# REGRESSION SET
*Three documents you have already done by hand, with the answers you know are
right sitting beside them. Twenty minutes to assemble. It is the difference
between tinkering and engineering.*

## How to use it

1. Keep the three source documents and the three correct outputs together.
2. Change one rule.
3. Re-run all three.
4. Compare field by field. A change that fixes one and breaks another is the
   global-rules-local-defects collision, caught before it ships.

## Variant A — the sample in `01_sample_scope.md`
Contains: absence (power), conflict (floor vs system of record), inference
(A-1101), multi-record (three ports), a Consequence (cabinets).

## Variant B — nothing to interpret
A submission where the narrative repeats what the fields already say. Correct
behaviour: every field MOVED, no NEEDS EYES, no consequences, one record.
**Catches:** a prompt that invents interpretation work to look useful.

## Variant C — narrative only
A submission where the form half is almost empty and everything is in prose.
Correct behaviour: Move fields NOT STATED rather than back-filled from the
narrative; Interpret fields all carry quotes.
**Catches:** a prompt that quietly promotes an interpreted value into a Move
field, which is the failure that looks most like success.

## Scoring

Count only three things per run:

- fields whose value changed from the known-correct answer
- fields whose STATUS changed
- consequences gained or lost

A run is a pass when all three are zero. Record the count every time, so
"it got better" stops being a feeling.