diff --git a/conditioning.py b/conditioning.py
index 9da9ccd5..de01de64 100644
--- a/conditioning.py
+++ b/conditioning.py
@@ -10208,7 +10208,9 @@ class FunPackVideoRefinerV2(FunPackVideoRefiner):
             # discard otherwise. Separate from the value-function block above because
             # wrong_appearance is deliberately NOT reward-admissible there (repair
             # semantics), yet it is the canonical intrusion signal here.
-            if has_previous_run and refinement_key and not learning_profile.get("skip_learning"):
+            # NOT on H3 -- dynashift (the only reader of this bank) is forced off there, and
+            # the sampler no longer captures a pending latent for it either.
+            if has_previous_run and refinement_key and not learning_profile.get("skip_learning") and not _is_h3:
                 try:
                     try:
                         from .negative_memory import consume_pending, is_negative_profile
diff --git a/samplers.py b/samplers.py
index 782e828f..155cdbee 100644
--- a/samplers.py
+++ b/samplers.py
@@ -8266,7 +8266,12 @@ class FunPackLTXAVSceneChainSampler:
             if self._is_nested(_snap):
                 _parts = [t for t in _snap.unbind() if isinstance(t, torch.Tensor) and t.numel() > 0]
                 _snap = max(_parts, key=lambda t: t.numel()) if _parts else None
-            if isinstance(_snap, torch.Tensor):
+            if isinstance(_snap, torch.Tensor) and not self._is_h3:
+                # NOT on H3: this snapshot only ever feeds output_guidance's value function
+                # and DynaShift's negative bank, both forced off for H3 (see self._is_h3
+                # override above) — capturing it here anyway would just be banking video
+                # latents nothing ever reads, the same waste the "disable everything unrelated
+                # to REINS" pass was for.
                 self._save_output_value_snapshot(refinement_key_input, _snap, None)
                 # DynaShift pending candidate: same run/rating pairing as the snapshot, but
                 # the RAW video latent (fp16) — the rating decides whether it becomes a