Reorganize into narrative-ordered experiment directories

01_tactile_delta_heatmaps_p01_vs_p01rand/README.md

@@ -0,0 +1,12 @@
+# 01 — Tactile Δ heatmaps: p01 fixed vs p01rand pool
+
+**Purpose**: motivate the reference-frame choice for `tactile_delta_to_ref` training.
+Compare fixed episode-level p01 anchor vs random no-contact frame sampling.
+
+**Format**: `<episode>_<side>_<mode>.png`, 5 episodes × 2 sides × 2 modes = 20 PNGs.
+
+**Per PNG**: 6 rows (top-4 high-Δ contact + bottom-2 low-Δ background), 5 columns:
+`raw t | ref RGB | ‖z_t‖ 16×16 | ‖z_ref‖ 16×16 | ‖Δz‖ 16×16`.
+
+**Read**: contact regions concentrate in ~5×5 latent tokens (10% of 16×16 map).
+p01rand pool sizes: 8-19% of episode frames.

02_videovae_tactile_debug/README.md

@@ -0,0 +1,15 @@
+# 02 — VideoVAE+ tactile-finetuned encoder artifact diagnosis
+
+3 tests to pin down the right-bottom hotspot + ti=0 boundary bleed.
+
+- **A_synthetic_input_bias**: encode uniform gray (0/64/128/192/255) + noise, look at ‖z‖.
+  gray-192 argmax = **(15, 15) right-bottom corner** — encoder positional bias on
+  non-zero uniform input.
+- **B_per_row_vmax_ep000**: same 3+3 sample frames rendered with global vs
+  per-row vmax vs log-scale — proves the "contact frame right-bottom disappears"
+  observation is a colormap artifact (GroupNorm suppression).
+- **C_temporal_identity_ep000**: encode 16 identical p01 ref frames → 4 latent
+  frames. left/right per-frame max = [52.67, 16.73, 10.17, 10.32] / [57.26,
+  17.09, 10.18, 10.22]. **ti=0 is 5× baseline**, ti=1 is 1.7×. Boundary bleed
+  from temporal encoder's front-replicate pad + spatial encoder's
+  `TemporalAttention` position bias.

03_videovae_visual_debug_base_vae/README.md

@@ -0,0 +1,14 @@
+# 03 — Base VideoVAE+ (visual) artifact diagnosis
+
+Same 3 tests as `02_videovae_tactile_debug/`, but with the base
+(never-finetuned) VideoVAE+ (`sota-4-16z.ckpt`) on `view_left` data.
+
+**Confirms both artifacts are architectural**, not tactile-finetune-specific:
+
+- gray-192 argmax = (15,15) — right-bottom hotspot ALSO on base VAE, and
+  **more severe** (max 20.51 vs tactile 16.94).
+- Test C: base + visual_left = [37.39, 11.21, 7.20, 6.91] → same [5.4, 1.6, 1, 1]
+  temporal bleed pattern.
+
+**Implication**: our whole vision+tactile latent pipeline suffers from these
+issues, not just tactile.

04_videovae_full_latents_tactile_ep000/README.md

@@ -0,0 +1,10 @@
+# 04 — VideoVAE+ full-episode tactile latent video
+
+Every latent frame of `2026-05-10__episode_000.pt` (both tactile sides),
+rendered from **non-overlapping T=16 raw windows** → 4 latent frames per window.
+
+Per video frame: 3 panels: `raw RGB | ‖z‖ GLOBAL vmax | ‖z‖ PER-FRAME vmax`
++ caption with `wi/ti/raw_idx/|z|max/contact_score`. Duration ~69 s at 25 fps.
+
+Watch for: ti=0 anchor (window-boundary bleed) + right-bottom hotspot on
+background frames + spatial concentration on contact frames.

05_videovae_full_latents_visual_ep000/README.md

@@ -0,0 +1,5 @@
+# 05 — VideoVAE+ full-episode visual latent video (base VAE)
+
+Same format as `04_...tactile...`, but for `view_left` of the same episode,
+encoded with the base VAE. Directly comparable to the tactile version at the
+same window/ti index. No contact_score column (visual has none).

