[Hazy142]

O3: REAL CUDA13 ADA WORKER INTEGRATED · 04.10.2026

Code: commit 5609f6d78ce9e746972319974542897947f41a80 · CI: https://github.com/Hazy142/resonarch-nodeos/actions/runs/37224230685 (SUCCESS)

The independently versioned Compute ONE v1 control contract is now end-to-end wired to a real, approved CUDA13 sm_89 worker on the owned Dell G15 RTX4060. The versioned bounded F09-like descriptor and SHA256 state-root are validated before hardware access. Gateway.dispatch(mode="cuda13_ada") requires a caller-supplied verifier, an exact selected Ada-bound lease, explicit administrator NVRTC DLL, bounded input and a replay-free workunit attempt. The only allowed hardware program is our own static CUDA C source in compute_one/cuda13_worker.py, compiled using the legitimately installed NVIDIA NVRTC13.2 API, then run via the official CUDA Driver API. No arbitrary user PTX, binary decompilation or driver modifications. The parent validates worker receipt schema/profile/SM/source SHA/root/tenant/epoch/attempt, output SHA and actual bytes against an independent CPU oracle.

Physical Dell evidence (not GitHub-hosted execution)

• Local NVIDIA Windows driver 610.88, CUDA driver API 13030, installed NVRTC 13.2, RTX4060 Laptop sm_89. No installation/reboot/USB changes.
• Three full leased WorkUnits physically completed: checksum32, xor_u8, cuda13_only_reverse; the latter demonstrates explicitly authorized preferred-Pascal → physical Ada routing (synthetic capability marker, not a CUDA12 inability claim).
• Fourth physical workload: 1,048,576 bytes in a bounded u8 tensor shape [16, 65536], 4,096 CUDA blocks, physical output byte-for-byte equal to CPU oracle; SHA256 5e8e6a984ed51e93f0b6889bb88b5322a9a52a109e2aa9b8636b11dbd050648c.
• Unsigned owner-lab workunit/kernel/PTX/hash observations in docs/evidence/dell-ada-compute-one-o3.json and docs/evidence/dell-ada-compute-one-o3-1mib.json. Opt-in hardware tests run on Dell: 10/10 PASS, including real three-op and max-boundary GPU path.

Reproducibility and CI

• python scripts/verify-layout.py: 32 discovered tests; 30 PASS, 2 expected hardware-only SKIP with no device configured; layout/JSON contracts PASS.
• GitHub Actions SUCCESS for head 5609f6d; Bash syntax and both pinned Buildroot configure-only profiles (Haswell GTX1070Ti, Dell RTX4060) PASS.
• Run on authorized Dell only: python -m compute_one.real_demo --real --large --nvrtc <INSTALLED_CUDA13_NVRTC_DLL>.

Not included in this proof

Still developer-HMAC-only enrollment (not production APE/mTLS/TPM); in-process replay cache (not durable cross-host); bounded JSON/base64 fixtures (not F09/F10 streaming/FiberFEC live integration); no signed image/hardware attestation, T4/Pascal physical worker, real GGML model inference or product performance claim. Worker subprocess separation is not an OS security sandbox. Linux NodeOS Secure Boot USB image and native .ko research remain independent, gated work. No Windows boot/storage, USB, firmware or NVIDIA driver was changed.

Stacked draft PR on existing Dell USB PR #2 (codex/nodeos-dell-g15-usb-headless), not a claim that either is merged to main.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.