deliver-cv-project

Deliver a previously evaluated CV capability as a replayable artifact and an explicit engineering handoff. The exit criterion is not “deployment requested.” It is a completed delivery contract whose artifact kind, acceptance evidence, commands, dependencies, data boundary, rollback, and remaining external checks are truthful and tested.

<entry_gate>

Read existing code, model metadata, .vision-delivery/eval-*.md, ledger rows, generated artifacts, and sample outputs before asking questions.

Require:

1. a model/pipeline identity and output schema;
2. a machine-readable acceptance artifact frozen before the candidate result, with its acceptance_id, project-relative path, and SHA-256 digest;
3. independent acceptance evidence for the selected version;
4. the intended consumer and runtime environment.

When the user says a hosted model works, verify its current entity/version and evaluation through the upstream platform resources. When the model is local, inspect its weights/config and rerun the recorded acceptance command. If acceptance evidence is absent or stale for the selected version, route back to the owning modality skill; do not turn an unmeasured model into production merely because it exists.

For a novice request such as “make this useful in my factory,” ask at most three plain questions:

• “What should happen when the model finds something?”
• “Where do the images come from: files, an application, or a live camera?”
• “Must data stay on this machine, or may it go to a hosted service?”

</entry_gate>

Platform execution boundary. Read ../../resources/roboflow-platform-lookup.md. Use installed official Roboflow skills or current MCP skill resources only for read-only discovery. For every platform-specific export, Workflow mutation, endpoint, device, deployment, telemetry mutation, active-learning action, paid action, or data movement, emit a sourced action brief and stop for execution through an external host/provider approval control. Sentinel never invokes those actions. Sentinel owns delivery selection, acceptance, artifact hardening, and the handoff record. If no upstream source is available, stop at a scaffold; do not guess an API, model ID, container tag, endpoint, or UI path.

Step 1 — Select the delivery contract

Classify the requested outcome:

Outcome	Artifact kind	Required proof
Calls a hosted provider endpoint	hosted-client	current upstream client shape, approved data movement, live request smoke
Runs exported weights on the user's machine	local-runtime	current upstream export, dependency lock, offline inference smoke
Contract only; transport/model unavailable	scaffold	deterministic self-test plus explicit missing live check
Managed service/device configuration	provider-managed handoff plus client kind	economic/spend consent, upstream status evidence, consumer smoke

Never call a hosted client local inference. Never call exported weights a local runtime until they load and infer with networking disabled.