init

.gitattributes

@@ -0,0 +1,60 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.avro filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.lz4 filter=lfs diff=lfs merge=lfs -text
+*.mds filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+# Audio files - uncompressed
+*.pcm filter=lfs diff=lfs merge=lfs -text
+*.sam filter=lfs diff=lfs merge=lfs -text
+*.raw filter=lfs diff=lfs merge=lfs -text
+# Audio files - compressed
+*.aac filter=lfs diff=lfs merge=lfs -text
+*.flac filter=lfs diff=lfs merge=lfs -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text
+*.ogg filter=lfs diff=lfs merge=lfs -text
+*.wav filter=lfs diff=lfs merge=lfs -text
+# Image files - uncompressed
+*.bmp filter=lfs diff=lfs merge=lfs -text
+*.gif filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.tiff filter=lfs diff=lfs merge=lfs -text
+# Image files - compressed
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+# Video files - compressed
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.webm filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,45 @@
+---
+pretty_name: SubcellPert
+language: [en]
+tags: [biology, spatial-transcriptomics, perturbation, subcellular]
+configs:
+- config_name: catalog
+  default: true
+  data_files:
+  - split: inventory
+    path: catalog/datasets.parquet
+---
+
+<h1 align="center">SubcellPert: Benchmarking Subcellular Perturbation Prediction</h1>
+
+<p align="center">SubcellPert benchmarks perturbation-response prediction across 14 imaging datasets and eight baselines, extending virtual cell evaluation from gene counts to intracellular RNA organization.</p>
+
+<p align="center">
+  <a href="https://personal-page.wjh.app/SubcellPert/">Project page</a> ·
+  <a href="https://www.biorxiv.org/content/10.64898/2026.10.11.758305v1">Paper</a> ·
+  <a href="https://github.com/king159/SubcellPert">Code</a>
+</p>
+
+## Dataset overview
+
+This snapshot contains **14 datasets, 13,880,723 cells, and
+3,653,986,724 retained RNA events**, with frozen benchmark splits and
+intervention descriptors. Each physical dataset is counted once. Seeds,
+baselines, and oracle/joint settings reuse the observations.
+
+| Dataset | Cells | RNA events | Genes |
+| --- | ---: | ---: | ---: |
+| DSS-Colitis | 1,439,542 | 119,553,365 | 940 |
+| ATM-RT | 790,166 | 147,722,071 | 298 |
+| Kidney-IRI | 1,374,915 | 168,519,194 | 300 |
+| Cardiac-MI | 225,023 | 61,134,068 | 475 |
+| Ddx54-PD1 | 1,155,233 | 494,265,984 | 5,006 |
+| Spinal-KO | 122,645 | 36,073,815 | 347 |
+| Adipose | 13,972 | 1,907,755 | 477 |
+| Breast-FUS | 880,565 | 102,102,171 | 1,000 |
+| GBM-oHSV | 2,401,940 | 520,974,197 | 480 |
+| Lung-Drug | 1,561,700 | 681,015,170 | 5,006 |
+| Gut-Microbiota | 2,057,563 | 189,197,264 | 1,815 |
+| Sickness | 1,765,443 | 1,101,298,121 | 941 |
+| PF-Astro | 11,454 | 10,670,839 | 485 |
+| UC-VDZ | 80,562 | 19,552,710 | 1,000 |

benchmark/intervention_descriptors/intervention_ood_compositional/gse246584_atm_radiation_xenium.json

@@ -0,0 +1,28 @@
+{
+  "artifact": "benchmark/intervention_descriptors/intervention_ood_compositional/gse246584_atm_radiation_xenium.npz",
+  "component_count": 2,
+  "component_normalization": "l2",
+  "component_texts": {
+    "ATM_loss": "ATM_loss",
+    "radiation": "radiation"
+  },
+  "condition_composition": "additive_sum",
+  "dataset": "gse246584_atm_radiation_xenium",
+  "dimension": 2,
+  "generation": {
+    "component_basis": [
+      "ATM_loss",
+      "radiation"
+    ],
+    "provider": "component_multihot",
+    "target_response_files_opened": [],
+    "target_response_rows_used": 0
+  },
+  "provider": "component_multihot",
+  "schema": "intervention_component_descriptors",
+  "schema_version": 1,
+  "split_id": "gse246584_atm_radiation_compositional",
+  "status": "complete",
+  "target_response_data_used": false,
+  "task_id": "intervention_ood_compositional"
+}

benchmark/intervention_descriptors/intervention_ood_compositional/gse246584_atm_radiation_xenium.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f549409d7dc661cff612a76f84e4211ccd599a9f703120201f4e56b20a12629d
+size 449

benchmark/intervention_descriptors/intervention_ood_compositional/gse285341_ddx54_pd1_xenium.json

@@ -0,0 +1,28 @@
+{
+  "artifact": "benchmark/intervention_descriptors/intervention_ood_compositional/gse285341_ddx54_pd1_xenium.npz",
+  "component_count": 2,
+  "component_normalization": "l2",
+  "component_texts": {
+    "Ddx54_KD": "Ddx54_KD",
+    "anti_PD1": "anti_PD1"
+  },
+  "condition_composition": "additive_sum",
+  "dataset": "gse285341_ddx54_pd1_xenium",
+  "dimension": 2,
+  "generation": {
+    "component_basis": [
+      "Ddx54_KD",
+      "anti_PD1"
+    ],
+    "provider": "component_multihot",
+    "target_response_files_opened": [],
+    "target_response_rows_used": 0
+  },
+  "provider": "component_multihot",
+  "schema": "intervention_component_descriptors",
+  "schema_version": 1,
+  "split_id": "gse285341_ddx54_pd1_compositional",
+  "status": "complete",
+  "target_response_data_used": false,
+  "task_id": "intervention_ood_compositional"
+}

benchmark/intervention_descriptors/intervention_ood_compositional/gse285341_ddx54_pd1_xenium.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b7a27c74ac51dc4632f6f0b0deb8ca4ebe8ad1d4829f66f9b46e5d6a340e079
+size 448

benchmark/intervention_descriptors/intervention_ood_main/perturbfish_astrocytes.json

