Eval guidance for cudaq-guide

Developer guidance for generating and refining evals.json. This outranks generated defaults during NV-BASE/NV-ACES generation and refinement.

How do I install CUDA-Q and confirm it works?
Write and run a minimal CUDA-Q program to verify my setup.
Which simulation target should I use for a circuit too large for one GPU?
How do I run a CUDA-Q kernel on real QPU hardware from a given provider?
How do I run many independent circuits in parallel across multiple GPUs?
What applications can I build with CUDA-Q?
(negative) Unrelated creative or general-programming requests.
(negative) Near-miss prompts that mention CUDA or "install" but are not about CUDA-Q (e.g. installing PyTorch with CUDA), to guard against over-routing.

The agent read skills/cudaq-guide/SKILL.md before acting.
The agent recommended the documented target/option for the scenario (nvidia, nvidia --target-option mgpu/mqpu, qpp-cpu, tensornet).
The agent followed the documented workflow (e.g. validate install with the Bell state example; for QPU, identify the provider technology and advise emulate=True before real hardware).

cudaq-guide is a documentation/onboarding skill with no executable script, so expected_script is null for every case and the agent should never run a script.
Ground truth is intentionally derived from SKILL.md content (the GPU target table, QPU two-step dialogue, parallelize mgpu/mqpu guidance), so cases remain answerable in an isolated workspace without staging the repo's docs/sphinx .rst files.
Keep the CI-gated dataset small (P0 smoke) for the 1-hour NV-CARPS limit. Deeper, doc-reading cases that require staging docs/sphinx/** can follow once the publish path is stable (would need skill_workspace.mode: group or fixtures under evals/files/).
Negative cases set expected_skill: null and should_trigger: false.

Third-party checks

No alerts19d3 checks · Risk SAFE

Gen Agent Trust Hub19d
The cudaq-guide skill is an onboarding tool for the NVIDIA CUDA-Q platform, providing instructions for installation, quantum programming, and hardware connectivity. It follows security best practices by advising environment-variable-based secret management and links to official documentation.
Socket19d
No alerts
Snyk19d
Risk: LOW · No issues

Capability

Other metadata

tags: [ "cuda-quantum", "quantum-computing", "onboarding", "getting-started", "authoring", "kernels", "nvidia" ]
tools: [ "Read", "Glob", "Grep" ]
compatibility: Python 3.10+, C++ 20
metadata: { "author": "CUDA-Q Team <cuda-quantum@nvidia.com>", "tags": [ "cuda-quantum", "quantum-computing", "onboarding", "getting-started", "nvidia" ], "languages": [ "python", "c++" ], "domain": "quantum" }

/cudaq-guide