diff --git a/src/projected_grpo/train.py b/src/projected_grpo/train.py
index 15f5084..846b21b 100644
--- a/src/projected_grpo/train.py
+++ b/src/projected_grpo/train.py
@@ -677,8 +677,8 @@ class StepLogger:
         # (the substrate); single-mode runs would just duplicate hack_s.
         self._modes = modes if len(modes) > 1 else []
         for m in self._modes:
-            cols.append(_Col(f"hk_{MODE_CODE[m]}", 6, f"hk_{MODE_CODE[m]}", "frac",
-                             f"cumulative student hacks of {m}"))
+            cols.append(_Col(f"hk_{MODE_CODE[m]}", 5, f"hk_{MODE_CODE[m]}", "d",
+                             f"student hacks of {m} THIS step (current batch, not cumulative)"))
         cols += [
             _Col("lp_s", 6, "lp_s↓", "+.2f", "mean student gen_logp (diagnostic)"),
             _Col("lp_t", 6, "lp_t↑", "+.2f", "mean teacher gen_logp; off-policy gap = lp_s-lp_t"),
@@ -1156,6 +1156,7 @@ def main(cfg: Config) -> int:
         step_rollouts: list[dict] = []  # student completions this step -> rollout_log_path
         agg_is_student: list[bool] = []
         agg_is_ablated: list[bool] = []  # deploy-mode (quarantine-ablated) student rows -> free per-step deploy proxy
+        step_mode_hacks: dict[str, int] = {}  # THIS step's student hacks per mode (the hk_<mode> columns; reset each step so they don't grow)
         agg_logp: list[float] = []  # per-rollout mean per-token gen_logp (student's logp on rollout tokens)
         agg_comp_lens, agg_finished, n_skipped = [], [], 0
         agg_loss = 0.0
@@ -1377,6 +1378,7 @@ def main(cfg: Config) -> int:
                 mode_rollouts[pm] = mode_rollouts.get(pm, 0) + 1
                 if r.exploited:
                     mode_hacks[pm] = mode_hacks.get(pm, 0) + 1
+                    step_mode_hacks[pm] = step_mode_hacks.get(pm, 0) + 1
                     if pm not in mode_first_step:
                         mode_first_step[pm] = step
                 # First full example of each hack mechanism -> verbose log (with
@@ -1882,11 +1884,11 @@ def main(cfg: Config) -> int:
             "gt_t": (gt_t_n, n_t) if n_t else (0, 0),
             "hack_s": (hack_s_n, n_s) if n_s else (0, 0),
             "hack_t": (hack_t_n, n_t) if n_t else (0, 0),
-            # Per-mode CUMULATIVE student exploit rate (which loophole classes are
-            # learnt, how strongly). From the running tallies, so it rises over the
-            # run; StepLogger only renders these on multi-mode (substrate) runs.
-            **{f"hk_{MODE_CODE[m]}": (mode_hacks.get(m, 0), mode_rollouts.get(m, 0))
-               for m in run_modes},
+            # Per-mode student hacks THIS step (current batch count, not cumulative --
+            # cumulative grew unboundedly and read as noise). The running mode_hacks/
+            # mode_rollouts tallies still feed the end-of-run substrate learning table.
+            # StepLogger only renders these on multi-mode (substrate) runs.
+            **{f"hk_{MODE_CODE[m]}": step_mode_hacks.get(m, 0) for m in run_modes},
             # Per-mechanism on student rollouts only. Used by final-tail BLUF for
             # cross-mechanism HACK_A / HACK_B; hidden from the per-step table to
             # avoid column bloat (rendered only in the markdown dump below).