r-data-validation

name	r-data-validation
description	Build auditable R data-validation pipelines with pointblank, branching-aware missingness rules, row-level extracts, and testable logic checks. Use for data-quality validators and business or clinical rule screening.
argument-hint	["validation target"]

R Data Validation

Turn data-quality rules into explicit, reviewable validation plans that separate missingness, business logic, and downstream action.

Quick start

1. Read references/pointblank-workflow.md.
2. Write down the validation contract before coding: required fields, branch-open fields, range rules, and cross-field logic rules.
3. Build one pointblank agent per source table.
4. Add missingness rules first, then logic and consistency rules.
5. Run interrogate(extract_failed = TRUE) and return the interrogated agent or the failed-row extracts.
6. Add targeted testthat coverage for happy path, branch-open failure, branch-closed pass, and invalid input.

Workflow

1. Define the validation contract

• Separate always-required fields from branch-dependent fields.
• Decide which failures should block analysis, which should warn, and which should feed a correction queue.
• Give every rule a stable step_id so failed rows remain traceable.

2. Build missingness checks first

• Use col_vals_not_null() for globally required fields.
• Use preconditions = to subset to branch-open rows before checking child fields.
• Keep missingness rules independent from logic rules so failures stay interpretable.

3. Add logic and integrity checks second

• Use type-specific validators when possible.
• Use col_vals_expr(~ ...) for cross-field logic.
• Decide na_pass intentionally based on whether missingness is already handled elsewhere.

4. Interrogate and extract failures

• Run interrogate(extract_failed = TRUE).
• Review all_passed(agent), agent[["validation_set"]], and get_data_extracts(agent).
• Return the interrogated agent when downstream code needs both the rule summary and row-level extracts.

5. Keep validator functions small

• Accept one data frame unless the task truly needs a broader interface.
• Return one stable validation object.
• Keep helper logic deterministic so test assertions remain durable.

Coordination notes

• Reach for /r-stats when the issue is mainly statistical cleaning or analysis design rather than validation.
• Reach for /r-advanced when repeated repo history shows the same validation or coercion failures coming back.

Resource

• references/pointblank-workflow.md