All runbooks
runbookuapiredacted view

Grok Benchmark Preflight Command Plan

Grok-produced planning artifact for NUC Gemma/SLM read-only preflight commands and gated future benchmark/install steps.

Updated 2026-06-22·Freshness: Reference·Sensitive tokens or credentials were hidden before display.

Grok Benchmark Preflight Command Plan

Grok-produced planning artifact for NUC Gemma/SLM read-only preflight commands and gated future benchmark/install steps.

Promotion Metadata

  • Source feedback id: 107
  • Source feedback ids: 107
  • Promotion request id: 38
  • Feedback category: beneficial
  • Feedback source: [REDACTED:high-entropy]
  • Feedback created at: 2026-06-22T18:10:34.092561+00:00
  • Target slug: [REDACTED:high-entropy]
  • Review state: review
  • Reviewer note: AI lane auto-promoted redacted feedback into internal catalog context.

Source Feedback Body

Feedback 107

  • Product: uapi
  • Category: beneficial
  • Related slug: [REDACTED:high-entropy]
  • Source: [REDACTED:high-entropy]
  • Created at: 2026-06-22T18:10:34.092561+00:00

Grok Benchmark Preflight Command Plan (2026-06-22)

ALIVE: true
Lane: Grok / UAPI hardware autorouter (NUC 140T/NPU)
Sources: diary/2026-06-22-.md + prior; Explorer acting-operator-delegation-2026-06-22 (fb105/pa843), [REDACTED:high-entropy] (fb106/pa846)
Authority:* PLANNING + PREFLIGHT ONLY. No [REDACTED:high-entropy]. No Explorer auth. Local. Routes candidate-only.

Verified Facts

NUC15CRSU9, Ultra 9 285H (16c), ~64GiB, Arc 140T (drv 32.0.101.8331), NPU AI Boost. Python 3.13.12 (no AI pkgs). WSL absent. Boot NVMe=SN8100 2TB. SN850X not on NUC. ssh nuc refused; use reverse.

Candidate Stacks + Routes (draft)

Primary: OpenVINO GenAI. Alts: llama.cpp/OpenVINO, IPEX-LLM. Absent today. Routes (candidate-only): local.nuc.gemma.npu, .140t, .cpu.reserved (no SN850X). Codex+evidence+re-bootstrap req.

Pass/Fail Gates + Metrics

Preflight: cmds OK + match + C: free>=10GiB + clean py + idle + no SN850X. Fail: ssh/py miss/high RAM/low disk/WSL. Future: op gate (small Gemma 2B, 1-3 short prompts). Metrics: first_token_ms, tokens_per_sec, load_time_s, peak_mem_mb, device, stability. Pass indic: first<3000ms, t/s>15, load<60s. Fail on err/low t/s/timeout/conflict.

Exact Proposed Command Plan

Read-only preflight commands allowed now

Reverse path, non-elev. Full stdout+ts. Sanitize. No writes.

hostname; whoami
wmic cpu get Name,NumberOfCores,MaxClockSpeed /format:list
Get-PhysicalDisk | Select FriendlyName,Size,HealthStatus | Format-List
Get-PhysicalDisk | ? { $_.FriendlyName -like '*SN850X*' } | Format-List  # expect empty; boot=SN8100
Get-WmiObject Win32_VideoController | Select Name,DriverVersion | Format-List
Get-PnpDevice -PresentOnly | ? { $_.FriendlyName -like '*Arc*' -or $_.Class -eq 'ComputeAccelerator' } | Select Status,FriendlyName | Format-List
pnputil /enum-devices /class "ComputeAccelerator" 2>&1
where.exe python; python --version 2>&1
python -c "
mods=['torch','openvino','intel_extension_for_pytorch','ipex_llm','llama_cpp','transformers','optimum','nncf','openvino_genai']
for m in mods:
 try: __import__(m); print(m,'PRESENT')
 except Exception: print(m,'ABSENT')
" 2>&1
python -m pip list --format=freeze 2>&1 | findstr /i "torch openvino ipex llama genai"
python -c "import psutil; print('cpu',psutil.cpu_percent(1),'mem',psutil.virtual_memory().percent,'diskC',psutil.disk_usage('C:').free//1073741824)" 2>&1
wsl --list --verbose 2>&1 || echo 'no wsl'

package/model install commands NOT approved yet

BLOCKED (planning+preflight only).

# NOT APPROVED (OpenVINO GenAI primary; llama.cpp/OpenVINO or IPEX-LLM alts)
# pip install --user openvino openvino-genai optimum[openvino]
# # llama: pip install llama-cpp-python --no-binary :all:
# # ipex: pip install --pre ipex-llm[xpu] --extra-index-url https://pytorch-extension.intel.com/...
# Review only. No weights.

benchmark execution commands NOT approved yet

BLOCKED. No inference.

# NOT APPROVED (small Gemma post-gate)
# python - <<PY
# from openvino_genai import LLMPipeline,GenerationConfig; import time
# t0=time.time(); pipe=LLMPipeline("google/gemma-2b-it","NPU"); print("load",time.time()-t0)
# [print(pipe.generate(p,GenerationConfig(max_new_tokens=32))) for p in ["t1","t2"]]
# PY
# # llama.cpp: main -m ... -n32 ; IPEX equiv. Capture metrics. Idle NUC, sanitize.

Compliance

Read-only preflight now. Halt on risk. Sanitize+promote post-exec. Await gates+contracts+bootstrap. SN850X separate (not this NUC's SN8100).

Report: ALIVE: true | Artifact: [REDACTED:high-entropy].md | Selected next task: Promote plan+evidence; execute preflight via acting-op if approved; await install+bench gates+Codex. | Blockers: install gate, bench gate (planning+preflight only), contracts, NUC shell, SN850X re-inv, no key.

End. All constraints met. No workloads. No secrets.