@@ -0,0 +1,788 @@
+{
+  "artifact": "benchmark/intervention_descriptors/intervention_ood_main/perturbfish_astrocytes.npz",
+  "component_count": 127,
+  "component_normalization": "l2",
+  "component_texts": {
+    "ADNP": "ADNP",
+    "AHDC1": "AHDC1",
+    "ANK2": "ANK2",
+    "ANKRD11": "ANKRD11",
+    "AP2S1": "AP2S1",
+    "ARID1B": "ARID1B",
+    "ASH1L": "ASH1L",
+    "ASXL3": "ASXL3",
+    "BCKDK": "BCKDK",
+    "BCL11A": "BCL11A",
+    "CACNA1E": "CACNA1E",
+    "CACNA2D3": "CACNA2D3",
+    "CASK": "CASK",
+    "CELF4": "CELF4",
+    "CHAMP1": "CHAMP1",
+    "CHD2": "CHD2",
+    "CHD8": "CHD8",
+    "CNOT3": "CNOT3",
+    "CORO1A": "CORO1A",
+    "CREBBP": "CREBBP",
+    "CSNK2A1": "CSNK2A1",
+    "CTCF": "CTCF",
+    "CTNNB1": "CTNNB1",
+    "DDX3X": "DDX3X",
+    "DEAF1": "DEAF1",
+    "DIP2A": "DIP2A",
+    "DNMT3A": "DNMT3A",
+    "DPYSL2": "DPYSL2",
+    "DSCAM": "DSCAM",
+    "DYNC1H1": "DYNC1H1",
+    "DYRK1A": "DYRK1A",
+    "EHMT1": "EHMT1",
+    "EIF3F": "EIF3F",
+    "EIF3G": "EIF3G",
+    "ELAVL3": "ELAVL3",
+    "EP300": "EP300",
+    "FOXG1": "FOXG1",
+    "FOXP1": "FOXP1",
+    "FOXP2": "FOXP2",
+    "GABRB2": "GABRB2",
+    "GABRB3": "GABRB3",
+    "GFAP": "GFAP",
+    "GIGYF1": "GIGYF1",
+    "GNAI1": "GNAI1",
+    "GRIA2": "GRIA2",
+    "GRIN2B": "GRIN2B",
+    "HDLBP": "HDLBP",
+    "HECTD4": "HECTD4",
+    "HIVEP2": "HIVEP2",
+    "HNRNPH2": "HNRNPH2",
+    "HNRNPU": "HNRNPU",
+    "IQSEC2": "IQSEC2",
+    "IRF2BPL": "IRF2BPL",
+    "ITPR3": "ITPR3",
+    "KCNMA1": "KCNMA1",
+    "KCNQ3": "KCNQ3",
+    "KDM5B": "KDM5B",
+    "KDM6B": "KDM6B",
+    "KIAA0232": "KIAA0232",
+    "KMT2A": "KMT2A",
+    "KMT2C": "KMT2C",
+    "KMT2E": "KMT2E",
+    "LDB1": "LDB1",
+    "LRRC4C": "LRRC4C",
+    "MAP1A": "MAP1A",
+    "MBD5": "MBD5",
+    "MED13L": "MED13L",
+    "MKX": "MKX",
+    "MYT1L": "MYT1L",
+    "NACC1": "NACC1",
+    "NCOA1": "NCOA1",
+    "NR3C2": "NR3C2",
+    "NRXN1": "NRXN1",
+    "NSD1": "NSD1",
+    "NUP155": "NUP155",
+    "PACS1": "PACS1",
+    "PAX5": "PAX5",
+    "PHF12": "PHF12",
+    "PHF2": "PHF2",
+    "PHF21A": "PHF21A",
+    "POGZ": "POGZ",
+    "PPP1R9B": "PPP1R9B",
+    "PPP2R5D": "PPP2R5D",
+    "PPP5C": "PPP5C",
+    "PRR12": "PRR12",
+    "PTEN": "PTEN",
+    "PTK7": "PTK7",
+    "RAI1": "RAI1",
+    "RFX3": "RFX3",
+    "RORB": "RORB",
+    "SATB1": "SATB1",
+    "SCN1A": "SCN1A",
+    "SCN2A": "SCN2A",
+    "SCN8A": "SCN8A",
+    "SETBP1": "SETBP1",
+    "SETD5": "SETD5",
+    "SHANK2": "SHANK2",
+    "SHANK3": "SHANK3",
+    "SIN3A": "SIN3A",
+    "SKI": "SKI",
+    "SLC6A1": "SLC6A1",
+    "SMARCA2": "SMARCA2",
+    "SMARCC2": "SMARCC2",
+    "SPAST": "SPAST",
+    "SRCAP": "SRCAP",
+    "SRPR": "SRPR",
+    "STXBP1": "STXBP1",
+    "SUV420H1": "SUV420H1",
+    "SYNGAP1": "SYNGAP1",
+    "TAOK1": "TAOK1",
+    "TBL1XR1": "TBL1XR1",
+    "TBR1": "TBR1",
+    "TCF20": "TCF20",
+    "TCF4": "TCF4",
+    "TCF7L2": "TCF7L2",
+    "TEK": "TEK",
+    "TLK2": "TLK2",
+    "TM9SF4": "TM9SF4",
+    "TRAF7": "TRAF7",
+    "TRIM23": "TRIM23",
+    "TRIO": "TRIO",
+    "TRIP12": "TRIP12",
+    "UBR1": "UBR1",
+    "USP9X": "USP9X",
+    "VEZF1": "VEZF1",
+    "WAC": "WAC",
+    "ZMYND8": "ZMYND8"
+  },
+  "condition_composition": "additive_sum",
+  "dataset": "perturbfish_astrocytes",
+  "dimension": 612,
+  "generation": {
+    "cache_manifest": "datasets/perturbfish_astrocytes/manifest.json",
+    "coexpression_basis": [
+      "ABCA1",
+      "ABHD3",
+      "ACAA2",
+      "ACKR3",
+      "ACO1",
+      "ACTN2",
+      "ADD1",
+      "ADNP",
+      "ADORA1",
+      "AEBP1",
+      "AFF1",
+      "AGT",
+      "AK3",
+      "AK4",
+      "AKR1C3",
+      "ALDH1A1",
+      "ALDH1L1",
+      "AMOT",
+      "ANAPC5",
+      "ANK2",
+      "ANK3",
+      "ANKRD11",
+      "ANKRD36",
+      "ANPEP",
+      "AP2S1",
+      "AQP4",
+      "ARHGAP42",
+      "ARHGAP44",
+      "ARHGAP5",
+      "ARHGEF26",
+      "ARID1B",
+      "ARL6",
+      "ARMCX6",
+      "ASAP3",
+      "ASCL1",
+      "ASH1L",
+      "ASXL3",
+      "ATP1A2",
+      "AXL",
+      "BCL11A",
+      "BCL6",
+      "BMP2K",
+      "BMPR1A",
+      "BMPR1B",
+      "C3orf70",
+      "CA12",
+      "CACHD1",
+      "CACNA1H",
+      "CACNA2D3",
+      "CADM2",
+      "CAMK4",
+      "CAMKK2",
+      "CARNS1",
+      "CASP7",
+      "CASR",
+      "CCDC18",
+      "CCDC8",
+      "CCNG1",
+      "CD300A",
+      "CD302",
+      "CD38",
+      "CD4",
+      "CD59",
+      "CDC37",
+      "CDC42EP4",
+      "CDH20",
+      "CDK5",
+      "CELF4",
+      "CEP112",
+      "CERCAM",
+      "CHCHD2",
+      "CHCHD5",
+      "CHD2",
+      "CHD8",
+      "CHPT1",
+      "CHRDL1",
+      "CIRBP",
+      "CLDN10",
+      "CNN1",
+      "CNN3",
+      "CNTFR",
+      "COL19A1",
+      "COL9A2",
+      "COLGALT1",
+      "CORO1A",
+      "COX20",
+      "CPEB3",
+      "CPNE3",
+      "CREBBP",
+      "CRIP2",
+      "CTNNB1",
+      "CTSC",
+      "CYP4F11",
+      "DDAH2",
+      "DEAF1",
+      "DES",
+      "DHRS3",
+      "DIP2A",
+      "DLC1",
+      "DNMT3A",
+      "DOCK7",
+      "DPY19L3",
+      "DPYSL2",
+      "DSCAM",
+      "DYNC1H1",
+      "DYRK1A",
+      "EGLN3",
+      "EHBP1L1",
+      "EIF2D",
+      "ELAVL3",
+      "EMX2",
+      "ENAH",
+      "ENO3",
+      "EPAS1",
+      "EPHB1",
+      "ERLIN2",
+      "EYA2",
+      "EZR",
+      "FABP7",
+      "FAH",
+      "FAHD2CP",
+      "FAM126A",
+      "FAM171B",
+      "FARP1",
+      "FBXL17",
+      "FBXO30",
+      "FBXO8",
+      "FDPS",
+      "FEM1C",
+      "FERMT2",
+      "FGD4",
+      "FGD6",
+      "FGF2",
+      "FGFRL1",
+      "FIGN",
+      "FKBP10",
+      "FMO1",
+      "FOXN3",
+      "FOXO1",
+      "FOXO3",
+      "FOXP1",
+      "FOXP2",
+      "FRZB",
+      "GABRB1",
+      "GABRB3",
+      "GABRG1",
+      "GALM",
+      "GALNT6",
+      "GARNL3",
+      "GDNF",
+      "GIGYF1",
+      "GJA1",
+      "GJA3",
+      "GJB6",
+      "GLI3",
+      "GLIS3",
+      "GNA13",
+      "GNAI1",
+      "GNG12",
+      "GNG5",
+      "GPATCH8",
+      "GPHN",
+      "GPIHBP1",
+      "GPM6A",
+      "GPNMB",
+      "GPR137B",
+      "GPR61",
+      "GPRC6A",
+      "GRAMD2B",
+      "GRIA1",
+      "GRIA2",
+      "GRID1",
+      "GRIN2A",
+      "GRIN2B",
+      "GRM3",
+      "GSK3B",
+      "HCLS1",
+      "HDAC4",
+      "HDLBP",
+      "HECTD4",
+      "HES1",
+      "HGF",
+      "HIF1A",
+      "HIF3A",
+      "HLA-B",
+      "HMBOX1",
+      "HMG20B",
+      "HMGN3",
+      "HS3ST5",
+      "IKZF2",
+      "IL17RB",
+      "IL33",
+      "IL6ST",
+      "INTU",
+      "IPO13",
+      "IRF2BPL",
+      "ITGAV",
+      "ITPR3",
+      "KAZN",
+      "KCNH8",
+      "KCNK13",
+      "KCNN3",
+      "KCNQ3",
+      "KDM5B",
+      "KDM6B",
+      "KIAA0232",
+      "KLF15",
+      "KLHL21",
+      "KLHL41",
+      "KMT2C",
+      "KMT2E",
+      "LAMA1",
+      "LAPTM4A",
+      "LCP1",
+      "LDB1",
+      "LGALS3",
+      "LHFPL1",
+      "LHX2",
+      "LILRB1",
+      "LINC00032",
+      "LMO3",
+      "LPAR1",
+      "LPP",
+      "LRBA",
+      "LRRC8C",
+      "LTBP1",
+      "LYPLAL1",
+      "MAOA",
+      "MAP1A",
+      "MAPK8",
+      "MAPKAPK3",
+      "MBNL1",
+      "MCC",
+      "MDFIC",
+      "MED13L",
+      "MEF2C-AS1",
+      "MEIS2",
+      "MERTK",
+      "MGST1",
+      "MID1",
+      "MLLT3",
+      "MOB1B",
+      "MON2",
+      "MPPED2",
+      "MSMO1",
+      "MT2A",
+      "MTMR2",
+      "MYH13",
+      "MYH6",
+      "MYH8",
+      "MYL4",
+      "MYL6B",
+      "MYL9",
+      "MYLPF",
+      "MYO1F",
+      "MYT1L",
+      "NABP1",
+      "NACC1",
+      "NACC2",
+      "NADK2",
+      "NCAM2",
+      "NCOA1",
+      "NCOA2",
+      "NEK6",
+      "NFAM1",
+      "NFE2L2",
+      "NFIB",
+      "NHLH2",
+      "NHSL1",
+      "NOTCH2",
+      "NOVA1",
+      "NPAS3",
+      "NQO1",
+      "NR2E1",
+      "NRF1",
+      "NRXN3",
+      "NSD1",
+      "NT5DC1",
+      "NTSR2",
+      "NUP155",
+      "OGT",
+      "OPHN1",
+      "OXR1",
+      "P2RX2",
+      "P2RX4",
+      "P2RX5",
+      "P2RX7",
+      "P2RY1",
+      "P2RY12",
+      "P2RY2",
+      "P2RY4",
+      "P4HA1",
+      "PABPC4",
+      "PACS1",
+      "PACSIN3",
+      "PALLD",
+      "PANX1",
+      "PARD3",
+      "PARVG",
+      "PATZ1",
+      "PAX6",
+      "PBXIP1",
+      "PCSK5",
+      "PDCD6IP",
+      "PDE10A",
+      "PDE4DIP",
+      "PDLIM3",
+      "PDLIM5",
+      "PHF12",
+      "PHF2",
+      "PHF21A",
+      "PLCB3",
+      "PLD1",
+      "PLEKHG1",
+      "PLLP",
+      "PLOD2",
+      "PLSCR1",
+      "PLSCR4",
+      "PLTP",
+      "POGZ",
+      "POU2F1",
+      "POU3F2",
+      "POU3F3",
+      "PPARA",
+      "PPFIA1",
+      "PPP1R14B",
+      "PPP1R9B",
+      "PPP2R5D",
+      "PPP5C",
+      "PRDM16",
+      "PRKCA",
+      "PRKCQ-AS1",
+      "PRR12",
+      "PRRX1",
+      "PSME4",
+      "PTAR1",
+      "PTEN",
+      "PTK2",
+      "PTK7",
+      "PTPRZ1",
+      "PTTG1IP",
+      "RAB31",
+      "RAI1",
+      "RAMP1",
+      "RASGRP1",
+      "RASSF8",
+      "RBM47",
+      "RELA",
+      "REPIN1",
+      "REPS1",
+      "RFX3",
+      "RFX4",
+      "RGS6",
+      "RHOBTB3",
+      "RHOC",
+      "RHPN2",
+      "RMST",
+      "RNASE1",
+      "RORB",
+      "RREB1",
+      "RXRA",
+      "RXRG",
+      "S100A16",
+      "S100A4",
+      "S100A6",
+      "S1PR1",
+      "SALL2",
+      "SAT2",
+      "SATB1",
+      "SCN2A",
+      "SDC2",
+      "SEC14L5",
+      "SERP1",
+      "SESTD1",
+      "SETD5",
+      "SGCD",
+      "SH3GLB1",
+      "SHANK2",
+      "SIN3A",
+      "SIRPB3P",
+      "SKI",
+      "SLC12A2",
+      "SLC16A9",
+      "SLC1A2",
+      "SLC1A3",
+      "SLC1A4",
+      "SLC22A15",
+      "SLC25A1",
+      "SLC25A20",
+      "SLC25A43",
+      "SLC2A12",
+      "SLC2A4RG",
+      "SLC35B3",
+      "SLC3A2",
+      "SLC44A3",
+      "SLC4A4",
+      "SLC5A6",
+      "SLC6A1",
+      "SLC6A11",
+      "SLC7A11",
+      "SLC7A11-AS1",
+      "SLC7A2",
+      "SMAD5",
+      "SMARCC2",
+      "SMO",
+      "SMURF2",
+      "SOX2",
+      "SOX5",
+      "SOX6",
+      "SOX9",
+      "SP1",
+      "SP110",
+      "SPAST",
+      "SPIN1",
+      "SREBF1",
+      "SREBF2",
+      "SRI",
+      "SRPRA",
+      "STAT3",
+      "STK17B",
+      "STK33",
+      "STOM",
+      "STON2",
+      "STXBP1",
+      "SYNGAP1",
+      "TAGLN",
+      "TAOK1",
+      "TBL1XR1",
+      "TBRG1",
+      "TCF20",
+      "TCF25",
+      "TCF4",
+      "TCF7L1",
+      "TCF7L2",
+      "TGFB2",
+      "TGFBR3",
+      "TGM2",
+      "TLK2",
+      "TLR4",
+      "TM9SF4",
+      "TMED5",
+      "TMEM161B-AS1",
+      "TMEM167B",
+      "TMEM169",
+      "TMEM47",
+      "TMSB4Y",
+      "TNNC1",
+      "TNNI1",
+      "TNRC6A",
+      "TNRC6B",
+      "TOPORS",
+      "TOX",
+      "TP53",
+      "TP53BP2",
+      "TPD52L1",
+      "TRAF7",
+      "TRIM23",
+      "TRIP12",
+      "TRPA1",
+      "TRPS1",
+      "TRPV4",
+      "TSPAN6",
+      "TSPAN9",
+      "TTLL11",
+      "TUBA4A",
+      "TUBB2B",
+      "TULP3",
+      "USP53",
+      "UVRAG",
+      "VEZF1",
+      "VIM",
+      "VPS13B",
+      "WAC",
+      "WBP2",
+      "WLS",
+      "YAP1",
+      "ZBTB1",
+      "ZFHX4",
+      "ZKSCAN8",
+      "ZMYND8",
+      "ZNF395",
+      "ZNF423",
+      "ZNF516",
+      "ZNF606",
+      "ZNF98"
+    ],
+    "coexpression_neighbors": 20,
+    "coexpression_nonzero_edges": 999,
+    "coexpression_threshold": 0.1,
+    "go_annotations": "external/scFoundation/GEARS/data/gene2go.pkl",
+    "go_basis": [
+      "ADNP",
+      "AHDC1",
+      "ANK2",
+      "ANKRD11",
+      "AP2S1",
+      "ARID1B",
+      "ASH1L",
+      "ASXL3",
+      "BCKDK",
+      "BCL11A",
+      "CACNA1E",
+      "CACNA2D3",
+      "CASK",
+      "CELF4",
+      "CHAMP1",
+      "CHD2",
+      "CHD8",
+      "CNOT3",
+      "CORO1A",
+      "CREBBP",
+      "CSNK2A1",
+      "CTCF",
+      "CTNNB1",
+      "DDX3X",
+      "DEAF1",
+      "DIP2A",
+      "DNMT3A",
+      "DPYSL2",
+      "DSCAM",
+      "DYNC1H1",
+      "DYRK1A",
+      "EHMT1",
+      "EIF3F",
+      "EIF3G",
+      "ELAVL3",
+      "EP300",
+      "FOXG1",
+      "FOXP1",
+      "FOXP2",
+      "GABRB2",
+      "GABRB3",
+      "GFAP",
+      "GIGYF1",
+      "GNAI1",
+      "GRIA2",
+      "GRIN2B",
+      "HDLBP",
+      "HECTD4",
+      "HIVEP2",
+      "HNRNPH2",
+      "HNRNPU",
+      "IQSEC2",
+      "IRF2BPL",
+      "ITPR3",
+      "KCNMA1",
+      "KCNQ3",
+      "KDM5B",
+      "KDM6B",
+      "KIAA0232",
+      "KMT2A",
+      "KMT2C",
+      "KMT2E",
+      "LDB1",
+      "LRRC4C",
+      "MAP1A",
+      "MBD5",
+      "MED13L",
+      "MKX",
+      "MYT1L",
+      "NACC1",
+      "NCOA1",
+      "NR3C2",
+      "NRXN1",
+      "NSD1",
+      "NUP155",
+      "PACS1",
+      "PAX5",
+      "PHF12",
+      "PHF2",
+      "PHF21A",
+      "POGZ",
+      "PPP1R9B",
+      "PPP2R5D",
+      "PPP5C",
+      "PRR12",
+      "PTEN",
+      "PTK7",
+      "RAI1",
+      "RFX3",
+      "RORB",
+      "SATB1",
+      "SCN1A",
+      "SCN2A",
+      "SCN8A",
+      "SETBP1",
+      "SETD5",
+      "SHANK2",
+      "SHANK3",
+      "SIN3A",
+      "SKI",
+      "SLC6A1",
+      "SMARCA2",
+      "SMARCC2",
+      "SPAST",
+      "SRCAP",
+      "SRPR",
+      "STXBP1",
+      "SUV420H1",
+      "SYNGAP1",
+      "TAOK1",
+      "TBL1XR1",
+      "TBR1",
+      "TCF20",
+      "TCF4",
+      "TCF7L2",
+      "TEK",
+      "TLK2",
+      "TM9SF4",
+      "TRAF7",
+      "TRIM23",
+      "TRIO",
+      "TRIP12",
+      "UBR1",
+      "USP9X",
+      "VEZF1",
+      "WAC",
+      "ZMYND8"
+    ],
+    "go_neighbors": 20,
+    "go_nonzero_edges": 2539,
+    "population_manifest": "datasets/perturbfish_astrocytes/population/manifest.json",
+    "provider": "gears_go_source_coexpression",
+    "source_cell_files": [
+      "inference_cells/shards/0000-sample1.parquet",
+      "inference_cells/shards/0001-sample2.parquet",
+      "inference_cells/shards/0002-sample5.parquet"
+    ],
+    "source_cells": 2036,
+    "source_conditions": [
+      "ctrl"
+    ],
+    "source_molecules": 2131199,
+    "source_normalization": "library_size_10000_log1p",
+    "source_population_files": [
+      "shards/0000-sample1/populations.parquet",
+      "shards/0001-sample2/populations.parquet",
+      "shards/0002-sample5/populations.parquet"
+    ],
+    "target_response_rows_used": 0
+  },
+  "provider": "gears_go_source_coexpression",
+  "schema": "intervention_component_descriptors",
+  "schema_version": 1,
+  "split_id": "perturbfish_astrocytes_intervention_ood",
+  "status": "complete",
+  "target_response_data_used": false,
+  "task_id": "intervention_ood_main"
+}

benchmark/intervention_descriptors/intervention_ood_main/perturbfish_astrocytes.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b3d85a139b316515e1501f55527649da1f8a8a30b90f4fffaf5cdc2bdab7d15
+size 19530

benchmark/protocols/context_ood_joint.yaml

@@ -0,0 +1,139 @@
+schema_version: 1
+id: context_ood_joint
+generalization_axis: cellular_context
+evaluation_mode: joint
+name: Cellular-context OOD joint RNA abundance and localization prediction
+status: planned
+data_tier: strict
+observation_unit: context_transition_population
+paired_counterfactual: false
+required_columns:
+- dataset_id
+- unit_id
+- sample_id
+- condition
+- cell_id
+- cell_type
+- gene
+- x
+- y
+split:
+  group_key: context
+  task_manifests:
+    context: benchmark/tasks/context_ood_lung.yaml
+    expansion: benchmark/tasks/context_ood_expansion.yaml
+    biological_unit: benchmark/tasks/context_ood_biological_unit.yaml
+    reaudited: benchmark/tasks/context_ood_reaudited.yaml
+  reference: scPerturBench-inspired Cellular_context_generalization/o.o.d.; local
+    full-vocabulary metrics are numerically non-equivalent
+  rule: jointly train the preregistered known-transition family on all non-test contexts
+    and hide every target response assigned to the held-out context group
+  fit_group: independent checkpoint unit; one or more response folds may share it
+  source_cells_in_test_context_available: true
+  source_and_target_population_support_gate: source enforced before inference; sealed
+    target enforced by evaluator
+  executable_context_keys:
+  - cell_type
+  - patient_id
+  - donor_id
+  transition_specific: false
+  transition_family_policy: frozen per fit group
+  biological_unit_role: leakage_audit_and_clustered_uncertainty
+  prohibited:
+  - held_out_context_target_cells_in_fit
+  - held_out_context_target_rna_in_fit
+  - target_population_size_at_inference
+count_policy:
+  mode: predicted_distribution
+  zeros: include every declared source cell without the evaluated gene
+  target_cell_count_used: false
+  source_population: held-out-context source-condition cells
+intervention_descriptor:
+  role: known_intervention_covariate_not_an_OOD_identity_prior
+  applies_to: learned_baselines_only
+  source: task-bound known-condition descriptor, either a frozen design-only artifact
+    or deterministic one-hot identity
+  target_response_data_used: false
+foundation_context:
+  conditioning: source_only
+  source_cell_policy: all_declared_source_cells_no_subsampling
+  target_response_data_used: false
+coordinate_policy:
+  eligible_semantics:
+  - decoded_molecule
+  - capture_site
+  capture_site_feature_types:
+  - dnb
+  - bead
+  - polony
+  - sequencing_cluster
+  - grid
+  event_representation: unit_event_rows
+  unit_event_value: 1
+  capture_site_gate: deterministic pre-normalization expansion plus artifact hashes,
+    global lineage, and exact sample/unit/condition/cell/gene reconciliation; no frozen
+    capture-site importer yet
+  aggregation: coordinate semantics and capture feature type are explicit non-pooled
+    strata
+  required_axes:
+  - x
+  - y
+  optional_axes:
+  - z
+  frame: cell_centric
+  factorization: predict_counts_then_localization
+  cell_weighting: equal_weight_per_nonempty_cell
+metrics:
+  primary: centroid_rmse_log1p
+  secondary:
+  - source_delta_pearson
+  - marginal_wasserstein_log1p
+  - multivariate_sliced_wasserstein_log1p
+  - dependency_rmse
+  - cell_balanced_spatial_swd
+  - spatial_target_gene_coverage
+  ranking: Count and spatial columns remain separate. Report unit and gene coverage;
+    no cross-dataset composite. Biological-unit uncertainty and global aggregation
+    are subsequent analyses.
+uncertainty:
+  implementation_status: planned_before_final_reporting
+  resampling_unit: biological_unit
+  single_unit_policy: NA_with_unit_coverage
+  prohibited_pseudoreplicates:
+  - genes
+  - cells
+  - molecules
+  - folds
+  - model_seeds
+eligible_datasets:
+- gse291460_spinal_ko_xenium
+- gse269719_kidney_iri_xenium
+- gut_microbiota_merfish
+- gse280376_cardiac_repair_xenium
+- gse296517_breast_fus_cosmx
+- gse296577_gbm_ohsv_xenium
+- gse325253_lung_drug_xenium
+- merfish_sickness
+- dss_colitis_merfish
+- gse295862_adipose_weightloss_xenium
+- uc_vedolizumab_cosmx
+candidate_datasets: []
+eligible_baselines: []
+success_artifacts:
+- result_json
+- exact_config
+- split_manifest
+- cache_manifest
+- seed
+- code_revision
+execution_protocol: null
+execution_note: Information contract only; no active execution entrypoint for this
+  setting.
+candidate_baselines:
+- count_matched_resampling
+- gaussian_affine_transport
+- radial_quantile_transport
+- scfoundation
+- scgpt
+- scgpt_spatial
+- spatialformer

benchmark/protocols/context_ood_oracle.yaml

@@ -0,0 +1,144 @@
+schema_version: 1
+id: context_ood_oracle
+generalization_axis: cellular_context
+evaluation_mode: oracle_count
+name: Cellular-context OOD conditional RNA localization with oracle target counts
+status: implemented
+data_tier: strict
+observation_unit: cell
+paired_counterfactual: false
+required_columns:
+- dataset_id
+- unit_id
+- sample_id
+- condition
+- cell_id
+- cell_type
+- gene
+- x
+- y
+split:
+  group_key: context
+  task_manifests:
+    context: benchmark/tasks/context_ood_lung.yaml
+    expansion: benchmark/tasks/context_ood_expansion.yaml
+    biological_unit: benchmark/tasks/context_ood_biological_unit.yaml
+    reaudited: benchmark/tasks/context_ood_reaudited.yaml
+  reference: scPerturBench-inspired Cellular_context_generalization/o.o.d.; local
+    full-vocabulary metrics are numerically non-equivalent
+  rule: jointly train the preregistered known-transition family on all non-test contexts
+    and hide every target response assigned to the held-out context group
+  fit_group: independent checkpoint unit; one or more response folds may share it
+  source_cells_in_test_context_available: true
+  source_and_target_population_support_gate: required
+  executable_context_keys:
+  - cell_type
+  - patient_id
+  - donor_id
+  transition_specific: false
+  transition_family_policy: frozen per fit group
+  biological_unit_role: leakage_audit_and_clustered_uncertainty
+  prohibited:
+  - held_out_context_target_cells_in_fit
+  - held_out_context_target_rna_in_fit
+  - random_cell_split
+  - random_molecule_split
+count_policy:
+  mode: exact_target_per_cell_gene
+  purpose: isolate conditional localization from abundance prediction
+  target_geometry_access: counts_only_physically_redacted
+foundation_context:
+  conditioning: source_only
+  source_cell_policy: all_declared_source_cells_no_subsampling
+  target_response_data_used: false
+  representation: individual_source_cell_embeddings
+coordinate_policy:
+  eligible_semantics:
+  - decoded_molecule
+  - capture_site
+  capture_site_feature_types:
+  - dnb
+  - bead
+  - polony
+  - sequencing_cluster
+  - grid
+  event_representation: unit_event_rows
+  unit_event_value: 1
+  capture_site_gate: deterministic pre-normalization expansion plus artifact hashes,
+    global lineage, and exact sample/unit/condition/cell/gene reconciliation; no frozen
+    capture-site importer yet
+  aggregation: coordinate semantics and capture feature type are explicit non-pooled
+    strata
+  required_axes:
+  - x
+  - y
+  optional_axes:
+  - z
+  frame: cell_centric
+  cell_weighting: equal_weight_per_nonempty_cell
+metrics:
+  primary: cell_balanced_sliced_wasserstein
+  secondary:
+  - spatial_centroid_rmse
+  - radial_wasserstein
+  - unpaired_cell_spatial_feature_sliced_wasserstein
+  - prediction_unit_coverage
+  ranking: average fine contexts within evaluation context group, groups within transition,
+    and transitions equally with metric-specific row coverage; no finite-only ranking
+    across unequal coverage; no composite score
+  expression_prediction: not_applicable_oracle_count_integrity_gate_only
+  rna_subsampling: false
+  gene_panels:
+  - all_gene
+  - top_100_deg
+uncertainty:
+  implementation_status: planned_before_final_reporting
+  resampling_unit: biological_unit
+  single_unit_policy: NA_with_unit_coverage
+  prohibited_pseudoreplicates:
+  - genes
+  - cells
+  - molecules
+  - folds
+  - model_seeds
+eligible_datasets:
+- gse291460_spinal_ko_xenium
+- gse269719_kidney_iri_xenium
+- gut_microbiota_merfish
+- gse280376_cardiac_repair_xenium
+- gse296517_breast_fus_cosmx
+- gse296577_gbm_ohsv_xenium
+- gse325253_lung_drug_xenium
+- merfish_sickness
+- dss_colitis_merfish
+- gse295862_adipose_weightloss_xenium
+- uc_vedolizumab_cosmx
+candidate_datasets: []
+eligible_baselines:
+- count_matched_resampling
+- gaussian_affine_transport
+- radial_quantile_transport
+- scgpt_spatial
+- spatialformer
+- scgpt
+- scfoundation
+success_artifacts:
+- result_json
+- exact_config
+- split_manifest
+- cache_manifest
+- seed
+- code_revision
+execution_protocol: scperturbench_scfoundation_cell_oracle
+execution_note: All registered baselines use scripts/run_cell_oracle.py; the current
+  complete-cell smoke gate is required for scientific matrix execution.
+training:
+  configuration: configs/experiments/cell_oracle.yaml
+  sampling: benchmark/sampling/scperturbench_scfoundation.yaml
+  initialization: independent_from_joint_checkpoints
+  input: complete_cell_counts_and_all_RNA_coordinates
+  objective: unpaired_complete_cell_distribution
+  count_head: absent
+  target_geometry: development_only; test geometry after all prediction seals
+  checkpoint_selection: final_fixed_15_epochs_on_frozen_development_union
+evaluation_unit: context_transition_cell_distribution

benchmark/protocols/intervention_ood_joint.yaml

@@ -0,0 +1,128 @@
+schema_version: 1
+id: intervention_ood_joint
+generalization_axis: perturbation_identity
+evaluation_mode: joint
+name: Intervention-OOD joint RNA abundance and localization prediction
+status: planned
+data_tier: strict
+observation_unit: intervention_population
+paired_counterfactual: false
+required_columns:
+- dataset_id
+- unit_id
+- sample_id
+- condition
+- cell_id
+- cell_type
+- gene
+- x
+- y
+split:
+  group_key: intervention_identity
+  task_manifests:
+    main: benchmark/tasks/intervention_ood_main.yaml
+    compositional: benchmark/tasks/intervention_ood_compositional.yaml
+  gears_interface:
+  - set2conditions
+  - test_perts
+  - test_subgroup
+  final_fit: train_plus_validation_after_hyperparameters_are_frozen
+  biological_unit_role: leakage_block_and_clustered_uncertainty
+  prohibited:
+  - test_target_cells_in_fit
+  - test_target_rna_in_fit
+  - target_population_size_at_inference
+count_policy:
+  mode: predicted_distribution
+  zeros: include every declared cell without the evaluated gene
+  target_cell_count_used: false
+  source_population: frozen source-condition population selected without target data
+intervention_descriptor:
+  applies_to: learned_baselines_only
+  provider: biomedbert_non_special_mean
+  model_revision: 1b3e7b435092c6f65dac0650f3d06df4866d6fd4
+  pooling: mean_last_hidden_non_special_tokens
+  component_normalization: l2
+  compositional_rule: additive_sum
+  reference_vector: zero
+  shared_identically_across_learned_baselines: true
+  target_response_data_used: false
+foundation_context:
+  conditioning: source_only
+  source_cell_policy: all_declared_source_cells_no_subsampling
+  batching_changes_selection: false
+  target_response_data_used: false
+coordinate_policy:
+  eligible_semantics:
+  - decoded_molecule
+  - capture_site
+  capture_site_feature_types:
+  - dnb
+  - bead
+  - polony
+  - sequencing_cluster
+  - grid
+  event_representation: unit_event_rows
+  unit_event_value: 1
+  capture_site_gate: deterministic pre-normalization expansion plus artifact hashes,
+    global lineage, and exact sample/unit/condition/cell/gene reconciliation; no frozen
+    capture-site importer yet
+  aggregation: coordinate semantics and capture feature type are explicit non-pooled
+    strata
+  required_axes:
+  - x
+  - y
+  optional_axes:
+  - z
+  frame: cell_centric
+  factorization: predict_counts_then_localization
+  cell_weighting: equal_weight_per_nonempty_cell
+metrics:
+  primary: centroid_rmse_log1p
+  secondary:
+  - source_delta_pearson
+  - marginal_wasserstein_log1p
+  - multivariate_sliced_wasserstein_log1p
+  - dependency_rmse
+  - cell_balanced_spatial_swd
+  - spatial_target_gene_coverage
+  ranking: Count and spatial columns remain separate. Report unit and gene coverage;
+    no cross-dataset composite. Biological-unit uncertainty and global aggregation
+    are subsequent analyses.
+uncertainty:
+  implementation_status: planned_before_final_reporting
+  resampling_unit: biological_unit
+  joint_resampling: independently_resample_source_prediction_and_target_units
+  single_unit_policy: NA_with_unit_coverage
+  prohibited_pseudoreplicates:
+  - genes
+  - cells
+  - molecules
+  - folds
+  - model_seeds
+eligible_datasets:
+- perturbfish_astrocytes
+- gse325253_lung_drug_xenium
+- merfish_sickness
+- gse246584_atm_radiation_xenium
+- gse285341_ddx54_pd1_xenium
+candidate_datasets: []
+eligible_baselines: []
+success_artifacts:
+- result_json
+- exact_config
+- split_manifest
+- cache_manifest
+- seed
+- code_revision
+execution_protocol: null
+execution_note: Information contract only; no active execution entrypoint for this
+  setting.
+candidate_baselines:
+- count_matched_resampling
+- gaussian_affine_transport
+- radial_quantile_transport
+- scfoundation
+- scgpt
+- scgpt_spatial
+- spatialformer

benchmark/protocols/intervention_ood_oracle.yaml

@@ -0,0 +1,140 @@
+schema_version: 1
+id: intervention_ood_oracle
+generalization_axis: perturbation_identity
+evaluation_mode: oracle_count
+name: Intervention-OOD conditional RNA localization with oracle target counts
+status: implemented
+data_tier: strict
+observation_unit: cell
+paired_counterfactual: false
+required_columns:
+- dataset_id
+- unit_id
+- sample_id
+- condition
+- cell_id
+- cell_type
+- gene
+- x
+- y
+split:
+  group_key: intervention_identity
+  task_manifests:
+    main: benchmark/tasks/intervention_ood_main.yaml
+    compositional: benchmark/tasks/intervention_ood_compositional.yaml
+  gears_interface:
+  - set2conditions
+  - test_perts
+  - test_subgroup
+  final_fit: train_plus_validation_after_hyperparameters_are_frozen
+  biological_unit_role: leakage_block_and_clustered_uncertainty
+  prohibited:
+  - test_target_cells_in_fit
+  - test_target_rna_in_fit
+  - random_cell
+  - random_molecule
+count_policy:
+  mode: exact_target_per_cell_gene
+  purpose: Isolate conditional localization from abundance prediction.
+  target_geometry_access: counts_only_physically_redacted
+intervention_descriptor:
+  applies_to: learned_baselines_only
+  provider: biomedbert_non_special_mean
+  model_revision: 1b3e7b435092c6f65dac0650f3d06df4866d6fd4
+  pooling: mean_last_hidden_non_special_tokens
+  component_normalization: l2
+  compositional_rule: additive_sum
+  reference_vector: zero
+  shared_identically_across_learned_baselines: true
+  target_response_data_used: false
+foundation_context:
+  conditioning: source_only
+  source_cell_policy: all_declared_source_cells_no_subsampling
+  batching_changes_selection: false
+  target_response_data_used: false
+  representation: individual_source_cell_embeddings
+coordinate_policy:
+  eligible_semantics:
+  - decoded_molecule
+  - capture_site
+  capture_site_feature_types:
+  - dnb
+  - bead
+  - polony
+  - sequencing_cluster
+  - grid
+  event_representation: unit_event_rows
+  unit_event_value: 1
+  capture_site_gate: deterministic pre-normalization expansion plus artifact hashes,
+    global lineage, and exact sample/unit/condition/cell/gene reconciliation; no frozen
+    capture-site importer yet
+  aggregation: coordinate semantics and capture feature type are explicit non-pooled
+    strata
+  required_axes:
+  - x
+  - y
+  optional_axes:
+  - z
+  frame: cell_centric
+  normalization: independently center and scale each cell by its own eligible-RNA
+    centroid and xy RMS radius; no cross-cell or test-fitted parameters
+  cell_weighting: equal_weight_per_nonempty_cell
+metrics:
+  primary: cell_balanced_sliced_wasserstein
+  secondary:
+  - spatial_centroid_rmse
+  - radial_wasserstein
+  - unpaired_cell_spatial_feature_sliced_wasserstein
+  - prediction_unit_coverage
+  ranking: Spatial fidelity and intervention specificity are reported separately.
+  expression_prediction: not_applicable_oracle_count_integrity_gate_only
+  rna_subsampling: false
+  gene_panels:
+  - all_gene
+  - top_100_deg
+uncertainty:
+  implementation_status: planned_before_final_reporting
+  resampling_unit: biological_unit
+  oracle_resampling: paired_prediction_and_truth_units
+  single_unit_policy: NA_with_unit_coverage
+  prohibited_pseudoreplicates:
+  - genes
+  - cells
+  - molecules
+  - folds
+  - model_seeds
+eligible_datasets:
+- perturbfish_astrocytes
+- gse325253_lung_drug_xenium
+- merfish_sickness
+- gse246584_atm_radiation_xenium
+- gse285341_ddx54_pd1_xenium
+candidate_datasets: []
+eligible_baselines:
+- count_matched_resampling
+- gaussian_affine_transport
+- radial_quantile_transport
+- scgpt_spatial
+- spatialformer
+- scgpt
+- scfoundation
+success_artifacts:
+- result_json
+- exact_config
+- split_manifest
+- cache_manifest
+- seed
+- code_revision
+execution_protocol: scperturbench_scfoundation_cell_oracle
+execution_note: All registered baselines use scripts/run_cell_oracle.py; the current
+  complete-cell smoke gate is required for scientific matrix execution.
+training:
+  configuration: configs/experiments/cell_oracle.yaml
+  sampling: benchmark/sampling/scperturbench_scfoundation.yaml
+  initialization: independent_from_joint_checkpoints
+  input: complete_cell_counts_and_all_RNA_coordinates
+  objective: unpaired_complete_cell_distribution
+  count_head: absent
+  target_geometry: development_only; test geometry after all prediction seals
+  checkpoint_selection: final_fixed_15_epochs_on_frozen_development_union
+evaluation_unit: context_transition_cell_distribution

