All runbooks
runbookuapiredacted view

UAPI NUC Gemma/SLM Smoke-Test Gate 2026-06-22

Acting-operator gate approving NUC benchmark planning and preflight only, with execution still gated.

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

UAPI NUC Gemma/SLM Smoke-Test Gate 2026-06-22

Acting-operator gate approving NUC benchmark planning and preflight only, with execution still gated.

Promotion Metadata

  • Source feedback id: 106
  • Source feedback ids: 106
  • Promotion request id: 37
  • Feedback category: beneficial
  • Feedback source: codex-acting-operator-2026-06-22
  • Feedback created at: 2026-06-22T18:05:08.015073+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 106

  • Product: uapi
  • Category: beneficial
  • Related slug: [REDACTED:high-entropy]
  • Source: codex-acting-operator-2026-06-22
  • Created at: 2026-06-22T18:05:08.015073+00:00

UAPI NUC Gemma/SLM Smoke-Test Gate 2026-06-22

Gate id: [REDACTED:high-entropy]. Acting operator: Codex under bounded delegation. Source evidence: NUC inventory context, Grok route matrix context, auto-pickup policy.

Status: APPROVED FOR PLANNING AND PREFLIGHT ONLY. Model execution is not approved by this gate.

Approved now: - Grok/Codex may prepare exact benchmark command plan and preflight checklist for local.nuc.gemma.npu and local.nuc.gemma.140t. - Read-only checks may verify Python, package availability, disk free, driver versions, process load, and route assumptions. - Route claims must remain candidate-only and must cite inventory evidence.

Not approved by this gate: - Installing packages, downloading model weights, running inference, changing drivers, enabling WSL, exposing endpoints, public routes, background jobs, sustained load, destructive actions, or secret access.

Minimum execution gate requirements before any smoke test: 1. Exact commands and package/model sources published to Explorer. 2. Max model size: small Gemma/SLM class only, preferably 2B or smaller/quantized. 3. Max run: one-shot human-supervised smoke test, 1-3 prompts, max-new-tokens 32-64, no user/private data. 4. Metrics: first-token latency, tokens/sec, load time, memory, device used, stability, fallback, and power/thermal if safely available. 5. Output sanitized and published to Explorer before any route is called production-capable. 6. NUC CPU remains reserved for Conductor/Foundry unless a later gate explicitly permits CPU-heavy inference.

Known facts: - NUC-IPEX: Core Ultra 9 285H, Arc 140T GPU, Intel AI Boost NPU, ~64 GiB RAM. - Default Python has no AI stack installed. - SN850X not observed on NUC; NUC boot/system NVMe is WD_BLACK SN8100 2000GB.

Next safe task: Grok should produce a benchmark preflight and exact command plan under 4000 chars. Codex then reviews and either publishes an execution gate or blocks.