diff --git a/scripts/b6_run.py b/scripts/b6_run.py
index 3acbc95..aa39589 100644
--- a/scripts/b6_run.py
+++ b/scripts/b6_run.py
@@ -83,12 +83,16 @@
 DEFAULT_PER_RUNG_H100_HR_RATE = 2.0  # Shadeform spot reference
 SEED_RETRY_OFFSET = 1000  # second-attempt seed = primary + 1000
 
-# Standard rungs for B6 — locked to the v0.10 ladder. Each rung
-# carries its dim + n_layers for audit reproducibility.
+# Standard rungs for B6 — a SINGLE S3 rung, matching the validated
+# ρ=0.614 transfer run (experiments/2026-06-transfer-credibility/run_b6_s3.sh).
+# Only the S3 cell feeds the Spearman score vector, so evaluating the same
+# checkpoint at S1/S2/S3 (which differ only by an advisory scale_label —
+# run_ladder_eval re-runs the same checkpoint+patch per rung) is 3× redundant
+# compute. `n_examples_per_task=100` caps per-task eval so the run stays inside
+# the validator time budget — without it, mode-0 evaluates ALL examples (e.g.
+# bigbench_qa_wikidata ≈ 20k rows), which blows the cap.
 _B6_STANDARD_RUNGS: tuple[LadderRungSpec, ...] = (
-    LadderRungSpec(scale_label="S1", dim=256, n_layers=4),
-    LadderRungSpec(scale_label="S2", dim=512, n_layers=12),
-    LadderRungSpec(scale_label="S3", dim=768, n_layers=12),
+    LadderRungSpec(scale_label="S3", dim=768, n_layers=12, n_examples_per_task=100),
 )
 
 
@@ -189,8 +193,8 @@ def estimate_cost(
     """Estimate USD cost of a recipe's wall-clock at the spot rate.
 
     The wall_clock_s is from the run_ladder_eval combined report; it's
-    the max across rungs (per merge_rung_reports), so a 3-rung recipe at
-    S3=300s would produce wall_clock_s ≈ 300s. Multiply by spot rate.
+    the max across rungs (per merge_rung_reports), so the single S3 rung at
+    ≈300s produces wall_clock_s ≈ 300s. Multiply by spot rate.
     """
     h100_hr = wall_clock_s / 3600.0
     return h100_hr * per_rung_h100_hr_rate
diff --git a/tests/test_b6_run.py b/tests/test_b6_run.py
index 1b1921d..bf826e7 100644
--- a/tests/test_b6_run.py
+++ b/tests/test_b6_run.py
@@ -42,11 +42,12 @@ def _set_test_mode(monkeypatch):
 
 
 def _b6_recipe_with_single_rung_only() -> tuple:
-    """Override the standard rungs to a single S3 rung for tests against
-    the synthetic CLI (which hard-codes its cell key).
+    """Pin the rungs to a bare single S3 rung (no per-task cap) for tests
+    against the synthetic CLI (which hard-codes its cell key).
 
-    NOTE: run_one_recipe + run_b6 use _B6_STANDARD_RUNGS (3-rung) by
-    contract. For tests we monkey-patch the constant.
+    The canonical `_B6_STANDARD_RUNGS` is also a single S3 rung but carries
+    `n_examples_per_task=100`; tests use synthetic fixtures where the cap is
+    irrelevant, so this keeps them decoupled from that value.
     """
     from validator.ladder import LadderRungSpec
     return (LadderRungSpec(scale_label="S3", dim=768, n_layers=12),)