benchmark/protocols/temporal_stage_ood_joint.yaml

@@ -0,0 +1,122 @@
+schema_version: 1
+id: temporal_stage_ood_joint
+generalization_axis: temporal_stage
+evaluation_mode: joint
+name: Temporal-stage OOD joint RNA abundance and localization prediction
+status: planned
+implementation:
+  split_loader: implemented
+  sealed_evaluator: sealed_manifest_interface_implemented
+  dataset_runner: pending_frozen_dataset_adapter
+data_tier: strict
+observation_unit: stage_context_population
+paired_counterfactual: false
+required_columns:
+- dataset_id
+- unit_id
+- sample_id
+- condition
+- cell_id
+- cell_type
+- gene
+- x
+- y
+split:
+  group_key: ordered_temporal_stage
+  reference: local extension beyond scPerturBench perturbation/context axes
+  rule: hold out one response stage while keeping the perturbation identity fixed
+    and visible at another post-perturbation stage
+  regimes:
+  - forward_extrapolation
+  - interpolation
+  forward_rule: no target-or-later response stage may enter fit
+  stage_descriptor: response-independent continuous numeric order; unseen one-hot
+    stage IDs are prohibited
+  biological_unit_role: leakage_audit_and_clustered_uncertainty
+  source_and_target_population_support_gate: required_by_sealed_interface
+  prohibited:
+  - held_out_stage_cells_in_fit
+  - held_out_stage_rna_in_fit
+  - target_population_size_at_inference
+  - target_or_later_stage_in_forward_fit
+count_policy:
+  mode: predicted_distribution
+  zeros: include every declared source cell without the evaluated gene
+  target_cell_count_used: false
+  source_population: frozen reference-stage population selected without target data
+coordinate_policy:
+  eligible_semantics:
+  - decoded_molecule
+  - capture_site
+  capture_site_feature_types:
+  - dnb
+  - bead
+  - polony
+  - sequencing_cluster
+  - grid
+  event_representation: unit_event_rows
+  unit_event_value: 1
+  capture_site_gate: deterministic pre-normalization expansion plus artifact hashes,
+    global lineage, and exact sample/unit/condition/cell/gene reconciliation; no frozen
+    capture-site importer yet
+  aggregation: coordinate semantics and capture feature type are explicit non-pooled
+    strata
+  required_axes:
+  - x
+  - y
+  optional_axes:
+  - z
+  frame: cell_centric_or_registered_subcellular_capture_site
+  factorization: predict_counts_then_localization
+  cell_weighting: equal_weight_per_nonempty_cell
+metrics:
+  primary: control_reference_pearson_delta
+  secondary:
+  - expression_mse
+  - expression_rmse
+  - log1p_count_energy_distance
+  - smoothed_count_symmetric_kl
+  - log1p_count_wasserstein
+  - expression_centroid_accuracy
+  - cell_balanced_joint_sliced_wasserstein
+  - spatial_centroid_accuracy
+  - spatial_residual_pearson
+  - radial_wasserstein
+  - predicted_empty_rate
+  - prediction_unit_coverage
+  ranking: expression and spatial columns remain separate; stages are macro-averaged
+    equally within forward/interpolation regime; report metric-specific row coverage
+    and do not rank finite-only values across unequal coverage
+uncertainty:
+  implementation_status: planned_before_final_reporting
+  resampling_unit: biological_unit
+  single_unit_policy: NA_with_unit_coverage
+  prohibited_pseudoreplicates:
+  - genes
+  - cells
+  - molecules
+  - folds
+  - model_seeds
+eligible_datasets: []
+candidate_datasets:
+- dss_colitis_merfish
+external_candidate_datasets: []
+eligible_baselines: []
+candidate_baselines:
+- count_matched_resampling
+- gaussian_affine_transport
+- radial_quantile_transport
+- scfoundation
+- scgpt
+- scgpt_spatial
+- spatialformer
+success_artifacts:
+- result_json
+- exact_config
+- split_manifest
+- cache_manifest
+- seed
+- code_revision
+execution_protocol: null
+execution_note: Information contract only; no active execution entrypoint for this
+  setting.

benchmark/protocols/temporal_stage_ood_oracle.yaml

@@ -0,0 +1,114 @@
+schema_version: 1
+id: temporal_stage_ood_oracle
+generalization_axis: temporal_stage
+evaluation_mode: oracle_count
+name: Temporal-stage OOD conditional RNA localization with oracle target counts
+status: planned
+implementation:
+  split_loader: implemented
+  sealed_evaluator: sealed_manifest_interface_implemented
+  dataset_runner: pending_frozen_dataset_adapter
+data_tier: strict
+observation_unit: stage_context_population
+paired_counterfactual: false
+required_columns:
+- dataset_id
+- unit_id
+- sample_id
+- condition
+- cell_id
+- cell_type
+- gene
+- x
+- y
+split:
+  group_key: ordered_temporal_stage
+  reference: local extension beyond scPerturBench perturbation/context axes
+  rule: hold out one response stage while keeping the perturbation identity fixed
+    and visible at another post-perturbation stage
+  regimes:
+  - forward_extrapolation
+  - interpolation
+  forward_rule: no target-or-later response stage may enter fit
+  stage_descriptor: response-independent continuous numeric order; unseen one-hot
+    stage IDs are prohibited
+  biological_unit_role: leakage_audit_and_clustered_uncertainty
+  source_and_target_population_support_gate: required_by_sealed_interface
+  prohibited:
+  - held_out_stage_cells_in_fit
+  - held_out_stage_rna_in_fit
+  - target_or_later_stage_in_forward_fit
+  - random_cell_split
+  - random_molecule_split
+count_policy:
+  mode: exact_target_per_cell_gene
+  purpose: isolate localization from abundance prediction
+  target_geometry_access: counts_only_physically_redacted_required_of_future_runner
+coordinate_policy:
+  eligible_semantics:
+  - decoded_molecule
+  - capture_site
+  capture_site_feature_types:
+  - dnb
+  - bead
+  - polony
+  - sequencing_cluster
+  - grid
+  event_representation: unit_event_rows
+  unit_event_value: 1
+  capture_site_gate: deterministic pre-normalization expansion plus artifact hashes,
+    global lineage, and exact sample/unit/condition/cell/gene reconciliation; no frozen
+    capture-site importer yet
+  aggregation: coordinate semantics and capture feature type are explicit non-pooled
+    strata
+  required_axes:
+  - x
+  - y
+  optional_axes:
+  - z
+  frame: cell_centric_or_registered_subcellular_capture_site
+  cell_weighting: equal_weight_per_nonempty_cell
+metrics:
+  primary: cell_balanced_sliced_wasserstein
+  secondary:
+  - spatial_centroid_accuracy
+  - spatial_residual_pearson
+  - radial_wasserstein
+  - predicted_empty_rate
+  - prediction_unit_coverage
+  expression_prediction: not_applicable_oracle_count_integrity_gate_only
+  ranking: report forward-extrapolation and interpolation separately with metric-specific
+    row coverage; no finite-only ranking across unequal coverage; no composite score
+uncertainty:
+  implementation_status: planned_before_final_reporting
+  resampling_unit: biological_unit
+  single_unit_policy: NA_with_unit_coverage
+  prohibited_pseudoreplicates:
+  - genes
+  - cells
+  - molecules
+  - folds
+  - model_seeds
+eligible_datasets: []
+candidate_datasets:
+- dss_colitis_merfish
+external_candidate_datasets: []
+eligible_baselines: []
+candidate_baselines:
+- count_matched_resampling
+- gaussian_affine_transport
+- radial_quantile_transport
+- scfoundation
+- scgpt
+- scgpt_spatial
+- spatialformer
+success_artifacts:
+- result_json
+- exact_config
+- split_manifest
+- cache_manifest
+- seed
+- code_revision
+execution_protocol: null
+execution_note: Information contract only; no active execution entrypoint for this
+  setting.

benchmark/registry/baselines.yaml

@@ -0,0 +1,188 @@
+schema_version: 1
+baselines:
+- id: count_matched_resampling
+  name: Count-matched control-coordinate resampling
+  family: mandatory_null
+  native_to_task: true
+  implementation_status: adapter_ready
+  evidence: model_test_only
+  intervention_ood_evidence: model_test_only
+  entrypoint: src/subcellpert/main.py
+  trainer: src/subcellpert/models/count_matched_resampling.py
+  protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  experiment_ids: []
+  notes: Independent joint/oracle fits share the complete-cell interface. Data-split
+    seeds are the only repetition axis; planned jobs do not establish training completion.
+  planned_protocols: []
+  execution_protocol: scperturbench_scfoundation_cell_oracle
+  configuration: configs/experiments/cell_oracle.yaml
+  model: src/subcellpert/models/count_matched_resampling.py
+  joint_configuration: configs/experiments/cell_joint.yaml
+- id: gaussian_affine_transport
+  name: Per-mark Gaussian affine transport
+  family: classical_ml_transport
+  native_to_task: true
+  implementation_status: adapter_ready
+  evidence: model_test_only
+  intervention_ood_evidence: model_test_only
+  entrypoint: src/subcellpert/main.py
+  trainer: src/subcellpert/models/gaussian_affine_transport.py
+  protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  experiment_ids: []
+  notes: Independent joint/oracle fits share the complete-cell interface. Data-split
+    seeds are the only repetition axis; planned jobs do not establish training completion.
+  planned_protocols: []
+  execution_protocol: scperturbench_scfoundation_cell_oracle
+  configuration: configs/experiments/cell_oracle.yaml
+  model: src/subcellpert/models/gaussian_affine_transport.py
+  joint_configuration: configs/experiments/cell_joint.yaml
+- id: radial_quantile_transport
+  name: Per-mark radial quantile transport
+  family: classical_ml_transport
+  native_to_task: true
+  implementation_status: adapter_ready
+  evidence: model_test_only
+  intervention_ood_evidence: model_test_only
+  entrypoint: src/subcellpert/main.py
+  trainer: src/subcellpert/models/radial_quantile_transport.py
+  protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  experiment_ids: []
+  notes: Independent joint/oracle fits share the complete-cell interface. Data-split
+    seeds are the only repetition axis; planned jobs do not establish training completion.
+  planned_protocols: []
+  execution_protocol: scperturbench_scfoundation_cell_oracle
+  configuration: configs/experiments/cell_oracle.yaml
+  model: src/subcellpert/models/radial_quantile_transport.py
+  joint_configuration: configs/experiments/cell_joint.yaml
+- id: scgpt_spatial
+  name: scGPT-spatial frozen-context transport
+  family: subcellular_foundation_adapter
+  native_to_task: false
+  implementation_status: adapter_ready
+  evidence: model_test_only
+  intervention_ood_evidence: model_test_only
+  entrypoint: src/subcellpert/main.py
+  trainer: src/subcellpert/train_one_step.py
+  source_cell_policy: all_declared_source_cells_no_subsampling
+  protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  experiment_ids: []
+  notes: Independent joint/oracle fits share the complete-cell interface. Data-split
+    seeds are the only repetition axis; planned jobs do not establish training completion.
+  planned_protocols: []
+  execution_protocol: scperturbench_scfoundation_cell_oracle
+  configuration: configs/experiments/cell_oracle.yaml
+  foundation_representation: frozen_source_cell_per_gene_features
+  model: src/subcellpert/models/scgpt_spatial.py
+  joint_configuration: configs/experiments/cell_joint.yaml
+- id: spatialformer
+  name: SpatialFormer frozen-context transport
+  family: subcellular_foundation_adapter
+  native_to_task: false
+  implementation_status: adapter_ready
+  evidence: model_test_only
+  intervention_ood_evidence: model_test_only
+  entrypoint: src/subcellpert/main.py
+  trainer: src/subcellpert/train_one_step.py
+  source_cell_policy: all_declared_source_cells_no_subsampling
+  protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  experiment_ids: []
+  notes: Independent joint/oracle fits share the complete-cell interface. Data-split
+    seeds are the only repetition axis; planned jobs do not establish training completion.
+  planned_protocols: []
+  execution_protocol: scperturbench_scfoundation_cell_oracle
+  configuration: configs/experiments/cell_oracle.yaml
+  foundation_representation: frozen_source_cell_per_gene_features
+  model: src/subcellpert/models/spatialformer.py
+  joint_configuration: configs/experiments/cell_joint.yaml
+- id: scgpt
+  name: scGPT frozen-context transport
+  family: gene_expression_foundation_adapter
+  native_to_task: false
+  implementation_status: adapter_ready
+  evidence: model_test_only
+  intervention_ood_evidence: model_test_only
+  entrypoint: src/subcellpert/main.py
+  trainer: src/subcellpert/train_one_step.py
+  source_cell_policy: all_declared_source_cells_no_subsampling
+  protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  experiment_ids: []
+  notes: Independent joint/oracle fits share the complete-cell interface. Data-split
+    seeds are the only repetition axis; planned jobs do not establish training completion.
+  planned_protocols: []
+  execution_protocol: scperturbench_scfoundation_cell_oracle
+  configuration: configs/experiments/cell_oracle.yaml
+  foundation_representation: frozen_source_cell_per_gene_features
+  model: src/subcellpert/models/scgpt.py
+  joint_configuration: configs/experiments/cell_joint.yaml
+- id: scfoundation
+  name: scFoundation frozen-context transport
+  family: gene_expression_foundation_adapter
+  native_to_task: false
+  implementation_status: adapter_ready
+  evidence: model_test_only
+  intervention_ood_evidence: model_test_only
+  entrypoint: src/subcellpert/main.py
+  trainer: src/subcellpert/train_one_step.py
+  source_cell_policy: all_declared_source_cells_no_subsampling
+  protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  experiment_ids: []
+  notes: Independent joint/oracle fits share the complete-cell interface. Data-split
+    seeds are the only repetition axis; planned jobs do not establish training completion.
+  planned_protocols: []
+  execution_protocol: scperturbench_scfoundation_cell_oracle
+  configuration: configs/experiments/cell_oracle.yaml
+  foundation_representation: frozen_source_cell_per_gene_features
+  model: src/subcellpert/models/scfoundation.py
+  joint_configuration: configs/experiments/cell_joint.yaml
+- id: state_se
+  name: STATE-SE frozen per-gene encoder with GEARS spatial transport
+  family: gene_expression_foundation_adapter
+  native_to_task: false
+  implementation_status: adapter_ready
+  evidence: model_test_only
+  intervention_ood_evidence: model_test_only
+  entrypoint: src/subcellpert/main.py
+  trainer: src/subcellpert/train_one_step.py
+  source_cell_policy: all_declared_source_cells_no_subsampling
+  protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  experiment_ids: []
+  notes: Independent joint/oracle fits share the complete-cell interface. Data-split
+    seeds are the only repetition axis; planned jobs do not establish training completion.
+  planned_protocols: []
+  execution_protocol: scperturbench_scfoundation_cell_oracle
+  configuration: configs/experiments/cell_oracle.yaml
+  foundation_representation: frozen_source_cell_per_gene_features
+  model: src/subcellpert/models/state_se.py
+  joint_configuration: configs/experiments/cell_joint.yaml

