# Online course delivery checklist

The six written labs, teaching interactions and workbook are authored. This is a production checklist, not a claim that recordings or learner cloud accounts already exist.

## Before enrolment

- Publish only a reviewed repository revision and bounded Pages artifact. Inspect Git history separately.
- Test student commands on the supported Python version in a clean environment. Record the revision and dependency lock.
- Have a second learner follow the tasks with their own account; preserve actual errors and completion times.
- Confirm model native-tool compatibility and each telemetry destination independently. Budget paid calls explicitly.
- Supply a browser-only lane and an honest offline alternative for students without provider access.

## Repeatable recording plan for every part

Record an introduction (2 minutes), problem and architecture (5), guided build (25), evidence inspection (10), failure/recovery (10), then recap and challenge (8). The student lab timings remain authoritative; edits can shorten the video without removing a required task.

Capture the visible app, selected source changes, exact terminal command, invented input, actual output and evidence classification. Provide captions/transcript, a chapter index, source revision and downloadable worksheet. Use the same synthetic case across all six parts. Keep the original terminal/output capture private until reviewed.

Never show a secrets file, shell history with credentials, private infrastructure identifiers, account selectors, unreviewed browser chrome or another user's data. Review pixels as well as text. Each real run needs its actual outcome; do not splice a successful result over a failed invocation.

## Per-part recording focus

1. Build the visible inbox, connect the prepared student-owned model and finish the real create → select → analyze journey within the hour. A disconnected AI state is an intermediate checkpoint, never the completed deliverable.
2. Start from the empty learner workspace, show the expected failing contracts, implement the engine and rerun unchanged contracts.
3. Bind an actual model, show one native tool round-trip and calculate a sourced estimate; preserve unknown values.
4. Follow one exact trace, exporter result and backend readback; show app/log correlation only for configured destinations.
5. Show the zero-generation policy stop, unknown-tool rejection, controlled failure and retest.
6. Review the deployment packet; record an authenticated hosted invocation only on an authorized target, then inspect telemetry and rollback readiness.

## Student completion rubric

Record each dimension as missing / design only / locally verified / provider verified, with an evidence reference: useful app interactions; bounded agent implementation; native model/tool call; token/cost treatment; trace/export/readback distinction; authorization before inference; safe failure; protocol/auth/hosting design; repeatable handoff.

Do not issue certification from a checkbox download. Review source and actual receipts privately. An OCI design-only submission can finish the written course but does not earn a live hosted-deployment claim.

## Operational handoff

Keep cloud target bindings, owner approvals, spend ceiling, retention, teardown list and exact recovery commands in private operator records. Public course materials use placeholders. No course page creates resources or collects student telemetry.