Off-the-shelf encoder scale is flat: 22M to 1.5B moves the bound 0.328 to 0.311HEADmain

Seven text encoders against a fixed DINOv2 vision side, priced by the
anchor bound rather than the retired field correlation:

  PPMI          0.289    BGE-large 335M   0.308
  MiniLM 22M    0.328    Qwen 0.5B        0.325
  mpnet 110M    0.341    Qwen 1.5B        0.311
  bert 110M     0.341    part-oracle ceiling 0.359

No correlation with scale and the largest model is not the best; a
retrieval-tuned 335M encoder is worse than bert-base. All seven sit in a
0.29-0.34 band against a part-oracle ceiling of 0.359, so swapping in a
stronger off-the-shelf encoder is closed as a route.

That the band is narrow across unrelated architectures and training sets
is itself the finding: what limits them is common to all of them. They
are generic text encoders whose similarity geometry is organised around
distinctions that do not line up with DINOv2's. Changing which one is
used cannot fix that; changing how both towers are trained might, and
that variable has never been touched.

Also records the VG part-level oracle: annotator-supplied box-to-phrase
correspondence buys 0.023 (0.336 -> 0.359), so part correspondence was
never the problem either.

Co-Authored-By: Claude <noreply@anthropic.com>

1 files changed, 10 insertions, 0 deletions

diff --git a/logs/encoder_matrix.log b/logs/encoder_matrix.log
new file mode 100644
index 0000000..5de4cdb
--- /dev/null
+++ b/logs/encoder_matrix.log
@@ -0,0 +1,10 @@
+256 evaluated scenes, 18496 phrases
+  PPMI (incumbent)       dim=128   rho=0.715  bound=0.2891
+`torch_dtype` is deprecated! Use `dtype` instead!
+  MiniLM-L6 (22M)        dim=384   rho=0.718  bound=0.3281
+  mpnet-base (110M)      dim=768   rho=0.726  bound=0.3406
+  bert-base (110M)       dim=768   rho=0.742  bound=0.3406
+  BGE-large (335M)       dim=1024  rho=0.728  bound=0.3078
+  Qwen2.5-0.5B           dim=896   rho=0.728  bound=0.3250
+  Qwen2.5-1.5B           dim=1536  rho=0.728  bound=0.3109
+{"done": true}