benchmark/registry/datasets.yaml

@@ -0,0 +1,368 @@
+schema_version: 1
+datasets:
+- id: gse291460_spinal_ko_xenium
+  organ: Spinal cord
+  perturbation_types:
+  - Genetic
+  diversity_note: Spinal cord; four developmental knockout responses.
+  name: Spinal developmental knockout Xenium
+  accession: GSE291460
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: animal
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/gse291460_spinal_ko_xenium/manifest.json
+  eligible_protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - cell_type
+  - animal
+  - anatomical_region
+  notes: All eight complete native samples match deposited cell-feature counts exactly.
+    Final author coordinate-table labels retain 132,930 cells and 38,473,529 endogenous
+    RNA across 347 genes. Animal namespaces precede barcode matching; Brach and Thor
+    controls remain anatomically distinct. Four known KO responses share each held-cell-type
+    fit. Missing author labels are not reconstructed with a local classifier, and
+    target types without same-region control support are excluded only from task views.
+    Neither unseen perturbation identity nor unseen animal is the OOD axis.
+- id: gse296577_gbm_ohsv_xenium
+  organ: Brain
+  perturbation_types:
+  - Biologic
+  diversity_note: Glioblastoma; oncolytic HSV treatment-associated recurrence.
+  name: Glioblastoma oncolytic HSV Xenium
+  accession: GSE296577
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: patient
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/gse296577_gbm_ohsv_xenium/manifest.json
+  eligible_protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - cell_type
+  - patient
+  notes: 'All twenty native ROIs were reconciled to the complete deposited Seurat
+    raw assay: 2,401,946 cells, 520,975,252 RNA and 480 genes, with zero cell-gene
+    differences and no unmatched author cells. One ROI has no cells in final author
+    QC. Pre/post samples remain grouped within each of eight patients. The known post-oHSV
+    recurrence response supports cell-type OOD; neither unseen-patient generalization
+    nor an isolated causal treatment effect is claimed.'
+- id: gse296517_breast_fus_cosmx
+  organ: Breast
+  perturbation_types:
+  - Ultrasound
+  - Chemical
+  diversity_note: Breast cancer; focused ultrasound with or without gemcitabine.
+  name: Breast cancer focused-ultrasound CosMx
+  accession: GSE296517
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: clinical_patient_pool
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/gse296517_breast_fus_cosmx/manifest.json
+  eligible_protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - cell_type
+  notes: All four runs exactly reconstruct 880,617 author-labelled cells and 102,102,989
+    endogenous RNA over 1,000 genes. Pooled untreated controls support known FUS and
+    FUS+GEM responses under donor_patient_split=false. Two pretreatment biopsies share
+    one run without a machine-readable patient crosswalk; clinical runs are conservatively
+    grouped into one storage unit. No patient OOD, subtype-matched causal effect,
+    or isolated GEM synergy is claimed. Exposure status is known even where lesion
+    pathology is uncertain.
+- id: gut_microbiota_merfish
+  organ: Intestine
+  perturbation_types:
+  - Microbiota
+  diversity_note: Small-intestine and colon regions; SPF to germ-free condition.
+  name: Gut microbiota MERFISH
+  accession: Dryad-10.5061/dryad.p5hqbzm0z
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: specimen
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/gut_microbiota_merfish/manifest.json
+  eligible_protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - cell_type
+  - specimen
+  - region
+  notes: All 2,060,051 deposited cells and 189,468,148 eligible RNA exactly match
+    the complete author 1,815-gene raw_counts axis after fixed contamination QC. Four
+    same-region SPF-to-GF transitions share each held-cell-type fit. The 28 acquisition/specimen
+    groups contain 105 slices; no undocumented animal crosswalk or independent-mouse
+    replicate count is asserted. Unsupported same-region target contexts are removed
+    from views only.
+- id: gse280376_cardiac_repair_xenium
+  organ: Heart
+  perturbation_types:
+  - Surgery
+  diversity_note: Heart; myocardial infarction induced by surgical coronary ligation.
+  name: Cardiac myocardial-infarction repair Xenium
+  accession: GSE280376
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: specimen
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/gse280376_cardiac_repair_xenium/manifest.json
+  eligible_protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - cell_type
+  - specimen
+  notes: Seven complete, correctly annotated samples retain 225,696 cells and 61,244,213
+    RNA. Deposited nico membership and its cluster-name crosswalk provide contexts.
+    GSM9460460 is excluded because its annotation duplicates another slide and does
+    not join its native barcodes. Sham-to-MI day7/day28 responses share each context
+    fit; target contexts lacking Sham are excluded only from task views. No independent
+    animal count is inferred from slide IDs.
+- id: gse269719_kidney_iri_xenium
+  organ: Kidney
+  perturbation_types:
+  - Surgery
+  diversity_note: Kidney; surgical ischemia-reperfusion injury.
+  name: Kidney ischemia-reperfusion injury Xenium
+  accession: GSE269719
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: mouse
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/gse269719_kidney_iri_xenium/manifest.json
+  eligible_protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - cell_type
+  - mouse
+  notes: 'All twelve kidneys pass exact native Baysor RNA versus author counts: 1,374,915
+    cells, 168,519,194 RNA, 300 genes and 20 deposited cell types. Sham is the source
+    for five known injury/recovery timepoints, not unseen perturbation identities.
+    L/R kidneys belong to the same mouse; six mice total, one per timepoint. Three
+    context seeds are frozen without rerunning the original nine datasets. Admission
+    does not claim completed model fits.'
+- id: perturbfish_astrocytes
+  organ: Brain
+  perturbation_types:
+  - Genetic
+  diversity_note: Brain-derived astrocyte culture; single-guide CRISPRi.
+  name: Perturb-FISH astrocytes
+  accession: BIL-10.35077/ace-gem-get
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: cell
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/perturbfish_astrocytes/manifest.json
+  eligible_protocols:
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  split_units:
+  - intervention_identity
+  - well
+  notes: Exact reconstruction matches all 14,926 published count vectors; the benchmark
+    cache retains 12,748 one-guide cells spanning all 127 single targets and excludes
+    every multi-guide cell. The active three seeds each hold 90/11/26 single-target
+    identities for training/validation/test; reference controls remain visible.
+- id: gse325253_lung_drug_xenium
+  organ: Lung
+  perturbation_types:
+  - Chemical
+  diversity_note: 'Formal drug family only: AZD3965 and VB124 added after bleomycin.'
+  name: Lung fibrosis drug Xenium
+  accession: GSE325253
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: section
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/gse325253_lung_drug_xenium/manifest.json
+  eligible_protocols:
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - intervention_identity
+  - cell_type
+  - section
+  notes: Eight sections, two per condition, with 11 level-2 cellular contexts. The
+    active three context seeds retain only drug (Bleomycin to both add-ons jointly);
+    injury results are archived. Historical exhaustive context and intervention holdouts
+    are not the current training schedule.
+- id: gse246584_atm_radiation_xenium
+  organ: Brain
+  perturbation_types:
+  - Genetic
+  - Radiation
+  diversity_note: Diffuse midline glioma; ATM loss and focal radiotherapy.
+  name: ATM loss and focal-radiation Xenium
+  accession: GSE246584
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: section
+  data_status: ready
+  adapter_status: implemented
+  paper_role: secondary
+  manifest: datasets/gse246584_atm_radiation_xenium/manifest.json
+  eligible_protocols:
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  split_units:
+  - intervention_component
+  - section
+  notes: Exact 2x2 factorial release; the combination is a single combo_seen2 combinatorial
+    stress test, not a replicated main leaderboard.
+- id: gse285341_ddx54_pd1_xenium
+  organ: Lung
+  perturbation_types:
+  - Genetic
+  - Biologic
+  diversity_note: Lung-tumor model; Ddx54 knockdown and anti-PD-1 antibody.
+  name: Ddx54 knockdown and anti-PD-1 Xenium
+  accession: GSE285341
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: section
+  data_status: ready
+  adapter_status: implemented
+  paper_role: secondary
+  manifest: datasets/gse285341_ddx54_pd1_xenium/manifest.json
+  eligible_protocols:
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  split_units:
+  - intervention_component
+  - section
+  notes: Exact 2x2 factorial release reconstructed from source masks; the combination
+    is a single combo_seen2 combinatorial stress test.
+- id: merfish_sickness
+  organ: Brain
+  perturbation_types:
+  - Chemical
+  - Biologic
+  diversity_note: Hypothalamus; DSS chemical challenge and LPS/poly(I:C)/PLA2/STAg
+    immune stimuli.
+  name: Sickness MERFISH
+  accession: Dryad-10.5061/dryad.v6wwpzh98
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: animal
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/merfish_sickness/manifest.json
+  eligible_protocols:
+  - intervention_ood_oracle
+  - intervention_ood_joint
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - intervention_identity
+  - cell_type
+  - animal
+  notes: After the author ignore_cell flag, the uncapped cache has 1,765,570 cells
+    and 1,101,347,616 RNA. Task admission excludes one unsupported PLA/Monocytes target
+    while retaining its raw observation. Three held-context fits jointly learn five
+    known challenge responses; historical exhaustive/zero-shot folds are not active.
+- id: dss_colitis_merfish
+  organ: Intestine
+  perturbation_types:
+  - Chemical
+  diversity_note: Colon; DSS-induced colitis and recovery.
+  name: DSS colitis time-course MERFISH
+  accession: Dryad-10.5061/dryad.rjdfn2zh3-v5
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: animal_section
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/dss_colitis_merfish/manifest.json
+  eligible_protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - cell_type
+  - day
+  - mouse
+  notes: The complete 58-slice cache has 1,439,542 official annotated cells and 119,553,365
+    RNA. Eight Tier-1 labels provide candidate contexts; the active three seeds jointly
+    fit known post-D0 responses and hold one context each. Adipose has no D35 target
+    cells, so that response is not evaluated. Temporal-stage holdouts and exhaustive
+    context folds are historical/auxiliary, not the active schedule.
+- id: gse295862_adipose_weightloss_xenium
+  organ: Adipose
+  perturbation_types:
+  - Surgery
+  diversity_note: Subcutaneous adipose before and after bariatric-surgery-associated
+    weight loss.
+  name: Adipose obesity and weight-loss Xenium
+  accession: GSE295862/SCP3117
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: paired_donor
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/gse295862_adipose_weightloss_xenium/manifest.json
+  eligible_protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - cell_type
+  - paired_donor
+  - condition
+  notes: Exact full-release reconstruction matches all 25,795 author cells and 3,795,009
+    assigned RNA. The strict Obese-to-Weightloss cache uses 14,007 cells, 1,911,887
+    RNA, 477 genes, four paired donors, and seven level-1 context labels. Three held-context
+    seeds are active; Lean is excluded from this directed task. Historical GPU smoke
+    receipts are not completion of the current formal fits.
+- id: uc_vedolizumab_cosmx
+  organ: Intestine
+  perturbation_types:
+  - Biologic
+  diversity_note: Colon; vedolizumab antibody treatment.
+  name: Ulcerative colitis vedolizumab longitudinal CosMx
+  accession: Figshare-23896959-v3
+  tier: strict
+  spatial_unit: molecule
+  intervention_unit: patient
+  data_status: ready
+  adapter_status: implemented
+  paper_role: main
+  manifest: datasets/uc_vedolizumab_cosmx/manifest.json
+  eligible_protocols:
+  - context_ood_oracle
+  - context_ood_joint
+  split_units:
+  - patient_id
+  - response_stratum
+  - pre_post
+  notes: The adapter and seven-patient compact population payload are complete for
+    separate responder and non-responder known-transition branches. Every biopsy and
+    FOV remains within patient; the non-responder branch has only three complete pairs.
+    The active context view uses 29 deposited cell-type labels and three held-context
+    seeds, not patient holdouts. Both retrospective outcome-conditioned responses
+    share each fit; future response strata are not a prospective prediction input.

benchmark/registry/metrics.yaml

@@ -0,0 +1,279 @@
+schema_version: 1
+metrics:
+- id: control_reference_pearson_delta
+  component: expression_context_response
+  status: implemented
+  direction: higher
+  aggregation: Pearson correlation between true and predicted zero-inclusive mean-log1p
+    response vectors after subtracting the held-out-context source vector; macro-average
+    cell types within transition and transitions equally.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: expression_mse
+  component: abundance
+  status: implemented
+  direction: lower
+  aggregation: Mean squared error between predicted and true held-out-context condition-by-gene
+    zero-inclusive mean-log1p count vectors; included as a scPerturBench-inspired
+    but numerically non-equivalent local diagnostic.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: perturbed_reference_pearson_delta
+  component: expression_specificity
+  status: implemented
+  direction: higher
+  aggregation: Systema Pearson correlation on zero-inclusive mean-log1p per-cell count
+    vectors after subtracting the equal-condition mean of development perturbation
+    centroids; macro-average first over held-out interventions.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: perturbed_reference_pearson_delta20
+  component: expression_specificity
+  status: implemented
+  direction: higher
+  aggregation: Systema Pearson-delta formula on the sealed evaluator's top 20 absolute
+    target-versus-source effects in the zero-inclusive mean-log1p count vector; macro-average
+    first over held-out interventions.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: expression_centroid_accuracy
+  component: expression_retrieval
+  status: implemented
+  direction: higher
+  aggregation: Fraction of other held-out true zero-inclusive mean-log1p expression
+    centroids farther from a prediction than its matching centroid; unique out-of-fold
+    conditions form a global retrieval set, with fold-local retrieval when an identity
+    recurs across simulation folds.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: expression_rmse
+  component: abundance
+  status: implemented
+  direction: lower
+  aggregation: RMSE between predicted and true held-out condition-by-gene zero-inclusive
+    mean-log1p per-cell count vectors.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: cell_balanced_sliced_wasserstein
+  component: localization
+  status: implemented
+  direction: lower
+  aggregation: Sliced 1-Wasserstein in normalized local xy with every nonempty cell
+    contributing equal total mass; macro-average first over intervention identity.
+  count_conditioning: exact_target_count
+  scope: scperturbench_scfoundation_cell_oracle
+  entrypoint: src/subcellpert/cell_evaluation.py
+- id: cell_balanced_joint_sliced_wasserstein
+  component: localization_under_predicted_abundance
+  status: implemented
+  direction: lower
+  aggregation: Cell-balanced sliced 1-Wasserstein between true points and coordinates
+    generated under predicted per-cell counts; reported separately from abundance
+    metrics.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: spatial_centroid_accuracy
+  component: localization_retrieval
+  status: implemented
+  direction: higher
+  aggregation: Fraction of other held-out intervention point patterns farther from
+    a prediction than its matching truth under the same cell-balanced spatial distance;
+    retrieval is fold-local when an identity recurs across simulation folds.
+  count_conditioning: both_protocol_specific
+  scope: utility_or_planned_diagnostic
+- id: spatial_residual_pearson
+  component: localization_specificity
+  status: implemented
+  direction: higher
+  aggregation: Pearson correlation between true and predicted fixed multiscale occupancy
+    vectors after subtracting the equal-condition development reference.
+  count_conditioning: both_protocol_specific
+  scope: utility_or_planned_diagnostic
+- id: predicted_empty_rate
+  component: localization_coverage
+  status: implemented
+  direction: lower
+  aggregation: Fraction of true-nonempty intervention/cell-type/gene groups for which
+    the predictor emits no spatial RNA; reported with spatial metric coverage and
+    never converted into a finite distance.
+  count_conditioning: both_protocol_specific
+  scope: utility_or_planned_diagnostic
+- id: prediction_unit_coverage
+  component: localization_coverage
+  status: implemented
+  direction: higher
+  aggregation: Fraction of relevant biological-unit prediction populations with defined
+    generated geometry; target-empty oracle units are excluded from the denominator
+    and no cell or gene is filtered.
+  count_conditioning: exact_target_count
+  scope: scperturbench_scfoundation_cell_oracle
+  entrypoint: src/subcellpert/cell_evaluation.py
+- id: systematic_variation_gap
+  component: expression_diagnostic
+  status: implemented
+  direction: diagnostic
+  aggregation: Perturbed-reference Pearson-delta minus control-reference Pearson-delta,
+    reported to expose generic all-perturbed shifts.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: macro_sliced_wasserstein
+  component: localization
+  status: implemented
+  direction: lower
+  aggregation: Mean over eligible intervention-gene groups after per-cell normalization.
+  count_conditioning: exact_target_count
+  scope: utility_or_planned_diagnostic
+- id: relative_swd_improvement
+  component: localization
+  status: implemented
+  direction: higher
+  aggregation: Relative improvement over count-matched resampling with group bootstrap
+    confidence interval.
+  count_conditioning: exact_target_count
+  scope: utility_or_planned_diagnostic
+- id: radial_wasserstein
+  component: localization
+  status: implemented
+  direction: lower
+  aggregation: Cell-balanced one-dimensional Wasserstein distance between per-cell
+    molecule radii; macro-average first over intervention identity.
+  count_conditioning: exact_target_count
+  scope: scperturbench_scfoundation_cell_oracle
+  entrypoint: src/subcellpert/cell_evaluation.py
+- id: count_wasserstein
+  component: abundance
+  status: implemented_diagnostic
+  direction: lower
+  aggregation: Wasserstein distance between per-cell count distributions.
+  count_conditioning: not_applicable
+  scope: utility_or_planned_diagnostic
+- id: log1p_count_wasserstein
+  component: abundance
+  status: implemented
+  direction: lower
+  aggregation: Macro Wasserstein distance between log1p per-cell count distributions,
+    including zero-expression cells.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: log1p_count_energy_distance
+  component: abundance_distribution
+  status: implemented
+  direction: lower
+  aggregation: Exact one-dimensional energy distance between predicted and true log1p
+    per-cell counts for each gene, including zeros; average the complete frozen gene
+    vocabulary and then macro-average the held-out OOD identities without evaluation-cell
+    subsampling.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: smoothed_count_symmetric_kl
+  component: abundance_distribution
+  status: implemented
+  direction: lower
+  aggregation: Symmetric KL between empirical integer per-cell count histograms with
+    a fixed Jeffreys pseudocount of 0.5; average the complete frozen gene vocabulary
+    and then macro-average held-out OOD identities.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: joint_sliced_wasserstein
+  component: localization_under_predicted_abundance
+  status: implemented
+  direction: lower
+  aggregation: Macro sliced-Wasserstein distance between target coordinates and the
+    point population generated from predicted per-cell counts.
+  count_conditioning: predicted
+  scope: utility_or_planned_diagnostic
+- id: energy_distance
+  component: localization
+  status: planned
+  direction: lower
+  aggregation: Sample-aware macro average with bootstrap confidence interval.
+  count_conditioning: exact_target_count
+  scope: utility_or_planned_diagnostic
+- id: multiscale_occupancy_mmd
+  component: localization
+  status: implemented_legacy_diagnostic
+  direction: lower
+  aggregation: Legacy pilot diagnostic; excluded from intervention-OOD because pseudo-control
+    calibration is incomplete.
+  count_conditioning: exact_target_count
+  scope: utility_or_planned_diagnostic
+- id: cross_k_discrepancy
+  component: colocalization
+  status: implemented_diagnostic
+  direction: lower
+  aggregation: Mean absolute discrepancy over the predefined marked cross-K radii,
+    aggregated over intervention groups and biological wells.
+  count_conditioning: exact_target_count
+  scope: utility_or_planned_diagnostic
+- id: centroid_rmse_log1p
+  component: abundance
+  status: planned
+  direction: lower
+  aggregation: RMSE between equal-unit, equal-cell mean log1p count vectors.
+  count_conditioning: predicted
+  scope: planned_joint_setting
+- id: source_delta_pearson
+  component: expression_context_response
+  status: planned
+  direction: higher
+  aggregation: Pearson correlation of predicted and observed mean log1p shifts from
+    source; constants are NA; not Systema perturbed-reference delta.
+  count_conditioning: predicted
+  scope: planned_joint_setting
+- id: marginal_wasserstein_log1p
+  component: abundance_distribution
+  status: planned
+  direction: lower
+  aggregation: Mean exact weighted empirical W1 over genes; supports unequal cell
+    counts.
+  count_conditioning: predicted
+  scope: planned_joint_setting
+- id: multivariate_sliced_wasserstein_log1p
+  component: abundance_distribution
+  status: planned
+  direction: lower
+  aggregation: Mean weighted W1 over 64 frozen full-panel random projections.
+  count_conditioning: predicted
+  scope: planned_joint_setting
+- id: dependency_rmse
+  component: abundance_dependency
+  status: planned
+  direction: lower
+  aggregation: Population-standardized correlation discrepancy over genes variable
+    in both populations; report gene coverage.
+  count_conditioning: predicted
+  scope: planned_joint_setting
+- id: cell_balanced_spatial_swd
+  component: localization
+  status: planned
+  direction: lower
+  aggregation: Conditional spatial SWD with equal supported-unit, expressing-cell
+    and RNA weighting; 128 points per gene, 64 angles.
+  count_conditioning: predicted
+  scope: planned_joint_setting
+- id: spatial_target_gene_coverage
+  component: localization_coverage
+  status: planned
+  direction: higher
+  aggregation: Fraction of target-present genes with predicted geometry; incomplete
+    coverage prevents complete spatial ranking.
+  count_conditioning: predicted
+  scope: planned_joint_setting
+- id: spatial_centroid_rmse
+  component: localization
+  status: implemented
+  direction: lower
+  count_conditioning: exact_target_count
+  scope: scperturbench_scfoundation_cell_oracle
+  entrypoint: src/subcellpert/cell_evaluation.py
+  aggregation: Exact RNA-weighted gene centroid RMSE, equal supported units then expressing
+    cells then RNA.
+- id: unpaired_cell_spatial_feature_sliced_wasserstein
+  component: within_cell_multigene_geometry
+  status: implemented
+  direction: lower
+  count_conditioning: exact_target_count
+  scope: scperturbench_scfoundation_cell_oracle
+  entrypoint: src/subcellpert/cell_evaluation.py
+  aggregation: Weighted sliced W2 squared between complete-cell full-gene moment/Fourier
+    features; secondary to spatial W1.

