# Example protocol: read model output by name, and check how it is stored

Illustrative resource from the Intuitionist Edu Project 01 demonstration
(https://intuitionist-edu.lampbotics.com/demo-project1/). Written for public
reading and adapted from guidance the research team recorded during a
hospitality-management replication. It is not the team's internal skill file,
and recording it is not evidence that it improves later work. That is what the
planned evaluation will test.

## The rule

1. Select every reported value by its label, never by its position. Print the
   labels beside the values before using either.
2. Find out how the software stores each kind of value. Some ordered-outcome
   models store the first cut point followed by the log of each gap between cut
   points (statsmodels 0.15.0 `OrderedModel` does this). Convert with the
   library's own function before labelling or testing.
3. Do not apply a conversion that the documentation does not call for. Values
   that are already cut points become wrong when converted again.
4. If the output does not say how values are stored, stop and find out. Do not
   guess, even when a guess would make the check pass.
5. When a check fails, open the line that computed it before reporting the
   failure. When it passes, keep a run in which it failed on a known error, or
   say that it has only ever passed.

## Why the team adopted it

Two agents accepted the same failed threshold check because they shared one
assumption about where the cut points sat and how they were stored. Their
agreement added confidence, not evidence. Review that opened the code, an
external audit, and questions from the original author and the investigator
changed the diagnosis and the extraction procedure.

## Where it applies

- Reading estimates from any statistical library.
- Reviewing a teammate's pass/fail summary or a saved true/false flag.

## Where it should be rejected or adapted

- The library already returns labelled cut points: skip step 2's conversion.
- The model has no cut points (for example, linear regression): the labelling
  habit still applies; the conversion does not.
- A different library or version stores values differently: the specific
  conversion does not carry over, only the habit of checking.

## Try it

`check.py` runs three hand-written readers on the synthetic cases in
`fixtures.json`. See `README.md`.
