Activate bounded ZeroGPU qualification after hosting slot release

README.md

@@ -1,15 +1,27 @@
 ---
-title: TAO GPU Training Lab
+title: TAO ZeroGPU Qualification
 emoji: 🏎️
 colorFrom: green
 colorTo: blue
-sdk: docker
-app_port: 7860
+sdk: gradio
+sdk_version: 6.29.1
+python_version: '3.10'
+app_file: app.py
 pinned: false
 ---
 
-Private CUDA deployment candidate pinned to Zorro909/driving_sim_rs d8ab74c25df70b6a86ed101080e33d5fb785ea6f, matching ald-tao-cpu. Builds unchanged upstream Rust CLI and CUDA library together. Explicit proton math profile; no fast-math override. CUDA 13.0 sm_75/sm_86 plus PTX for compatible newer devices. Driver/device acceptance remains unverified.
+Private native CUDA compatibility probe, not a qualified trainer.
+Uses included ZeroGPU allowance through a 45-second leased call. No paid Jobs
+or dedicated hardware requested. Check the authenticated caller's included quota
+before each call; the app itself does not know account billing balances.
 
-Hardware stays CPU Basic until separately authorized. This is NOT a verified GPU trainer yet. Without a usable GPU library/device, jobs fail preflight; CPU fallback is disabled. Only positive-mutation train-scratch jobs are supported by this wrapper; frozen replay stays on ald-tao-cpu. Existing Formula-only, bounded network/sensor and steering/acceleration restrictions remain. No upstream WebSocket service is exposed. CPU/GPU matched-output and benchmark acceptance remain pending.
+Original Docker source and dashboard remain in this repository but are inactive.
+Rollback: restore README.md from revision ba07e35a39211af7a0deda0ebd9adb2c554b4f38
+and request CPU Basic. BINARY_PROVENANCE.json pins the two exact HF-built binaries
+to that revision and upstream d8ab74c25df70b6a86ed101080e33d5fb785ea6f.
 
-Open upstream PRs 19 and 21 are not included. Keep use private and bounded. No original-game parity claim. Storage is ephemeral: download checkpoints and receipts before restart. PACKAGE_MANIFEST.json covers deployment files; UPSTREAM.json records provenance. No local saved-car library is uploaded.
+No CUDA initialization outside the lease. First check device discovery and a
+three-value sine kernel in process, then the normal Rust gpu-info subprocess.
+No fallback, training, trajectory parity, speedup or H200 claim. A second lease
+must independently pass before testing training. Retained PACKAGE_MANIFEST.json
+describes the prior Docker snapshot only; ZEROGPU_MANIFEST.json covers this overlay.

ZEROGPU_MANIFEST.json

@@ -1,5 +1,5 @@
 {
-  "README.md": "e7ed341a32da89a48dd5194ede7d3066fea0700aab2752f1886f23b7970d6ca8",
+  "README.md": "2ed9310903d06964947ba9008a34f522a5cba48318b183318af938a859911b37",
   "app.py": "60bbb765a5783bb9786aca0b3b910a42eb570fffc6b3072d5220895be978450a",
   "probe_core.py": "97a36ad8cf8efec09d84bb4b2c1ee6794658ceedfd643cd9849ff3c6d85aca9d",
   "requirements.txt": "d4a97517b799a94b3370ea2e11aee3e41304aa12b77bab9abecd0b37e961acaf",