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/manifest.json

@@ -0,0 +1,2220 @@
+{
+  "canonical_cache": "data/benchmark/intervention_ood/dss_colitis_merfish",
+  "cell_count_dtype": "uint16",
+  "cell_count_dtypes": [
+    "uint16"
+  ],
+  "cell_count_observed_max": 323,
+  "cell_filter_policy": {
+    "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+    "minimum_eligible_rna_per_cell": 1,
+    "zero_rna_cells_retained": false,
+    "population_support": {
+      "policy": "dataset_context_condition_min50_v1",
+      "minimum_cells": 50,
+      "grouping": [
+        "dataset",
+        "benchmark_cellular_context",
+        "condition"
+      ],
+      "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+      "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+      "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+      "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+    }
+  },
+  "cell_identity_view": {
+    "donor_patient_split": false,
+    "physical_array_index": "datasets/dss_colitis_merfish/population",
+    "schema": "cell_identity_view",
+    "selected_cells": 3751
+  },
+  "cell_key_dtype": "uint32",
+  "cell_key_semantics": "shard_local_lexical_cell_rank",
+  "cell_pointer_dtype": "uint64",
+  "context_column": "cell_type",
+  "coordinate_dimensions": [
+    "x",
+    "y"
+  ],
+  "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+  "dataset": "dss_colitis_merfish",
+  "elapsed_seconds": 122.89457134599797,
+  "format": "cell_major_rna_csr_numpy_parquet_v3",
+  "implementation": {
+    "builder": "scripts/build_population_index.py"
+  },
+  "inventory": {
+    "conditions": [
+      "D0",
+      "D21",
+      "D3",
+      "D35",
+      "D9"
+    ],
+    "contexts": [
+      "Adipose",
+      "Endothelial",
+      "EntericNervous",
+      "Epithelial",
+      "Fibroblast",
+      "ICC",
+      "Immune",
+      "Smooth Muscle Cells"
+    ],
+    "genes": [
+      "Ackr1",
+      "Ackr2",
+      "Ackr3",
+      "Ackr4",
+      "Acod1",
+      "Acvr1",
+      "Acvr1b",
+      "Acvr1c",
+      "Acvr2a",
+      "Acvr2b",
+      "Acvrl1",
+      "Ada",
+      "Adam12",
+      "Adam19",
+      "Adam23",
+      "Adamdec1",
+      "Adamts10",
+      "Adamts12",
+      "Adamts17",
+      "Adamts19",
+      "Adamts2",
+      "Adamts3",
+      "Adamts4",
+      "Adamts5",
+      "Adamts6",
+      "Adamts7",
+      "Adcyap1",
+      "Adgrd1",
+      "Adgre1",
+      "Adipoq",
+      "Adora1",
+      "Adora2a",
+      "Adrb2",
+      "Ahr",
+      "Aicda",
+      "Alcam",
+      "Alox5",
+      "Amh",
+      "Amhr2",
+      "Amotl2",
+      "Ankrd1",
+      "Aoah",
+      "Apod",
+      "Areg",
+      "Arg1",
+      "Asap1",
+      "Ascl2",
+      "Ash1l",
+      "Atf3",
+      "Atoh1",
+      "Axin2",
+      "Axl",
+      "Batf3",
+      "Bcl6",
+      "Bcr",
+      "Bicc1",
+      "Bmi1",
+      "Bmp1",
+      "Bmp10",
+      "Bmp15",
+      "Bmp2",
+      "Bmp3",
+      "Bmp4",
+      "Bmp5",
+      "Bmp6",
+      "Bmp7",
+      "Bmp8a",
+      "Bmp8b",
+      "Bmpr1a",
+      "Bmpr1b",
+      "Bmpr2",
+      "Btk",
+      "Btla",
+      "C1qtnf12",
+      "C1qtnf3",
+      "C3",
+      "C7",
+      "Cadm1",
+      "Cadm3",
+      "Calcrl",
+      "Casp12",
+      "Ccl1",
+      "Ccl11",
+      "Ccl12",
+      "Ccl17",
+      "Ccl19",
+      "Ccl2",
+      "Ccl20",
+      "Ccl21a",
+      "Ccl22",
+      "Ccl24",
+      "Ccl25",
+      "Ccl26",
+      "Ccl27a",
+      "Ccl28",
+      "Ccl3",
+      "Ccl4",
+      "Ccl5",
+      "Ccl6",
+      "Ccl7",
+      "Ccl8",
+      "Ccl9",
+      "Ccn1",
+      "Ccn2",
+      "Ccn3",
+      "Ccn4",
+      "Ccr1",
+      "Ccr10",
+      "Ccr1l1",
+      "Ccr2",
+      "Ccr3",
+      "Ccr4",
+      "Ccr5",
+      "Ccr6",
+      "Ccr7",
+      "Ccr8",
+      "Ccr9",
+      "Ccrl2",
+      "Cd14",
+      "Cd163",
+      "Cd19",
+      "Cd1d1",
+      "Cd2",
+      "Cd200",
+      "Cd200r1",
+      "Cd200r3",
+      "Cd22",
+      "Cd226",
+      "Cd248",
+      "Cd27",
+      "Cd274",
+      "Cd276",
+      "Cd28",
+      "Cd302",
+      "Cd34",
+      "Cd3d",
+      "Cd3e",
+      "Cd4",
+      "Cd40",
+      "Cd40lg",
+      "Cd44",
+      "Cd5",
+      "Cd5l",
+      "Cd6",
+      "Cd68",
+      "Cd69",
+      "Cd79a",
+      "Cd79b",
+      "Cd80",
+      "Cd83",
+      "Cd86",
+      "Cd8a",
+      "Cd8b1",
+      "Cd99l2",
+      "Cdh11",
+      "Cdh13",
+      "Cdh15",
+      "Cdh2",
+      "Cdh3",
+      "Cdh4",
+      "Cdh5",
+      "Cenpf",
+      "Cer1",
+      "Cercam",
+      "Chat",
+      "Chil3",
+      "Chrd",
+      "Chrna7",
+      "Clcf1",
+      "Cldn1",
+      "Cldn10",
+      "Cldn11",
+      "Cldn13",
+      "Cldn14",
+      "Cldn15",
+      "Cldn16",
+      "Cldn17",
+      "Cldn18",
+      "Cldn19",
+      "Cldn20",
+      "Cldn22",
+      "Cldn23",
+      "Cldn25",
+      "Cldn34-ps",
+      "Cldn34c4",
+      "Cldn34d",
+      "Cldn4",
+      "Cldn5",
+      "Cldn6",
+      "Cldn9",
+      "Clec10a",
+      "Clec9a",
+      "Clu",
+      "Cma1",
+      "Cnr2",
+      "Cntf",
+      "Cntfr",
+      "Cntn1",
+      "Cntn2",
+      "Cntnap1",
+      "Cntnap2",
+      "Col11a1",
+      "Col12a1",
+      "Col14a1",
+      "Col15a1",
+      "Col16a1",
+      "Col18a1",
+      "Col1a2",
+      "Col27a1",
+      "Col4a3",
+      "Col4a5",
+      "Col4a6",
+      "Col5a1",
+      "Col6a1",
+      "Col6a4",
+      "Col8a1",
+      "Cr2",
+      "Crabp1",
+      "Crim1",
+      "Crlf2",
+      "Csf1",
+      "Csf1r",
+      "Csf2",
+      "Csf2ra",
+      "Csf2rb",
+      "Csf2rb2",
+      "Csf3",
+      "Csf3r",
+      "Cspg4",
+      "Cst7",
+      "Ctf1",
+      "Ctf2",
+      "Ctla4",
+      "Cx3cl1",
+      "Cx3cr1",
+      "Cxcl1",
+      "Cxcl10",
+      "Cxcl11",
+      "Cxcl12",
+      "Cxcl13",
+      "Cxcl14",
+      "Cxcl15",
+      "Cxcl16",
+      "Cxcl17",
+      "Cxcl2",
+      "Cxcl3",
+      "Cxcl5",
+      "Cxcl9",
+      "Cxcr1",
+      "Cxcr2",
+      "Cxcr3",
+      "Cxcr4",
+      "Cxcr5",
+      "Cxcr6",
+      "Cyba",
+      "Cybb",
+      "Cygb",
+      "Cyp19a1",
+      "Ddah1",
+      "Ddit4",
+      "Ddx58",
+      "Dkk1",
+      "Dkk2",
+      "Dkk3",
+      "Dkk4",
+      "Dlc1",
+      "Dll4",
+      "Dpt",
+      "Dram1",
+      "Dusp1",
+      "Dusp10",
+      "Dusp2",
+      "Dusp4",
+      "Dusp5",
+      "Dysf",
+      "Ebi3",
+      "Eda2r",
+      "Edar",
+      "Edn1",
+      "Ednra",
+      "Ednrb",
+      "Egr1",
+      "Egr2",
+      "Egr3",
+      "Enpp3",
+      "Eomes",
+      "Esam",
+      "Ets1",
+      "Etv5",
+      "F3",
+      "Fap",
+      "Fas",
+      "Fasl",
+      "Fbln5",
+      "Fbn1",
+      "Fcer1a",
+      "Fcer2a",
+      "Fcgr1",
+      "Fcgr2b",
+      "Fcgr4",
+      "Fcna",
+      "Fem1a",
+      "Fem1al",
+      "Fer1l5",
+      "Fgf10",
+      "Fgf12",
+      "Fgf2",
+      "Fgf7",
+      "Fgf9",
+      "Fgfr1",
+      "Fgfr3",
+      "Fgfr4",
+      "Fjx1",
+      "Flnc",
+      "Flt4",
+      "Fmo2",
+      "Fmod",
+      "Fndc4",
+      "Fos",
+      "Fosb",
+      "Fosl1",
+      "Fosl2",
+      "Foxc2",
+      "Foxf1",
+      "Foxf2",
+      "Foxk1",
+      "Foxl1",
+      "Foxm1",
+      "Foxn2",
+      "Foxn3",
+      "Foxo1",
+      "Foxo3",
+      "Foxo4",
+      "Foxp1",
+      "Foxp2",
+      "Foxp3",
+      "Foxp4",
+      "Fpr1",
+      "Frzb",
+      "Fscn1",
+      "Fst",
+      "Fzd1",
+      "Fzd10",
+      "Fzd2",
+      "Fzd3",
+      "Fzd4",
+      "Fzd5",
+      "Fzd6",
+      "Fzd7",
+      "Fzd8",
+      "Fzd9",
+      "Gadd45a",
+      "Gadd45b",
+      "Gata3",
+      "Gdf10",
+      "Gdf11",
+      "Gdf15",
+      "Gdf2",
+      "Gdf3",
+      "Gdf5",
+      "Gdf6",
+      "Gdf7",
+      "Gdf9",
+      "Gfap",
+      "Gfpt2",
+      "Gfra3",
+      "Ghrl",
+      "Ghsr",
+      "Glg1",
+      "Gli1",
+      "Glycam1",
+      "Gna12",
+      "Gna13",
+      "Gna14",
+      "Gp2",
+      "Gper1",
+      "Gpr17",
+      "Gpr183",
+      "Gpr31b",
+      "Gpr65",
+      "Gps2",
+      "Gpx3",
+      "Grem1",
+      "Grem2",
+      "Gzma",
+      "Gzmb",
+      "Hand2",
+      "Havcr1",
+      "Havcr2",
+      "Hbegf",
+      "Hcar1",
+      "Hgf",
+      "Hif1a",
+      "Hmmr",
+      "Icam1",
+      "Icam2",
+      "Icos",
+      "Icosl",
+      "Ido1",
+      "Ier3",
+      "Ier5",
+      "Ifi35",
+      "Ifih1",
+      "Ifit1",
+      "Ifit2",
+      "Ifit3",
+      "Ifnar1",
+      "Ifnar2",
+      "Ifnb1",
+      "Ifng",
+      "Ifngr2",
+      "Ifnl2",
+      "Ifnl3",
+      "Ifnlr1",
+      "Igf1",
+      "Igf1r",
+      "Igf2",
+      "Igfbp1",
+      "Igfbp3",
+      "Igfbp5",
+      "Igfbp6",
+      "Igsf11",
+      "Ikzf2",
+      "Il10",
+      "Il10ra",
+      "Il10rb",
+      "Il11",
+      "Il11ra1",
+      "Il12a",
+      "Il12b",
+      "Il12rb1",
+      "Il12rb2",
+      "Il13",
+      "Il13ra1",
+      "Il13ra2",
+      "Il15",
+      "Il15ra",
+      "Il16",
+      "Il17a",
+      "Il17b",
+      "Il17c",
+      "Il17d",
+      "Il17f",
+      "Il17ra",
+      "Il17rb",
+      "Il17rc",
+      "Il17rd",
+      "Il17re",
+      "Il18",
+      "Il18r1",
+      "Il18rap",
+      "Il19",
+      "Il1a",
+      "Il1b",
+      "Il1f5",
+      "Il1f6",
+      "Il1f8",
+      "Il1f9",
+      "Il1r1",
+      "Il1r2",
+      "Il1rap",
+      "Il1rl1",
+      "Il1rl2",
+      "Il1rn",
+      "Il2",
+      "Il20",
+      "Il20ra",
+      "Il20rb",
+      "Il21",
+      "Il21r",
+      "Il22",
+      "Il22ra1",
+      "Il22ra2",
+      "Il23a",
+      "Il23r",
+      "Il24",
+      "Il25",
+      "Il27",
+      "Il27ra",
+      "Il2ra",
+      "Il2rb",
+      "Il2rg",
+      "Il3",
+      "Il31",
+      "Il31ra",
+      "Il33",
+      "Il34",
+      "Il3ra",
+      "Il4",
+      "Il4ra",
+      "Il5",
+      "Il5ra",
+      "Il6",
+      "Il6ra",
+      "Il6st",
+      "Il7",
+      "Il7r",
+      "Il9",
+      "Il9r",
+      "Iltifb",
+      "Inha",
+      "Inhba",
+      "Inhbb",
+      "Inhbc",
+      "Inhbe",
+      "Irf3",
+      "Irf4",
+      "Irf5",
+      "Irf8",
+      "Irgm1",
+      "Irs1",
+      "Irs2",
+      "Irs3",
+      "Isg15",
+      "Isg20",
+      "Isl1",
+      "Itga1",
+      "Itga10",
+      "Itga11",
+      "Itga2",
+      "Itga2b",
+      "Itga4",
+      "Itga5",
+      "Itga7",
+      "Itga8",
+      "Itga9",
+      "Itgae",
+      "Itgal",
+      "Itgam",
+      "Itgav",
+      "Itgax",
+      "Itgb2",
+      "Itgb2l",
+      "Itgb3",
+      "Itgb5",
+      "Itgb6",
+      "Itgb7",
+      "Itgb8",
+      "Itln1",
+      "Jag1",
+      "Jag2",
+      "Jam2",
+      "Jam3",
+      "Kdm6b",
+      "Kdr",
+      "Kit",
+      "Klk1b4",
+      "Klrb1c",
+      "Klrc2",
+      "Klrk1",
+      "Krt1",
+      "Kynu",
+      "L1cam",
+      "Lag3",
+      "Lbh",
+      "Lbp",
+      "Lefty1",
+      "Lefty2",
+      "Lepr",
+      "Lgals2",
+      "Lgr4",
+      "Lgr5",
+      "Lgr6",
+      "Lif",
+      "Lifr",
+      "Lmx1a",
+      "Lpar1",
+      "Lrfn5",
+      "Lrp5",
+      "Lrp6",
+      "Lrrc32",
+      "Lrrc4",
+      "Lrrc4b",
+      "Lrrc4c",
+      "Lta",
+      "Ltb",
+      "Ltbp1",
+      "Ltbr",
+      "Lum",
+      "Lyve1",
+      "Macir",
+      "Madcam1",
+      "Maf",
+      "Mafa",
+      "Mafb",
+      "Maff",
+      "Mafk",
+      "Mag",
+      "Map3k2",
+      "Map3k8",
+      "Mcm2",
+      "Mdk",
+      "Mefv",
+      "Mfap4",
+      "Mfhas1",
+      "Mki67",
+      "Mmp10",
+      "Mmp11",
+      "Mmp12",
+      "Mmp13",
+      "Mmp14",
+      "Mmp15",
+      "Mmp16",
+      "Mmp17",
+      "Mmp19",
+      "Mmp1a",
+      "Mmp1b",
+      "Mmp2",
+      "Mmp20",
+      "Mmp21",
+      "Mmp23",
+      "Mmp24",
+      "Mmp25",
+      "Mmp27",
+      "Mmp28",
+      "Mmp3",
+      "Mmp7",
+      "Mmp8",
+      "Mmp9",
+      "Mpo",
+      "Mpz",
+      "Mpzl1",
+      "Mrc1",
+      "Ms4a1",
+      "Msc",
+      "Mst1",
+      "Mstn",
+      "Muc1",
+      "Muc4",
+      "Mvd",
+      "Mvk",
+      "Mx1",
+      "Mx2",
+      "Myc",
+      "Myd88",
+      "Nampt",
+      "Ncam1",
+      "Ncam2",
+      "Ncr1",
+      "Nectin1",
+      "Negr1",
+      "Neo1",
+      "Neurod1",
+      "Neurog3",
+      "Nf2",
+      "Nfasc",
+      "Nfat5",
+      "Nfil3",
+      "Nfkbid",
+      "Nfkbie",
+      "Ngf",
+      "Ngfr",
+      "Ninj1",
+      "Ninj2",
+      "Nlgn1",
+      "Nlgn2",
+      "Nlgn3",
+      "Nlrc3",
+      "Nlrc4",
+      "Nlrc5",
+      "Nlrp10",
+      "Nlrp12",
+      "Nlrp1a",
+      "Nlrp1b",
+      "Nlrp3",
+      "Nlrp4a",
+      "Nlrp5",
+      "Nlrp9a",
+      "Nlrp9b",
+      "Nlrp9c",
+      "Nlrx1",
+      "Nod2",
+      "Nodal",
+      "Nog",
+      "Nos1",
+      "Nos2",
+      "Nos3",
+      "Notch1",
+      "Notch2",
+      "Notch3",
+      "Notch4",
+      "Notum",
+      "Nox1",
+      "Nox3",
+      "Nox4",
+      "Npy",
+      "Npy5r",
+      "Nr1d2",
+      "Nr1h3",
+      "Nr4a1",
+      "Nr4a2",
+      "Nr4a3",
+      "Nrcam",
+      "Nrxn1",
+      "Nrxn2",
+      "Nrxn3",
+      "Nt5e",
+      "Ntng1",
+      "Ntng2",
+      "Nudt1",
+      "Oasl1",
+      "Olr1",
+      "Osm",
+      "Osmr",
+      "Otulin",
+      "P2ry6",
+      "Pax5",
+      "Pbk",
+      "Pcna",
+      "Pdcd1",
+      "Pdcd1lg2",
+      "Pde4b",
+      "Pdgfra",
+      "Pdgfrb",
+      "Pdpn",
+      "Pecam1",
+      "Per1",
+      "Pf4",
+      "Pfkfb3",
+      "Phgdh",
+      "Pi16",
+      "Pla2g5",
+      "Plau",
+      "Plaur",
+      "Plcg2",
+      "Plek",
+      "Plk2",
+      "Pou2af1",
+      "Pou2f3",
+      "Pou6f1",
+      "Ppara",
+      "Ppbp",
+      "Ppp1r15a",
+      "Prdm1",
+      "Prf1",
+      "Prkcb",
+      "Prkce",
+      "Proc",
+      "Prox1",
+      "Ptger1",
+      "Ptger2",
+      "Ptger3",
+      "Ptger4",
+      "Ptgis",
+      "Ptgr1",
+      "Ptgs1",
+      "Ptgs2",
+      "Ptpn2",
+      "Ptprc",
+      "Ptpre",
+      "Ptprm",
+      "Ptx3",
+      "Pvr",
+      "Rac2",
+      "Rb1",
+      "Rbp1",
+      "Rbp4",
+      "Ripk2",
+      "Rnf43",
+      "Rora",
+      "Rorc",
+      "Rspo1",
+      "Rspo2",
+      "Rspo3",
+      "S1pr2",
+      "S1pr3",
+      "Sav1",
+      "Sbno2",
+      "Schip1",
+      "Sdc1",
+      "Sdc2",
+      "Sdc3",
+      "Sele",
+      "Sell",
+      "Selp",
+      "Selplg",
+      "Sema4a",
+      "Serpina3g",
+      "Serpinb8",
+      "Serpine1",
+      "Serpinf1",
+      "Sfrp1",
+      "Sfrp2",
+      "Shcbp1",
+      "Siglec1",
+      "Siglece",
+      "Siglecf",
+      "Siglecg",
+      "Sik1",
+      "Sirpa",
+      "Slamf8",
+      "Slc12a3",
+      "Slc2a3",
+      "Slc2a6",
+      "Slit2",
+      "Smgc",
+      "Smoc2",
+      "Smpdl3b",
+      "Snn",
+      "Socs1",
+      "Socs3",
+      "Socs5",
+      "Sox4",
+      "Sox6",
+      "Sparc",
+      "Spdef",
+      "Sphk1",
+      "Spi1",
+      "Spn",
+      "Spp1",
+      "Spry1",
+      "Stat2",
+      "Stat4",
+      "Stat5a",
+      "Stat5b",
+      "Sting1",
+      "Stk3",
+      "Sucnr1",
+      "Syk",
+      "Syt11",
+      "Tafa3",
+      "Tagln",
+      "Taz",
+      "Tbx21",
+      "Tcf21",
+      "Tcf7",
+      "Tead1",
+      "Tead2",
+      "Tead3",
+      "Tead4",
+      "Tff2",
+      "Tgfa",
+      "Tgfb1",
+      "Tgfb2",
+      "Tgfb3",
+      "Tgfbr1",
+      "Tgfbr2",
+      "Tgfbr3",
+      "Thbs1",
+      "Thy1",
+      "Ticam1",
+      "Tigit",
+      "Timp1",
+      "Timp2",
+      "Timp3",
+      "Timp4",
+      "Tlr1",
+      "Tlr2",
+      "Tlr3",
+      "Tlr4",
+      "Tlr5",
+      "Tlr6",
+      "Tlr7",
+      "Tlr8",
+      "Tlr9",
+      "Tnc",
+      "Tnf",
+      "Tnfaip2",
+      "Tnfaip3",
+      "Tnfaip6",
+      "Tnfaip8l2",
+      "Tnfrsf10b",
+      "Tnfrsf11a",
+      "Tnfrsf11b",
+      "Tnfrsf12a",
+      "Tnfrsf13b",
+      "Tnfrsf13c",
+      "Tnfrsf14",
+      "Tnfrsf17",
+      "Tnfrsf18",
+      "Tnfrsf1a",
+      "Tnfrsf1b",
+      "Tnfrsf21",
+      "Tnfrsf25",
+      "Tnfrsf4",
+      "Tnfrsf8",
+      "Tnfrsf9",
+      "Tnfsf10",
+      "Tnfsf11",
+      "Tnfsf12",
+      "Tnfsf13",
+      "Tnfsf13b",
+      "Tnfsf14",
+      "Tnfsf15",
+      "Tnfsf18",
+      "Tnfsf4",
+      "Tnfsf8",
+      "Tnfsf9",
+      "Tnip2",
+      "Tnnt2",
+      "Traf1",
+      "Traf2",
+      "Traf3",
+      "Traf5",
+      "Traf6",
+      "Trdc",
+      "Trem1",
+      "Trem2",
+      "Trex1",
+      "Trpa1",
+      "Trpv1",
+      "Tslp",
+      "Tubb2a",
+      "Tubb3",
+      "Tyro3",
+      "Vcam1",
+      "Vcan",
+      "Vcl",
+      "Vegfa",
+      "Vil1",
+      "Vsir",
+      "Vtcn1",
+      "Wfdc1",
+      "Wif1",
+      "Wnt1",
+      "Wnt10a",
+      "Wnt10b",
+      "Wnt11",
+      "Wnt16",
+      "Wnt2",
+      "Wnt2b",
+      "Wnt3",
+      "Wnt3a",
+      "Wnt4",
+      "Wnt5a",
+      "Wnt5b",
+      "Wnt6",
+      "Wnt7a",
+      "Wnt7b",
+      "Wnt8a",
+      "Wnt8b",
+      "Wnt9a",
+      "Wnt9b",
+      "Xcl1",
+      "Xcr1",
+      "Yap1",
+      "Zbtb10",
+      "Zbtb46",
+      "Zc3h12a",
+      "Znrf3"
+    ],
+    "units": [
+      "062221_D9_m3",
+      "062921_D0_m3a",
+      "062921_D9_m2a",
+      "062921_D9_m5",
+      "072523_D35_m11",
+      "072523_D35_m6",
+      "082421_D0_m6",
+      "082421_D0_m7",
+      "082421_D21_m1",
+      "082421_D21_m2",
+      "092421_D3_m1",
+      "092421_D3_m2",
+      "092421_D3_m3",
+      "092421_D3_m4",
+      "100221_D9_m2",
+      "100221_D9_m3",
+      "100221_D9_m5"
+    ]
+  },
+  "loaded_cell_key_dtype": "uint64",
+  "loaded_cell_key_semantics": "unit_index_uint32_shifted_plus_lexical_cell_rank",
+  "point_dtype": "float32",
+  "schema_version": 3,
+  "shards": [
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 126,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0000-062221_D9_m3/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0000-062221_D9_m3/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0000-062221_D9_m3/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0000-062221_D9_m3/cell_rna_ptr.npy",
+      "cells": "shards/0000-062221_D9_m3/cells.parquet",
+      "conditions": [
+        "D9"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 82479,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 5430714,
+      "nonempty_cell_gene_rows": 4063364,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0000-062221_D9_m3/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0000-062221_D9_m3/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 69,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062221_D9_m3_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062221_D9_m3_2_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062221_D9_m3_2_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062221_D9_m3_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062221_D9_m3_2_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062221_D9_m3_2_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "062221_D9_m3",
+      "unit_index": 0
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 323,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0001-062921_D0_m3a/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0001-062921_D0_m3a/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0001-062921_D0_m3a/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0001-062921_D0_m3a/cell_rna_ptr.npy",
+      "cells": "shards/0001-062921_D0_m3a/cells.parquet",
+      "conditions": [
+        "D0"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 133081,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 15952163,
+      "nonempty_cell_gene_rows": 9067663,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0001-062921_D0_m3a/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0001-062921_D0_m3a/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 361,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D0_m3a_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D0_m3a_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D0_m3a_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D0_m3a_2_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D0_m3a_2_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D0_m3a_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D0_m3a_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D0_m3a_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D0_m3a_2_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D0_m3a_2_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "062921_D0_m3a",
+      "unit_index": 1
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 139,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0002-062921_D9_m2a/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0002-062921_D9_m2a/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0002-062921_D9_m2a/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0002-062921_D9_m2a/cell_rna_ptr.npy",
+      "cells": "shards/0002-062921_D9_m2a/cells.parquet",
+      "conditions": [
+        "D9"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 74125,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 6804895,
+      "nonempty_cell_gene_rows": 4435903,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0002-062921_D9_m2a/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0002-062921_D9_m2a/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 243,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D9_m2a_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D9_m2a_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D9_m2a_2_slice_2.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D9_m2a_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D9_m2a_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D9_m2a_2_slice_2.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "062921_D9_m2a",
