PreviousNext
AI-assisted engineering/Guardrails

Screen survey themes before a manager sees them

Agent buildEasy10 minAnthropic style

Your task

Your developer survey promised anonymity, and some free-text answers name names or describe situations only one person could have lived. An agent groups the answers into themes, and a screen decides what may reach a manager.

Write screen_survey_themes(theme_names, respondent_counts, has_identifiable_detail, k_floor): the screen every theme passes through.

  • theme_names (list of str): the theme names.
  • respondent_counts (list of int): how many different people raised each theme.
  • has_identifiable_detail (list of bool): the theme's quotes contain something that could identify a person.
  • k_floor (int): the fewest respondents a theme needs before it can be shown. This is a k-anonymity floor: a theme from fewer people could point to one of them.
  • Return one verdict per theme, in order, written "<name>:<verdict>".

Rules:

  1. Index i in each list is one theme. Stop at the end of the shortest list.
  2. A floor below 1 counts as 1.
  3. Trim the name. An empty name becomes unnamed.
  4. Fewer respondents than the floor: "suppress".
  5. Otherwise, identifiable detail: "paraphrase".
  6. Otherwise: "release".

Examples

screen_survey_themes(theme_names: list[str], respondent_counts: list[int], has_identifiable_detail: list[bool], k_floor: int) -> list[str]
Example 1
Inputscreen_survey_themes(["flaky ci", "laptop specs", "review turnaround"], [12, 7, 3], [False, False, False], 5)
Output["flaky ci:release", "laptop specs:release", "review turnaround:suppress"]
Example 2
Inputscreen_survey_themes(["oncall load", "manager named in quote"], [9, 6], [False, True], 5)
Output["oncall load:release", "manager named in quote:paraphrase"]
Example 3
Inputscreen_survey_themes([], [], [], 5)
Output[]