Remove validation record and restore model card

README.md

@@ -23,12 +23,10 @@ library as ExecuTorch `.pte` programs, ready to run on device.
 | --- | --- | --- | --- |
 | `e2b/mlx/gemma_4_e2b_mlx_int4.pte` | mlx | int4 | 2,048 |
 | `e2b/vulkan/gemma_4_e2b_vulkan_8da4w.pte` | vulkan | 8da4w | 2,048 |
-| `e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte` | xnnpack | 8da4w | 4,096 |
+| `e2b/xnnpack/gemma_4_e2b_xnnpack_8da4w.pte` | xnnpack | 8da4w | 4,048 |
 
 
 
-XNNPACK accepts prefill chunks of up to **512 tokens** within its 4,096-token context. The chunk bound prevents sliding-window ring-cache eviction during long prefills. The exported binary, embedded model schema, and companion config use the same bound.
-
 ## Repository structure
 
 ```

e2b/xnnpack/validation.json

@@ -1,40 +0,0 @@
-{
-  "passed": true,
-  "metadata_verified": true,
-  "sha256": "a5799ecbed3db94029a5c2ce847b138159c1d2b6c0f09f2f153a790211eb6409",
-  "date": "2026-10-15",
-  "device": {
-    "model": "Samsung SM-G986B",
-    "soc": "Exynos 990",
-    "gpu": "Mali-G77"
-  },
-  "backend": "XNNPACK",
-  "max_prefill_tokens": 512,
-  "max_context_tokens": 4096,
-  "transfer_sha256_verified": true,
-  "prefill_513_rejected": true,
-  "checks": {
-    "omni_tool": {
-      "finite_logits": true,
-      "output": "The weather in Warsaw is 18 degrees and sunny.<turn|>",
-      "first_token": 818,
-      "termination_token": 106
-    },
-    "omni_long": {
-      "finite_logits": true,
-      "output": "The weather in Warsaw is 8 degrees and sunny.<turn|>",
-      "first_token": 818,
-      "termination_token": 106
-    },
-    "transcription": {
-      "finite_logits": true,
-      "output": "And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.<turn|>",
-      "first_token": 3133,
-      "termination_token": 106
-    }
-  },
-  "limitations": [
-    "The 2,102-token repetitive tool fixture answers 8 degrees instead of the supplied 18 degrees, also observed with the previous published model. This verifies continuation and finite logits, not factual accuracy.",
-    "These fixtures did not reproduce the reported immediate EOS; they do not establish correctness for every prompt."
-  ]
-}