+      "unit_index": 2
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 148,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0003-062921_D9_m5/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0003-062921_D9_m5/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0003-062921_D9_m5/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0003-062921_D9_m5/cell_rna_ptr.npy",
+      "cells": "shards/0003-062921_D9_m5/cells.parquet",
+      "conditions": [
+        "D9"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 182804,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 12878989,
+      "nonempty_cell_gene_rows": 9416552,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0003-062921_D9_m5/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0003-062921_D9_m5/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 61,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D9_m5_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D9_m5_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D9_m5_1_slice_3.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D9_m5_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D9_m5_2_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/062921_D9_m5_2_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D9_m5_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D9_m5_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D9_m5_1_slice_3.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D9_m5_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D9_m5_2_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/062921_D9_m5_2_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "062921_D9_m5",
+      "unit_index": 3
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 171,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0004-072523_D35_m11/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0004-072523_D35_m11/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0004-072523_D35_m11/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0004-072523_D35_m11/cell_rna_ptr.npy",
+      "cells": "shards/0004-072523_D35_m11/cells.parquet",
+      "conditions": [
+        "D35"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 42870,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 3301544,
+      "nonempty_cell_gene_rows": 2347163,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0004-072523_D35_m11/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0004-072523_D35_m11/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 520,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/072523_D35_m11_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/072523_D35_m11_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/072523_D35_m11_1_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/072523_D35_m11_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/072523_D35_m11_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/072523_D35_m11_1_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "072523_D35_m11",
+      "unit_index": 4
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 111,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0005-072523_D35_m6/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0005-072523_D35_m6/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0005-072523_D35_m6/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0005-072523_D35_m6/cell_rna_ptr.npy",
+      "cells": "shards/0005-072523_D35_m6/cells.parquet",
+      "conditions": [
+        "D35"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 48090,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 3019291,
+      "nonempty_cell_gene_rows": 2333791,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0005-072523_D35_m6/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0005-072523_D35_m6/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 122,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/072523_D35_m6_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/072523_D35_m6_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/072523_D35_m6_1_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/072523_D35_m6_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/072523_D35_m6_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/072523_D35_m6_1_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "072523_D35_m6",
+      "unit_index": 5
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 227,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0006-082421_D0_m6/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0006-082421_D0_m6/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0006-082421_D0_m6/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0006-082421_D0_m6/cell_rna_ptr.npy",
+      "cells": "shards/0006-082421_D0_m6/cells.parquet",
+      "conditions": [
+        "D0"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 70000,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 6824033,
+      "nonempty_cell_gene_rows": 4305026,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0006-082421_D0_m6/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0006-082421_D0_m6/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 268,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/082421_D0_m6_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/082421_D0_m6_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/082421_D0_m6_1_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/082421_D0_m6_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/082421_D0_m6_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/082421_D0_m6_1_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "082421_D0_m6",
+      "unit_index": 6
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 190,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0007-082421_D0_m7/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0007-082421_D0_m7/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0007-082421_D0_m7/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0007-082421_D0_m7/cell_rna_ptr.npy",
+      "cells": "shards/0007-082421_D0_m7/cells.parquet",
+      "conditions": [
+        "D0"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 36369,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 4666025,
+      "nonempty_cell_gene_rows": 2795305,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0007-082421_D0_m7/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0007-082421_D0_m7/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 150,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/082421_D0_m7_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/082421_D0_m7_1_slice_2.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/082421_D0_m7_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/082421_D0_m7_1_slice_2.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "082421_D0_m7",
+      "unit_index": 7
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 235,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0008-082421_D21_m1/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0008-082421_D21_m1/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0008-082421_D21_m1/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0008-082421_D21_m1/cell_rna_ptr.npy",
+      "cells": "shards/0008-082421_D21_m1/cells.parquet",
+      "conditions": [
+        "D21"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 88089,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 9172622,
+      "nonempty_cell_gene_rows": 5714746,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0008-082421_D21_m1/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0008-082421_D21_m1/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 103,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/082421_D21_m1_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/082421_D21_m1_1_slice_2.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/082421_D21_m1_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/082421_D21_m1_1_slice_2.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "082421_D21_m1",
+      "unit_index": 8
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 143,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0009-082421_D21_m2/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0009-082421_D21_m2/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0009-082421_D21_m2/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0009-082421_D21_m2/cell_rna_ptr.npy",
+      "cells": "shards/0009-082421_D21_m2/cells.parquet",
+      "conditions": [
+        "D21"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 60792,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 3328322,
+      "nonempty_cell_gene_rows": 2598079,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0009-082421_D21_m2/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0009-082421_D21_m2/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 31,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/082421_D21_m2_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/082421_D21_m2_1_slice_2.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/082421_D21_m2_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/082421_D21_m2_1_slice_2.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "082421_D21_m2",
+      "unit_index": 9
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 187,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0010-092421_D3_m1/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0010-092421_D3_m1/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0010-092421_D3_m1/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0010-092421_D3_m1/cell_rna_ptr.npy",
+      "cells": "shards/0010-092421_D3_m1/cells.parquet",
+      "conditions": [
+        "D3"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 83828,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 6410107,
+      "nonempty_cell_gene_rows": 4239283,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0010-092421_D3_m1/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0010-092421_D3_m1/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 268,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m1_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m1_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m1_1_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m1_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m1_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m1_1_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "092421_D3_m1",
+      "unit_index": 10
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 167,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0011-092421_D3_m2/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0011-092421_D3_m2/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0011-092421_D3_m2/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0011-092421_D3_m2/cell_rna_ptr.npy",
+      "cells": "shards/0011-092421_D3_m2/cells.parquet",
+      "conditions": [
+        "D3"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 84850,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 6268312,
+      "nonempty_cell_gene_rows": 4339466,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0011-092421_D3_m2/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0011-092421_D3_m2/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 290,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m2_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m2_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m2_1_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m2_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m2_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m2_1_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "092421_D3_m2",
+      "unit_index": 11
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 221,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0012-092421_D3_m3/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0012-092421_D3_m3/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0012-092421_D3_m3/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0012-092421_D3_m3/cell_rna_ptr.npy",
+      "cells": "shards/0012-092421_D3_m3/cells.parquet",
+      "conditions": [
+        "D3"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 70949,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 6903494,
+      "nonempty_cell_gene_rows": 4205455,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0012-092421_D3_m3/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0012-092421_D3_m3/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 534,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m3_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m3_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m3_1_slice_3.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m3_1_slice_4.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m3_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m3_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m3_1_slice_3.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m3_1_slice_4.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "092421_D3_m3",
+      "unit_index": 12
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 303,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0013-092421_D3_m4/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0013-092421_D3_m4/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0013-092421_D3_m4/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0013-092421_D3_m4/cell_rna_ptr.npy",
+      "cells": "shards/0013-092421_D3_m4/cells.parquet",
+      "conditions": [
+        "D3"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 59647,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 7497574,
+      "nonempty_cell_gene_rows": 4384975,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0013-092421_D3_m4/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0013-092421_D3_m4/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 344,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m4_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m4_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/092421_D3_m4_1_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m4_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m4_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/092421_D3_m4_1_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "092421_D3_m4",
+      "unit_index": 13
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 190,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0014-100221_D9_m2/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0014-100221_D9_m2/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0014-100221_D9_m2/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0014-100221_D9_m2/cell_rna_ptr.npy",
+      "cells": "shards/0014-100221_D9_m2/cells.parquet",
+      "conditions": [
+        "D9"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 59103,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 5581016,
+      "nonempty_cell_gene_rows": 3527816,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0014-100221_D9_m2/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0014-100221_D9_m2/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 188,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m2_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m2_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m2_1_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m2_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m2_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m2_1_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "100221_D9_m2",
+      "unit_index": 14
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 159,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0015-100221_D9_m3/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0015-100221_D9_m3/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0015-100221_D9_m3/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0015-100221_D9_m3/cell_rna_ptr.npy",
+      "cells": "shards/0015-100221_D9_m3/cells.parquet",
+      "conditions": [
+        "D9"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 89335,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 5605952,
+      "nonempty_cell_gene_rows": 4341454,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0015-100221_D9_m3/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0015-100221_D9_m3/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 52,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m3_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m3_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m3_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m3_2_slice_2.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m3_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m3_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m3_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m3_2_slice_2.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "100221_D9_m3",
+      "unit_index": 15
+    },
+    {
+      "cell_count_dtype": "uint16",
+      "cell_count_observed_max": 167,
+      "cell_counts": "../../../../../../datasets/dss_colitis_merfish/population/shards/0016-100221_D9_m5/counts.npy",
+      "cell_filter_policy": {
+        "cell_filter_stage": "after_eligible_RNA_filtering_before_population_indexing",
+        "minimum_eligible_rna_per_cell": 1,
+        "zero_rna_cells_retained": false,
+        "population_support": {
+          "policy": "dataset_context_condition_min50_v1",
+          "minimum_cells": 50,
+          "grouping": [
+            "dataset",
+            "benchmark_cellular_context",
+            "condition"
+          ],
+          "aggregation": "Across eligible units in the same dataset; anatomy encoded in condition labels is preserved.",
+          "stage": "after_nonempty_RNA_reconstruction_before_release_and_split_access",
+          "response_rule": "Prediction responses additionally require >=50 matched source and >=50 target cells.",
+          "context_note": "UC-VDZ uses the native cellular-context annotations already used in frozen task views; patient identity is retained separately. Cells without a recoverable cellular-context annotation are excluded."
+        }
+      },
+      "cell_gene_dtype": "uint16",
+      "cell_genes": "../../../../../../datasets/dss_colitis_merfish/population/shards/0016-100221_D9_m5/gene_ids.npy",
+      "cell_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0016-100221_D9_m5/cell_gene_ptr.npy",
+      "cell_key_dtype": "uint32",
+      "cell_key_semantics": "shard_local_lexical_cell_rank",
+      "cell_pointer_dtype": "uint64",
+      "cell_rna_indptr": "../../../../../../datasets/dss_colitis_merfish/population/shards/0016-100221_D9_m5/cell_rna_ptr.npy",
+      "cells": "shards/0016-100221_D9_m5/cells.parquet",
+      "conditions": [
+        "D9"
+      ],
+      "context_column": "cell_type",
+      "contexts": [
+        "Adipose",
+        "Endothelial",
+        "EntericNervous",
+        "Epithelial",
+        "Fibroblast",
+        "ICC",
+        "Immune",
+        "Smooth Muscle Cells"
+      ],
+      "coordinate_order": "cell_key_then_gene_id_then_original_RNA_order",
+      "declared_cells": 173131,
+      "format": "cell_major_rna_csr_numpy_parquet_v3",
+      "molecules": 9908312,
+      "nonempty_cell_gene_rows": 7266786,
+      "points": "../../../../../../datasets/dss_colitis_merfish/population/shards/0016-100221_D9_m5/coordinates.npy",
+      "population_rows": 1,
+      "populations": "shards/0016-100221_D9_m5/groups.parquet",
+      "schema_version": 3,
+      "selected_cells": 147,
+      "source_cells": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m5_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m5_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m5_1_slice_3.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m5_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m5_2_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/cells/100221_D9_m5_2_slice_3.parquet"
+      ],
+      "source_molecules": [
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m5_1_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m5_1_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m5_1_slice_3.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m5_2_slice_1.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m5_2_slice_2.parquet",
+        "data/benchmark/intervention_ood/dss_colitis_merfish/molecules/100221_D9_m5_2_slice_3.parquet"
+      ],
+      "status": "complete",
+      "unit_id": "100221_D9_m5",
+      "unit_index": 16
+    }
+  ],
+  "status": "complete",
+  "storage_migration": {
+    "every_RNA_round_trip_equal": true,
+    "method": "lossless_cell_major_permutation"
+  },
+  "totals": {
+    "declared_cells": 1439542,
+    "molecules": 119553365,
+    "nonempty_cell_gene_rows": 79382827,
+    "population_rows": 133,
+    "units": 17
+  },
+  "zero_rna_cell_filter": {
+    "RNA_arrays_verified_unchanged": true,
+    "removed_cells": 0
+  },
+  "release_filter_provenance": {
+    "frozen_membership_unchanged": true,
+    "physical_keys_remapped": true
+  }
+}

