# Build record and authorship

Miguel Rodriguez selected the project, defined the portfolio and hiring objective,
and authorized implementation, public website integration, and private GitHub
backup. Implementation, documentation, and automated verification were produced
with Codex assistance. No claim is made that Miguel manually wrote every line or
has independently completed the interview walkthrough yet.

## Decisions in the first implementation

- Use a real PostgreSQL queue and a separate HTTP/worker process before adding AI.
- Keep the unsafe worker explicit so the failure can be reproduced.
- Check full record identity and values, including refunds and tenant-scoped IDs.
- Use independent expected records created before fault injection.
- Branch actual database state and verify the untouched source.
- Measure one local model decision against a strong deterministic baseline.
- Preserve unsuccessful model decisions and timeouts.
- Keep public interaction as labelled replay; preserve the established site brand.

## Ownership walkthrough to complete with Miguel

1. Explain why a restart duplicates a write at the broken acknowledgement boundary.
2. Show why an order ID is insufficient for deduplication.
3. Change a business invariant and identify which tests should fail.
4. Explain a model failure without blaming it on wording alone.
5. Reproduce a recorded recovery and discuss the limits of the evidence.

These exercises and later resume/application work are explicitly deferred. They
should inform the eventual interview narrative; they are not completed user-test
or hiring outcomes.