# P539 final submission checklist

Use this list before you submit Milestone 5, and when you review a classmate's project at the fourth exercise review. The required items are on the course assignments page: <https://p539.jima.me/assignments/#milestone-5>

Exercise numbers refer to the required text.

## The project runs (Exercise 10.1.2)

- [ ] A README explains what the project does, where the data come from, and the order in which to run the steps.
- [ ] The environment is pinned to exact package versions, and the language version and operating system are recorded with it.
- [ ] Every stochastic step is seeded, and the seeds are recorded. The analysis was run under at least three seeds and the spread is reported.
- [ ] The data snapshot the analysis reads is saved as a dated file, with the query that produced it.
- [ ] The README states the Gundersen reproducibility type (R1 to R4) the project reaches and what the next one still needs.
- [ ] A classmate has cloned or unpacked the project and tried to run it. What they could not reproduce is noted.

## The AI-assisted coding can be audited (Exercise 10.2)

- [ ] **Knowledge.** The promptbook under `kb/codebooks/` holds the full instrument: system prompt, prompt template, examples, output schema, model identifier, and settings.
- [ ] **Memory.** `memory/PROMPTLOG.md` records each substantive prompt revision and why it was made. `memory/runs/` has a dated note for each coding run, naming the promptbook version it used.
- [ ] **Rules.** `AGENTS.md` states what every coding run must record and what must be validated before output enters the analysis.
- [ ] The validation report gives the sample size, the agreement metric, results by class or subgroup, and the cases where the model failed.
- [ ] The prompt-and-output pairs are saved, so a reader can check labels the model can no longer be made to reproduce.

## The artifact works for its audience (Exercise 10.3.4)

- [ ] The audience is named, with the three or four questions they would ask.
- [ ] At least three coordinated views answer related questions about the same records. It is more than a folder of figures.
- [ ] A short methods note documents the choices behind each view: filters, exclusions, aggregations, and the subset each number is computed on.
- [ ] A static fallback exists for readers who cannot use the interactive version, and the artifact can be navigated by keyboard.
- [ ] A maintenance note says what could break the artifact in three to five years, and separates the durable analysis from the disposable interface.

## Uncertainty is shown (Exercise 10.3.5)

- [ ] One result is shown with its uncertainty, estimated from at least three seeds or three specifications, or from an analytical or bootstrap interval.
- [ ] The caption states what the uncertainty represents and what it does not cover.
- [ ] If the result depends on labels a model assigned, a note says that the model's own confidence is not calibrated and that accuracy against human coding is the standard.

## The ethics review is done (Chapter 11)

- [ ] It says which uses of the corpus preserve the norms under which its records were produced.
- [ ] It reports whose cases the measurement errors fell on, using the marginal-cases audit from Milestone 3.
- [ ] It includes an AI-disclosure paragraph that points a reader to a record they can check.
- [ ] It says where accountability sits when an AI-coded variable is wrong, and which project files let someone reconstruct what happened.

## Before you upload

- [ ] One PDF report contains the written pieces in the order the assignments page lists them, with an AI-use disclosure at the end.
- [ ] The project folder is a zip archive or a link to a specific commit, not a folder that will keep changing.
- [ ] No credentials, API keys, or restricted material are in the folder.
- [ ] The slides or link for the presentation, and the final reflection, are ready to submit to their own Canvas assignments.