Upload folder using huggingface_hub

data.json

@@ -1,5 +1,16 @@
 {
-  "generated_utc": "2026-07-31 07:18:02Z",
+  "generated_utc": "2026-07-31 07:55:33Z",
+  "seeds": {
+    "n": 3,
+    "incr_mean": -0.009137061718228529,
+    "incr_std": 0.005614727687208816,
+    "incr_all": [
+      -0.010297063318744483,
+      -0.01408118656974655,
+      -0.003032935266194556
+    ],
+    "sign_stable": true
+  },
   "tokenizer": "cosmos-cv4x8x8",
   "zero_baseline": 1.2195221489913883,
   "verdict": "NO TRANSFER",

index.html

@@ -18,7 +18,7 @@
       margin:20px 0;font-size:14px}
 </style>
 <h1>Stage-1 gate — does the action-motion encoder transfer to Cosmos?</h1>
-<div class=sub>tokenizer <code>cosmos-cv4x8x8</code> &middot; generated 2026-07-31 07:18:02Z</div>
+<div class=sub>tokenizer <code>cosmos-cv4x8x8</code> &middot; generated 2026-07-31 07:55:33Z</div>
 <p><span class=verdict>NO TRANSFER</span> &nbsp; the encoder learns well (D beats zero by 32.3%, C by 33.0%), but the action map adds NOTHING over tactile-only (D is 1.0% worse than C; Wan2.2 had D +3.3% over C). The conditioning signal Stages 2-3 inject is redundant here.</p>
 <div class=note>
  The encoder predicts per-latent-transition tactile flow. <b>Active-pixel EPE</b> is the
@@ -42,7 +42,7 @@
  <b>Judge the gate on the ordering</b> (D &lt; C &lt; B &lt; zero) <b>and on D's
  relative margin over zero</b>, both measured entirely within the Cosmos column.
 </div>
-<div class=note><b>Single seed — read the margin with care.</b> The verdict turns on a D-vs-C difference of 1.0%, which is small enough to sit inside run-to-run variation. The supported claim is "the action map shows no <i>measurable</i> increment", not that D is reliably worse than C. Confirming the sign needs repeated seeds; for the decision this gate exists to make — whether to spend ~150 epochs on Stages 2-3 — the absence of a measurable increment is already the answer.</div>
+<div class=note><b>Seeds.</b> 3 seeds per ablation. The D-vs-C increment is -0.9% &plusmn; 0.6% (std). The sign is stable across seeds.</div>
 <h2 style="font-size:17px;margin-top:32px">What the two ablations actually see</h2>
 <p>D's input is a strict superset of C's: the same tactile latent, plus the action
 map. <b>D &minus; C is therefore exactly what knowing the sensor's own motion buys