diff --git a/applications/dynacell/src/dynacell/evaluation/cross_condition_probe.py b/applications/dynacell/src/dynacell/evaluation/cross_condition_probe.py
index 0b554931b..2ea5d4e3d 100644
--- a/applications/dynacell/src/dynacell/evaluation/cross_condition_probe.py
+++ b/applications/dynacell/src/dynacell/evaluation/cross_condition_probe.py
@@ -40,6 +40,20 @@
 _SOURCES = ("pred", "gt")
 _CONDITION_TOKENS = ("mock", "denv", "zikv")
 _DEFAULT_PAIRS = (("mock", "denv"), ("mock", "zikv"))
+_FIELDNAMES = (
+    "feature_type",
+    "pair",
+    "source",
+    "n_cells_c0",
+    "n_cells_c1",
+    "n_fovs",
+    "n_folds",
+    "auroc_mean",
+    "auroc_std",
+    "skipped_reason",
+)
+#: Filename written into each infected condition's eval dir by :func:`run_for_group`.
+GROUP_PROBE_FILENAME = "cross_condition_probe.csv"
 
 
 def _detect_condition(eval_dir: Path) -> str:
@@ -135,6 +149,68 @@ def _probe_pair(
     return row
 
 
+def _write_rows(out_path: Path, rows: list[dict]) -> None:
+    """Write probe rows as a CSV with the canonical field order."""
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    with out_path.open("w", newline="") as f:
+        writer = csv.DictWriter(f, fieldnames=_FIELDNAMES)
+        writer.writeheader()
+        writer.writerows(rows)
+
+
+def run_for_group(
+    eval_dirs: list[Path],
+    n_splits: int = 5,
+    rng_seed: int = 2020,
+) -> list[Path]:
+    """Probe each infected condition against mock and write a per-condition CSV.
+
+    Unlike :func:`run` (long-form CSV over all pairs at one ``out_path``),
+    this writes one :data:`GROUP_PROBE_FILENAME` into *each infected
+    condition's* eval dir, holding only that condition's ``mock_vs_<cond>``
+    rows (every feature × {pred, gt}). This colocates the probe with the eval
+    dir the reporting layer already resolves per (model, pool, organelle,
+    condition), so the table generator can read it without knowing about
+    sibling conditions.
+
+    Requires a ``mock`` reference dir plus at least one infected dir; returns
+    the list of CSV paths written (empty when the group has no mock or no
+    infected condition, e.g. the in-distribution iPSC eval).
+
+    Parameters
+    ----------
+    eval_dirs : list[Path]
+        Per-condition eval dirs of one (model, pool, organelle) group. The
+        condition is inferred from each dir's trailing ``_{mock,denv,zikv}``;
+        dirs without a recognized token are ignored.
+    n_splits, rng_seed : int
+        Forwarded to :func:`fov_stratified_auroc`.
+    """
+    by_condition: dict[str, Path] = {}
+    for d in eval_dirs:
+        try:
+            cond = _detect_condition(d)
+        except ValueError:
+            continue
+        by_condition[cond] = d
+    if "mock" not in by_condition:
+        return []
+
+    written: list[Path] = []
+    for ref, cond in _DEFAULT_PAIRS:  # ref == "mock" for every default pair
+        if cond not in by_condition:
+            continue
+        rows = [
+            _probe_pair(by_condition, (ref, cond), feature, source, n_splits, rng_seed)
+            for feature in _FEATURE_TYPES
+            for source in _SOURCES
+        ]
+        out_path = by_condition[cond] / GROUP_PROBE_FILENAME
+        _write_rows(out_path, rows)
+        written.append(out_path)
+    return written
+
+
 def run(
     eval_dirs: list[Path],
     out_path: Path,
@@ -169,23 +245,7 @@ def run(
             for source in _SOURCES:
                 rows.append(_probe_pair(eval_dirs_by_condition, pair, feature, source, n_splits, rng_seed))
 
-    out_path.parent.mkdir(parents=True, exist_ok=True)
-    fieldnames = [
-        "feature_type",
-        "pair",
-        "source",
-        "n_cells_c0",
-        "n_cells_c1",
-        "n_fovs",
-        "n_folds",
-        "auroc_mean",
-        "auroc_std",
-        "skipped_reason",
-    ]
-    with out_path.open("w", newline="") as f:
-        writer = csv.DictWriter(f, fieldnames=fieldnames)
-        writer.writeheader()
-        writer.writerows(rows)
+    _write_rows(out_path, rows)
     return out_path
 
 
diff --git a/applications/dynacell/tests/test_benchmark_config_composition.py b/applications/dynacell/tests/test_benchmark_config_composition.py
index d1ebaedc7..8a4252e90 100644
--- a/applications/dynacell/tests/test_benchmark_config_composition.py
+++ b/applications/dynacell/tests/test_benchmark_config_composition.py
@@ -327,8 +327,8 @@ def test_a549_predict_leaf_composes(
     - dataset_ref.{dataset,target} carry through composition (proves the
       gene-keyed predict_set + targets overlays took effect).
     - experiment_id reflects the cross-eval pairing (per condition).
-    - sbatch.constraint inherits "h200" from hardware_h200_single (these
-      are single-GPU predict leaves).
+    - sbatch.constraint is unset (None): predict leaves use
+      hardware_predict_any_gpu (any GPU), not the H200 pin.
     """
     monkeypatch.setattr("sys.argv", ["dynacell", "predict"])
     leaf = BENCHMARKS / organelle / model / "ipsc_confocal" / f"predict__a549_mantis_{condition}.yml"
@@ -347,8 +347,10 @@ def test_a549_predict_leaf_composes(
     assert bench["dataset_ref"]["target"] == gene_slug
     assert bench["experiment_id"] == f"{organelle}__ipsc_confocal__{model}__a549_mantis_{gene_slug}_{condition}"
 
-    # Single-GPU predict topology: h200 only, not the 4-GPU alternation.
-    assert cfg["launcher"]["sbatch"].get("constraint") == "h200"
+    # Single-GPU predict topology, any GPU: the any-GPU profile composes an
+    # explicit constraint=null. Subscript (not .get) so a dropped hardware
+    # profile surfaces as a failure instead of silently passing.
+    assert cfg["launcher"]["sbatch"]["constraint"] is None
 
 
 def test_manifest_spacing_propagates(monkeypatch) -> None: