# Mei's Revenue-Cycle Prep for AI — Schema

Flat CSV in `mei-revenue-cycle.csv`. One row per CFO question.

## Columns

| Column | Type | Description |
| --- | --- | --- |
| `id` | integer | Question id (1..5) in the order Mei asked. |
| `question` | text (quoted) | The CFO-shaped question, verbatim as asked. |
| `before_verdict` | text | One of `worked`, `ambiguous`, `wrong`, `refused`. |
| `before_notes` | text (quoted) | What the before-state outcome was and why. |
| `after_verdict` | text | Same allowed set as before_verdict. |
| `after_notes` | text (quoted) | What the after-state outcome was and why. |
| `prep_component` | text (quoted) | Which Prep-for-AI component closed the gap (empty for questions that already worked before Prep). |

The `question`, `before_notes`, `after_notes`, and `prep_component` fields are CSV-safe quoted strings.

## Row inventory

- 5 rows — one per CFO question.

Total: 5 rows.

## Deliberate design choices

The five-question sample size is small on purpose: Chapter 11’s point is that on a healthy hand-built model, the *distribution* of failures is what matters (some work, some are ambiguous, some are wrong, some are unanswerable) — not the raw count. Five questions is the smallest set that produces this full failure taxonomy.

The `prep_component` column names the specific fix per question so an instructor teaching the Prep-for-AI pass can walk each question’s fix in isolation. Composite values (e.g., `verified_answers + ai_data_schema`) capture the questions where multiple components contribute.

## License

Released for teaching use alongside *The Defensible Decision*. Use freely.