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0000-062221_D9_m3/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f13640492d143bc770cb60761e2130b1f02367fd7e088454721a11e7a817fec
+size 7395

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0000-062221_D9_m3/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff4128f42c2062be902faa6fdd8376572779ccfd94cd918c193a67dc8aadbb16
+size 1916

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0001-062921_D0_m3a/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7db18d9f05199804f425b20cce9b7d74612e87a785fbe6e19290d48e94905e29
+size 19352

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0001-062921_D0_m3a/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fdb97c5d114a5d14cc29cb5e36cc17bb5d0c009a2f97f064d328097c4cfcfdee
+size 1925

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0002-062921_D9_m2a/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:07f9d6f5f5e2f4a410d3de10c99300680a4adada91ceb289be2bdf7ba813d3d0
+size 14548

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0002-062921_D9_m2a/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3a222f588b746d4f12bfdcd4c70b40f440c1451e7f8393f191287e1795ad016b
+size 1925

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0003-062921_D9_m5/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c61d94892879c482f88ea5ec66acacc6bd1b248b51fccf49f91b862f85ac718b
+size 6948

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0003-062921_D9_m5/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:436e4ad169871067b440698ab404ae2c8301d3621d849a1cf7042efbb7579067
+size 1916

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0004-072523_D35_m11/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ca0b3b89167bf0d9686c4a0f22bcd3c7378f5b8ffa2755748ea9fa31666911b
+size 24679

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0004-072523_D35_m11/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec5e09be5c0d2f2bd932640e63bb46480c40fcad9583caed35c6eb8c780aa910
+size 1943

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0005-072523_D35_m6/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b924350ca0221fdf51d6a5e7bbdbc3d8986517576d85a2e8ebc6e53e8e8dbe15
+size 9623

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0005-072523_D35_m6/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61512de4221e8234b23dc54d28cd80ba332093c451d5e389c101853da1ee4ead
+size 1934

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0006-082421_D0_m6/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17142e159a7e1df8c26b5595d403a51fb852bd650260ca779f6aa1b5409fee33
+size 15236

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0006-082421_D0_m6/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c192988f0ce15597b38f6cccfd7e950b3476573f804816ea0beb98db2e23b22b
+size 1916

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0007-082421_D0_m7/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e5781b4f4e951caaa8f1b931135aeafd7391055ffce2ee3c00f6835ed1a4515
+size 10717

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0007-082421_D0_m7/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40081169ac4fc155df0b423437dc4505cba6e93ca4b0da087478f98fc973a9f6
+size 1916

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0008-082421_D21_m1/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:692826629ec7294339abacdbe0cf7405b8b8ce69d311379bbe18fcff8d409603
+size 8912

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0008-082421_D21_m1/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d34583bd3f61fb131eebff2a71929f4bfc7ecb4c7b8174d395f68d4da213dafa
+size 1934

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0009-082421_D21_m2/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfbb5ce47ca15ac64d5899dc185cc57f3b80849359c041c64dbf489db5140fdc
+size 5587

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0009-082421_D21_m2/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a10d141c505bd28717a94aaa2d0222b5cdd3c370a86e66a668a0b91815fb861
+size 1934

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0010-092421_D3_m1/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86784cd5f818dea2188319c9b7f71ddb26f4fe877fd5e12ab32cc1daea8b4584
+size 15204

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0010-092421_D3_m1/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df931b7157b5b3a827394ebd4c5c04ea95e8812e5cdc68f45ff23f67772f8069
+size 1916

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0011-092421_D3_m2/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d636803480142b4ce8dfef3145826bef8bd35a08a015b14f2455d4df6567485
+size 16081

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0011-092421_D3_m2/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34f4186d0ba0f16d0fe10a3d5380734620f797bb3e6107ddeca91745c5ea8aff
+size 1916

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0012-092421_D3_m3/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30f7ca3ba3cf2639c4c0baeb33cbf6ad8556b88fcf101db91675fc436b9f41e4
+size 25929

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0012-092421_D3_m3/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:929ec4d83a26f87b569f02f5cad5c814caa86b6ceabb7e53a24cc503980c085b
+size 1916

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0013-092421_D3_m4/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5e837b582f9d966ea1588de8fb7d606fcba8dc21e119dd5632195ad041ff6617
+size 18953

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0013-092421_D3_m4/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1205a2b5d3e2e40ff451f2b17561ddfe65ffb88225abe4f45bdb9b0d43e0b58
+size 1916

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0014-100221_D9_m2/cells.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:928b90d96f0de83746442ee86a0453451655904d4a7066a0f017ca129a2c150f
+size 12284

benchmark/split_views/dss_colitis_merfish/seed1/same_context_dual_min50_v1/test/shards/0014-100221_D9_m2/groups.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60dc968d0a597d8dd3f03de3d2648f2d77f1db5602222fff046184bbec334072
+size 1916