06_cosmos_debug/README.md

@@ -0,0 +1,13 @@
+# 06 — Cosmos CV4x8x8 encoder verification
+
+Same tests as `02` + `03`, but with Cosmos CV4x8x8 (encoder JIT from
+`nvidia/Cosmos-0.1-Tokenizer-CV4x8x8`).
+
+- **A_synthetic_input_bias**: gray-192 max/mean ratio = **1.6×** (VideoVAE+
+  base was 4.5×). Right-bottom hotspot much weaker.
+- **C_temporal_identity_***: for each stream, encode 16 identical raw frames
+  → 5 Cosmos latent frames. **All 5 frames identical to 4 decimal digits**
+  (drift mean = 0.0000, bf16 precision noise). Perfect causal identity.
+
+**Verifies Cosmos is strictly causal** and has no temporal boundary bleed
+architectural artifact.

07_cosmos_vs_videovae_recon/README.md

@@ -0,0 +1,25 @@
+# 07 — Cosmos vs VideoVAE+ reconstruction comparison
+
+For each of 5 streams, encode a T=16 raw segment (ep000 frames [1000, 1016))
+with **both** tokenizers, decode back to pixels, compute PSNR/SSIM/L1 per frame.
+
+**Layout per MP4 frame**:
+`orig | Cosmos recon | Cosmos |Δ| | VideoVAE+ recon | VideoVAE+ |Δ|`.
+Each panel 128→256 upsample. Per-frame PSNR/SSIM in title, stream-level
+averages in caption.
+
+**Key result** (average over 16 frames):
+
+| stream | Cosmos PSNR | Cosmos SSIM | VVAE+ PSNR | VVAE+ SSIM |
+|---|---|---|---|---|
+| visual_left | 30.04 | 0.892 | **32.99** | **0.949** |
+| visual_middle | 29.94 | 0.885 | **33.20** | **0.945** |
+| visual_right | 28.82 | 0.889 | **32.10** | **0.948** |
+| tactile_left | **43.98** | **0.971** | 37.42 | 0.965 |
+| tactile_right | **46.68** | **0.983** | 38.83 | 0.979 |
+
+**Interpretation**:
+- VideoVAE+ wins visual by ~3 dB (expected — base VVAE+ was pretrained on
+  natural video too; Cosmos is competitive but slightly behind).
+- Cosmos wins tactile by 6-8 dB **zero-shot**, despite VideoVAE+ being
+  finetuned on our exact tactile distribution.

08_cosmos_full_latents_ep000/README.md

@@ -0,0 +1,27 @@
+# 08 — Cosmos full-episode latent visualization
+
+Cosmos CV4x8x8 encoding of episode_000 across all 5 streams
+(visual_left/middle/right + tactile_left/right).
+
+Two artifacts per stream:
+
+- `<stream>_all_latents.mp4` — non-overlapping T=16 raw windows encoded with
+  Cosmos → 5 latent frames per window. Every latent frame rendered as
+  `raw | ‖z‖ GLOBAL vmax | ‖z‖ PER-FRAME vmax`. Caption tags ti=0 as
+  `[ANCHOR ti=0]` (which by Cosmos design is a single-frame encoding of
+  raw[window_start]).
+- `<stream>_p01_delta.png` — same format as experiment 01: 6 rows × 5 cols
+  showing raw / ref RGB / ‖z_t‖ / ‖z_ref‖ / ‖Δz‖. Reference for tactile:
+  `_contact_meta['ref_p01_idx_<side>']` (same as experiment 01). Reference
+  for visual: raw[0] (episode start).
+
+**Reference latent computed via Cosmos causality**: replicate the ref raw
+frame 16× temporally → take ti=0 (which by strict causality depends only on
+raw[0]). Cleanest possible single-frame encoding.
+
+**Compare to experiments 01/04/05 (VideoVAE+ same episode)** to see:
+- No right-bottom hotspot
+- No ti=0 boundary bleed
+- Cleaner Δ maps on contact frames
+- 5-7× smaller latent magnitudes (closer to unit variance → better for
+  downstream diffusion training).