THE ASSIGNMENT
One hour. One policy director. Four working products.
Construct a defensible dataset and report how income and under-five mortality differed across World Bank income groups in 2022. Decide what evidence is trustworthy, document any material data problem, and state what the results do—and do not—show.
The project supplies both a draft briefing file and documented course data. Those files may not agree. Codex has two roles: first, it reviews the draft without changing it; later, it constructs and runs a bounded pipeline. In both roles, you decide which claims and outputs survive.
code/02-health-agent-pipeline.R
outputs/02-health-analysis.csv
outputs/02-health-briefing-by-income.csv
BRIEFING.mdCodex may help produce and check the analysis. The release decision and policy language remain yours.
| Minutes | Decision |
|---|---|
| 00–08 | Specify the unit, key, and a briefing-changing failure. |
| 08–18 | Review the draft and independently verify one important finding in R. |
| 18–40 | Commission, observe, and inspect the construction pipeline. |
| 40–48 | Accept, revise, or reject the constructed dataset. |
| 48–60 | Build the grouped table and decide what to tell the director. |
DECISION GATES 1 + 2
Can we trust the draft?
Define the intended economy–year observation before inspecting the records. Then ask Codex for a read-only review. Select the finding most likely to change a count, median, plot, or conclusion. Then write and run a small R check that tests it directly against the project files. A persuasive explanation is not evidence that the finding is correct.
What record problem could change the director’s briefing?
Require files, variables, economy–years, and observed values.
Use R to confirm, contradict, or qualify it from the records.
Accept, revise, or reject the agent’s claim.
A confident agent, a recent modification date, or a promising filename does not make a file authoritative. Ask which project documentation and provenance support the choice.
DECISION GATE 3
Can we trust the constructed dataset?
Commission Codex to choose and justify the input, create
code/02-health-agent-pipeline.R, run it, and save
outputs/02-health-analysis.csv. Observe the work:
which files were read, which actions required permission, which
source was selected, and which code actually ran?
| Judgment | Question | What it does not prove |
|---|---|---|
| Execution | Did the script run and save the requested file? | That the file represents the intended evidence. |
| Output evidence | Does the file have the required shape and match documented records? | That the rule used to produce it was defensible. |
| Method | Would the result survive a change in row order or unsupported source? | That the policy interpretation follows. |
Run the independent audit from
code/lab-2-starter.R, inspect the generated script,
and make an explicit accept, revise, or reject decision. Correct
output by accident is not a defensible method.
DECISION GATE 4
Can we trust the briefing?
Add the grouped calculation to the accepted pipeline. For 2022,
compare n() with
n_distinct(iso3c), explain why those quantities can
differ, and decide which belongs in the policy briefing. Then
report one row per current income group with the number of
economies, median GDP per capita (PPP), and median under-five
mortality.
One row represents a current income group in 2022.
Use unweighted counts and medians after the dataset survives review.
Codex checks the saved table without rewriting your code.
Give one numerical conclusion and refuse one unsupported claim.
Would you send this briefing to the policy director? Identify the strongest evidence supporting your decision, one failure you ruled out, one numerical conclusion the table supports, and one statement you would refuse to make.