# MRI-001: Methods & Evidence

This page is backed by the restored public evidence package at `mri/001/evidence/`, copied from the existing backup `mri/001.backup-20260817-1935/evidence/` without changing evidence contents.

## Primary evidence now surfaced

- `evidence/data/profile_a_sensitivity_results.csv`: full 109-row MRI Profile A ranking.
- `content/mri.js`: full 109-entry published BFCL ranking, joined into the public ranking table in `methods.html`.
- `evidence/data/profile_a_relevant_results.csv`: relevant Profile A component-score evidence.
- `evidence/data/bfcl_relevant_baseline_rows.csv`: relevant BFCL baseline component rows.
- `evidence/provenance/BFCL_DECISION_PROFILES.md`: exact work-profile weights and eligibility rules.
- `evidence/provenance/BFCL_PASS_FAIL_RULES.md`: pass, mixed, and fail conditions.
- `evidence/provenance/BFCL_BASELINE_RECONSTRUCTION_REPORT.md`: reconstruction formula and 109-model verification.
- `evidence/provenance/BFCL_CALCULATION_AND_MATERIALITY_AUDIT.md`: sensitivity checks and materiality audit.
- `evidence/pilot/03_RESULTS.csv`: 288 row-level center-pull pilot results.
- `evidence/pilot/04_MRI_METRICS.csv`: center-pull metrics.
- `evidence/pilot/05_TECHNICAL_REPORT.md` and `06_PLAIN_LANGUAGE_RESULT.md`: explicit PILOT FAIL result.
- `evidence/EVIDENCE_PACKAGE_MANIFEST.csv`, `evidence/provenance/BFCL_SNAPSHOT_MANIFEST.csv`, `evidence/code/`, and `evidence/mri-001-public-evidence.zip`: provenance and reproducibility files.

## Key public result

Profile A selects `Qwen3-14B (Prompt)`, published BFCL rank 47, as MRI rank 1. `Claude-Opus-4-5-20251101 (FC)`, the published BFCL rank 1 model, appears as MRI rank 28.

## Sensitivity result surfaced

The Profile A decision flip survives the full registered profile, cost removed, latency removed, and capability-only checks reported in `BFCL_CALCULATION_AND_MATERIALITY_AUDIT.md`.

## Pilot result surfaced

The center-pull pilot evidence is linked and summarized with the explicit pre-registered verdict: `PILOT FAIL`.

## AI-assisted research disclosure

MRI-001 used AI-assisted research, coding and analysis.

AI models, Python and Codex were used to help collect and organize evidence, implement calculations, process the model field, run sensitivity tests, preserve outputs, prepare tables and audit the analysis.

The research question, conceptual boundaries, test structure, interpretation, publication decisions and limits on what could be claimed remained human decisions.

Outputs that weakened the preferred interpretation were retained. This includes malformed responses, alternative winners under sensitivity tests and the failed center-pull pilot.

The evidence package is published so the distinction between the research judgment and the mechanical work can be inspected rather than taken on trust.