# Pilot File Map

This folder contains the smallest complete set found for the failed authority-pull pilot. Evidence values and raw responses are preserved; machine-local absolute paths were replaced with package-relative references.

## Files

| Copied file | Published/source reference | What it contains |
| --- | --- | --- |
| `run_mri_rank_before_work_pilot.py` | `pilot/run_mri_rank_before_work_pilot.py` | Source for the pilot conditions, prompts, randomized trials, response parsing, metric calculations, and verdict generation. |
| `00_PREREGISTRATION.md` | `mri_rank_before_work/00_PREREGISTRATION.md` | Pilot title, core question, scenario definitions, conditions, decision-agent plan, measures, and pass/mixed/fail rule. |
| `01_EXPERIMENTAL_CARDS.md` | `mri_rank_before_work/01_EXPERIMENTAL_CARDS.md` | Human-readable experimental cards, including EVIDENCE FIRST, CENTER FIRST, and MRI ORDER condition wording. |
| `02_RAW_RESPONSES.md` | `mri_rank_before_work/02_RAW_RESPONSES.md` | Preserved raw model/judge responses. |
| `03_RESULTS.csv` | `mri_rank_before_work/03_RESULTS.csv` | Row-level trial records for Profile A and Profile C, including selected model, preferred model, decision regret, citations, changes, and parsing/compliance errors. |
| `04_MRI_METRICS.csv` | `mri_rank_before_work/04_MRI_METRICS.csv` | Calculated center pull, task-fit accuracy, decision regret, variation suppression, correction capacity, and cross-family metrics. |
| `05_TECHNICAL_REPORT.md` | `mri_rank_before_work/05_TECHNICAL_REPORT.md` | Technical summary and generated final verdict. |
| `06_PLAIN_LANGUAGE_RESULT.md` | `mri_rank_before_work/06_PLAIN_LANGUAGE_RESULT.md` | Plain-language result and generated final verdict. |

## Run Counts

- Planned runs: 288 decision records (4 judge families x 2 scenarios x 3 presentation conditions x 12 trials).
- Completed runs recorded: 288 of 288 row-level records.
- Failed or malformed runs: 139 records with parsing or compliance errors.

## Formulas

- Center pull: `100 * (CENTER_FIRST aggregate-winner selection rate - EVIDENCE_FIRST aggregate-winner selection rate)`.
- Decision regret: `max(0, best frozen scenario-specific score - selected option frozen scenario-specific score)`.
- Task-fit accuracy: selection rate of the frozen scenario-specific preferred model.
- MRI ORDER correction capacity: frequency and direction of final-selection changes after overall-rank disclosure.

## Verdict Location

- The final verdict is generated by `verdict(metrics)` and written by `write_outputs(...)` in `pilot/run_mri_rank_before_work_pilot.py`.
- `PILOT FAIL` is present in `05_TECHNICAL_REPORT.md`: YES.

## Missing Pilot Items

- Complete pilot prompts and row-level results were found.
- The raw response file includes malformed or non-compliant responses as recorded in `03_RESULTS.csv`; these are preserved, not repaired.