Imported from Pein2017/CoordExp (.codex/skills/qwen3-vl-execution/SKILL.md). Install upstream with npx skills add Pein2017/CoordExp --skill qwen3-vl-execution. Copyright stays with the author.

Qwen3-VL Execution

Use this as a technical entry, not an additional review or launch gate. Work in the user-selected checkout and respect the requested read/write boundary. Resolve its actual model/runtime and scientific owner before choosing an API: the research base supports direct native execution; packed training and strict scored inference have different contracts. Do not move a research caller into either framework just to obtain a model or token scores.

Load only the relevant knowledge

• For architecture, processor/image grids, MRoPE, compact logits, cache or DeepStack hooks, read the matching section of the selected checkout's docs/standards/upstream/QWEN_VL.md. The maintained research copy is here. Its installed-version scope is explicit; inspect the actual loaded dependency when a version-sensitive behavior matters.
• For precision, normalization, causal alignment, batches or template assembly, use only the applicable section of execution checks. It points to code/spec/test owners rather than another execution framework.
• For candidate-set probability, row scoring, finite-set mass or reducers, read its Objective, denominator and distributed reduction section. Bind the real consumer schema and test empty support, null/finite representation and EOS behavior rather than inferring them from nonempty examples.
• For a training launch, discover the selected checkout's config/launcher; this skill owns no GPU count, batch size, precision default or launch budget.

Start from the named source/artifact and retrieve one relevant contract or counterexample. Familiar tensor operations do not require a tutorial or a full checklist. Correct shapes and a finite scalar do not by themselves prove correct target alignment, reduction, autograd or conditioning.

When implementing or running an already frozen research probe, also use the Frozen Probe Execution Packet for producer identity, attempt handling and dependent-artifact readiness.

Close the specific question

State the relevant input-to-output relationship and its owner, then inspect or exercise the smallest real caller that can distinguish the suspected mistake. Reuse existing tests and receipts; choose a real model check only when mocks or saved inputs cannot settle the claim. Keep generation, gradients and numerical parity claims separate. An existing diagnostic does not become a new acceptance gate unless the active contract makes it one.

Keep research populations, rewards, owner acceptance and stop rules in their direction. Use model-diagnosis for an observed behavioral symptom; model-innovation-risk-audit only for a decision-bearing mechanism-fidelity question; full-pipeline-smoke when the real entry/distributed/persistence path is the unresolved risk. Do not run them as a sequence of approvals.