Demonstrate the journey
This assessment is a self-review, not an independently proctored qualification. Complete the core code extension, the capstone, and the clearly labeled advanced design exercises before marking the full roadmap complete. The certificate uses the website's existing self-reported completion flow.
Explain without opening the lesson
- What did we train in PocketChat? Nothing; identify the pretrained model, system instruction, and supplied context separately.
- Why can a socket read not be treated as one JSON response? Explain SSE event framing and arbitrary network boundaries.
- Why does EOF without STOP fail? Distinguish a completed answer from interrupted partial text.
- Which operations must happen when the learner taps Stop? Trace attempt invalidation, coroutine cancellation, and HTTP cancellation.
- Which turns are eligible for context? Explain failed/stopped exclusions, complete pairs, chronology, and omission reporting.
- What survives rotation, backgrounding, and process death? Name each owner and the checkpoint limitation.
- Why does deleting a conversation leave a saved answer? Explain snapshot semantics and provider-side data boundaries.
- How does the Android Engineers guide differ from RAG? Explain static directory context and known-destination buttons.
- Why is a successful release build not a production chat deployment? Identify the missing backend boundary.
- What does a model's proposed tool call authorize? Nothing on its own; explain validation, approval, and idempotency.
Evidence checklist
| Deliverable | Passing evidence |
|---|---|
| Reproducible setup | Pinned revision, toolchain, successful build |
| Context-policy extension | Expected failing tests followed by a green full suite |
| Streaming behavior | Parser, EOF, cancellation, and retry evidence |
| Android state | Draft recreation and Room reopen checks |
| Branded guide | Known-link matching tests and honest recommendation evaluation |
| Capstone | Persisted per-conversation style and isolated request tests |
| Advanced designs | Invalid structured-output cases and a duplicate-safe tool execution table |
| Limitations | Explicit live/model/device paths not yet verified |
A learner without provider access can complete deterministic work and record that live evaluation remains unverified. A completion tick must not be presented in a portfolio as proof of live testing. Revisit any answer that relies only on “the library handles it.” Explain the application policy that remains your responsibility.