10",
-    "gamma_component_range": [1, 8],
-    "adaptive_component_search": false,
+    "snapshot": "0.2.3",
+    "gamma_component_range": [1, 24],
+    "adaptive_component_search": true,
     "shell_ranks": [4, 8, 16, 32],
     "core_rank": 4,
     "connectivity_rank": 4
   },
   "release_context": {
-    "public_version": "0.2.2",
+    "public_version": "0.2.3",
     "public_component_range": [1, 24],
     "public_adaptive_component_search": true,
-    "relationship": "The public release preserves the 0.2.1 algorithm defaults; manuscript results use the frozen fixed-range snapshot above."
+    "relationship": "The manuscript reports the 0.2.3 release robustness and completed serial scaling evidence separately from historical fixed-range synthetic, biological, and Gaia results."
   },
   "declared_endpoints": {
     "synthetic": "macro_target_f1",
@@ -30,7 +30,7 @@
     "biological_study_leave_one_out"
   ],
   "sha256": {
-    "manuscript/oedm2026/MDL-StrataSCAN_OEDM2026_draft.pdf": "fa85c74693abfe82b92f1bacf73ff81d983e0449c3da5a898d382a3bf4e7c19f",
+    "manuscript/oedm2026/MDL-StrataSCAN_OEDM2026_draft.pdf": "95fd4970b9925677081403ceecd0e26cea4fbcb5afed9077bb7c68b370322a15",
     "results/runs/publication_full_7synthetic_20260722/results.csv": "8c4ebadc9c1bde8f2e0b0ce69dd68d0daa62c5d5a3432b76fefd3fe317e3fa68",
     "results/runs/samusik_01/results.csv": "463931d1250ce4eb0e4505f231473d285288882cac4aa2bef6206365e9ac545e",
     "results/runs/samusik_all_samples_02_10/results.csv": "28839ba4921b63318a45fcce3033cbe478475b0c2851a0f5e5d00546bc4e1bf7",
@@ -44,12 +44,15 @@
     "benchmarks/protocol.optimization-v0.2.0-dev10.json": "8d81d8c76c6936dd84a1c54153251304da53d60d9914430cd9ea77f1096614cc",
     "benchmarks/protocol.optimization-v0.2.0-dev10-real-pilot.json": "8d975776ada0bda94a5a22ad4c26a45423fb71be957af067763802853362e622",
     "benchmarks/protocol.v0.2.1-release.json": "bee41c84b8d609ca3d8cbaf6a7dd0bfb68d785948203fbadb3258a0f2148c0ee",
-    "benchmarks/protocol.v0.2.2-release.json": "d4a16647fafea23c11e384fea241d861558c5981687ac6df3e2998c3c7eae279"
+    "benchmarks/protocol.v0.2.2-release.json": "d4a16647fafea23c11e384fea241d861558c5981687ac6df3e2998c3c7eae279",
+    "benchmarks/protocol.v0.2.3-release.json": "0a26b3a0a8bd0248864dc560d27aaf9dec4582d102ff2489f40ba67132585c4c",
+    "benchmarks/protocol.v0.2.3-seven-family-robustness.json": "f10d1488dcfd5930e7df810d7ecfcc3a3ef293c0a2c2daff7b1a6d0f434e0f8a",
+    "benchmarks/protocol.v0.2.3-seven-family-stratascan-scalability-5m.json": "1f7dbd5df7b4a2134d5fbb7f820778b08335821e5831b3be77900bd7cd8c7a77"
   },
   "notes": [
     "Frozen result files retain their historical method identifiers.",
     "Current replay protocols use explicit versioned identifiers so the mutable public StrataSCAN alias cannot select the wrong estimator.",
     "The target-background and target-background-structure harmonic F1, Pareto counts, and study sensitivity summaries were computed after the frozen predictions and are not predeclared endpoints.",
-    "Scalability records are attributed to their exact algorithm generation; only fixed-range MDL rows characterize the submitted snapshot."
+    "Scalability records are attributed to their exact algorithm generation and protocol; the completed 0.2.3 serial multidensity records through two million points are separated from historical rows and incomplete campaign cells."
   ]
 }