diff --git a/README_DEFERRED.md b/README_DEFERRED.md
index a7c0569..73f40d9 100644
--- a/README_DEFERRED.md
+++ b/README_DEFERRED.md
@@ -5,15 +5,6 @@ items under *Rejected* are decided unless new information arrives.
 
 ## Parked
 
-### Add `torchvision` (flips the preprocessing backend)
-
-NOTES.md, "Runtime log noise" item 3: without torchvision, transformers falls
-back to `CLIPImageProcessorPil` — slightly slower preprocessing, same results.
-Adding it to `pyproject.toml`/`uv.lock` switches the backend back, so runs
-from before/after must not be mixed (reproducibility); ship it with the next
-`uv.lock` change that happens anyway. Do it only if preprocessing ever becomes
-a bottleneck or a correctness question.
-
 ### Re-pin the pod image off the rc tag
 
 The pod image is stock `runpod/base:1.3.0-rc.164-ubuntu2404`, pinned by digest
@@ -29,6 +20,18 @@ or drop the section.
 
 ## Rejected
 
+### Add `torchvision` (flipped the preprocessing backend; resolved 2026-09-14)
+
+Parked here as "ship with the next `uv.lock` change, and don't mix runs from
+before/after" — then adding `dreamsim` for `dltb-similarity` pulled torchvision
+into `uv.lock` as a side effect, closing the item. The feared reproducibility
+break did not materialize: NOTES.md ("Runtime log noise" item 3) verified that
+the torchvision-backed transformers image processors are reachable only from
+the safety checker and image-encoder code paths, both dead in dltb's
+pipelines, so pixel output is unchanged. The caveat stands: re-check before
+ever enabling a safety checker or an image encoder, and compare deliberately
+against pre-change runs instead of mixing them if that happens.
+
 ### Build a custom pod image (retired 2026-09-13)
 
 The first commissioning baked the locked venv into a ~12 GB custom image
diff --git a/README_RUNPOD.md b/README_RUNPOD.md
index fdfe912..1356547 100644
--- a/README_RUNPOD.md
+++ b/README_RUNPOD.md
@@ -114,8 +114,11 @@ The image itself does **not** count against the container disk (`df
 | **all five** | **~87.5 GB** |
 
 The default cache policy keeps every swept model (all five ≈ 87.5 GB) + xet
-(≤10 GB) + outputs (a full video sweep is a few GB) — comfortable on the
-150 GB disk, and revisiting an earlier model costs no re-download. With
+(≤10 GB) + outputs (a full video sweep is a few GB) — plus ~2.7 GB of DreamSim
+weights in `untracked/models` when `dltb-similarity` / `SMOKE_SIMILARITY=1`
+is exercised (not part of bundles, so every fresh pod re-downloads them) —
+comfortable on the 150 GB disk, and revisiting an earlier model costs no
+re-download. With
 `EVICT_CACHE=1` (keep-one, `input/inputs.env`-configurable) the steady-state
 requirement drops to one model (~33 GB max) + xet + outputs ≈ **45 GB**, for
 smaller container disks.