recon README: concise + embed contact-fidelity images

cosmos_threshold_sweep/recon_contact_fidelity/README.md

@@ -1,22 +1,26 @@
-# Tokenizer contact-preservation diagnostic
+# Does Cosmos preserve tactile contact? — Yes (don't finetune)
 
-Does the encode→decode round-trip preserve tactile contact (area/intensity)?
-Cosmos = frozen zero-shot; VVAE+tac = VideoVAE+ finetuned on this tactile distribution.
-`ret` = mean(area_recon)/mean(area_orig) on a contact-rich 16-frame window (1.0 = perfect).
+Encode→decode a contact frame with **Cosmos** (frozen, zero-shot) vs **VideoVAE+**
+(finetuned on this tactile data), then recompute contact `area`. `ret = area(recon)
+/ area(orig)`; **1.0 = perfect**.
 
-| episode | side | seg | area orig | Cosmos (ret) | VVAE+tac (ret) | vis PSNR |
-|---|---|---|---|---|---|---|
-| episode_016 | left | contact | 0.1097 | 0.1136 (1.04) | 0.2972 (2.71) | 33.7 |
-| episode_016 | left | background | 0.0000 | 0.0045 (4501.34) | 0.1548 (154815.67) | 33.7 |
-| episode_016 | right | contact | 0.1723 | 0.1845 (1.07) | 0.2317 (1.34) | 33.7 |
-| episode_016 | right | background | 0.0000 | 0.0023 (2273.56) | 0.0620 (62046.05) | 33.7 |
-| episode_005 | left | contact | 0.3435 | 0.3485 (1.01) | 0.4721 (1.37) | 31.6 |
-| episode_005 | left | background | 0.0000 | 0.0046 (4585.27) | 0.1584 (158359.53) | 31.6 |
-| episode_005 | right | contact | 0.3389 | 0.3366 (0.99) | 0.3801 (1.12) | 31.6 |
-| episode_005 | right | background | 0.0002 | 0.0019 (8.54) | 0.0451 (207.61) | 31.6 |
-| episode_000 | left | contact | 0.2572 | 0.2530 (0.98) | 0.3666 (1.43) | 29.9 |
-| episode_000 | left | background | 0.0000 | 0.0063 (6328.58) | 0.1394 (139408.11) | 29.9 |
-| episode_000 | right | contact | 0.2047 | 0.2142 (1.05) | 0.2525 (1.23) | 29.9 |
-| episode_000 | right | background | 0.0000 | 0.0002 (228.88) | 0.0437 (43716.43) | 29.9 |
+| ep / side | Cosmos ret | VVAE+tac ret |
+|---|---|---|
+| ep016 R | **1.07** | 1.34 |
+| ep005 R | **0.99** | 1.12 |
+| ep000 R | **1.05** | 1.23 |
 
-**Read**: if Cosmos `ret` << VVAE+tac `ret` on contact segments, Cosmos loses contact detail → finetuning justified. Compare Cosmos tactile fidelity to its visual PSNR (in-distribution) to confirm the OOD hypothesis.
+**Cosmos keeps the contact area within ±7%** and its mask hugs the original; the
+*finetuned* VVAE+ actually **inflates** area (1.1–2.7) and sprays background
+artifacts. So the latent-floor problem was **not** lost contact detail — it was
+anchor-latent contamination (since fixed) + tokenizer noise-sensitivity (sidestepped
+by pixel-space labeling). **Finetuning is not justified.**
+
+Each panel: `orig | Cosmos recon | VVAE+tac recon | |orig−Cosmos|`, area labelled.
+
+![2026-05-10__episode_000_left_contact.png](https://huggingface.co/datasets/yxma/vtwm-tactile-latent-delta-viz/resolve/main/cosmos_threshold_sweep/recon_contact_fidelity/2026-05-10__episode_000_left_contact.png)
+![2026-05-10__episode_000_right_contact.png](https://huggingface.co/datasets/yxma/vtwm-tactile-latent-delta-viz/resolve/main/cosmos_threshold_sweep/recon_contact_fidelity/2026-05-10__episode_000_right_contact.png)
+![2026-05-11__episode_005_left_contact.png](https://huggingface.co/datasets/yxma/vtwm-tactile-latent-delta-viz/resolve/main/cosmos_threshold_sweep/recon_contact_fidelity/2026-05-11__episode_005_left_contact.png)
+![2026-05-11__episode_005_right_contact.png](https://huggingface.co/datasets/yxma/vtwm-tactile-latent-delta-viz/resolve/main/cosmos_threshold_sweep/recon_contact_fidelity/2026-05-11__episode_005_right_contact.png)
+![2026-05-11__episode_016_left_contact.png](https://huggingface.co/datasets/yxma/vtwm-tactile-latent-delta-viz/resolve/main/cosmos_threshold_sweep/recon_contact_fidelity/2026-05-11__episode_016_left_contact.png)
+![2026-05-11__episode_016_right_contact.png](https://huggingface.co/datasets/yxma/vtwm-tactile-latent-delta-viz/resolve/main/cosmos_threshold_sweep/recon_contact_fidelity/2026-05-11__episode_016_right_contact.png)