Upload README.md with huggingface_hub

README.md

@@ -27,8 +27,14 @@ Both graphs also ship in fp16, and both hold their accuracy there:
 
 Every file takes and returns fp32 tensors, so precision is a file swap — the fp16
 pair is 68.2 MB against 133.9 MB. Hiera is attention-heavy, which is why fp16 halves
-it cleanly; dynamic int8 was measured too and is not published, since it leaves the
-convolutional parts in fp32 and came out at the full 109.2 MB for the encoder.
+it cleanly.
+
+Dynamic int8 is measured and not published. It does shrink the encoder — 109.2 MB
+to 29.1 MB at worst-output correlation 0.9958 — but the number that decides is what
+the decoder makes of it. Feeding the int8 encoder's embeddings to the same fp32
+decoder, over 42 clicks the fp32 build answered, the worst decoded mask comes back
+at **0.9375 IoU** against the fp32 encoder's mask (median 0.9956), under the 0.95
+bar every mask on this shelf has to clear.
 
 If size is the binding constraint, look at
 [EdgeTAM](https://huggingface.co/mlboydaisuke/EdgeTAM-ExecuTorch) — same output
@@ -59,13 +65,3 @@ torch.export → to_edge_transform_and_lower(XnnpackPartitioner) → .pte, with:
 - identity `repeat_interleave(1, dim)` calls stripped (delegate mis-sizes their
   decomposition)
 - constant `dense_prompt_embeddings` (no-mask embedding) baked as a buffer
-
-<!-- funnel:v1 -->
-
----
-
-**More models in this format:** [ExecuTorch Model Zoo](https://huggingface.co/collections/mlboydaisuke/executorch-model-zoo-6a7ff328390b63075ffeae5e) — 31 models, each with the recipe that produced it.
-
-**Want a different model on-device?** [Open a request](https://github.com/john-rocky/on-device-requests) — free, open weights only; the export and its measured numbers get published publicly.
-
-<!-- /funnel:v1 -->