# Miru projected assessment: data and reproduction

This bundle includes the original independent-review-only scenario and the expanded author-side plus independent-review scenario. Both use the same category anchors, weights, penalties, cap and floor. The expanded scenario uses the newly published author-side evidence in `author-side/`; it does not relabel the original 53.8 result.

This bundle accompanies the projected assessment on https://descant.run/case-studies/miru#projected-assessment. It contains public GitHub records, a documented classification of review findings, the projected rubric inputs and a constructed scoring scenario. No alternative implementation was run or evaluated. The migration was stopped after the customer demonstration; the projected score assumes completion of the remaining agreed work.

## Reproduce offline

Download and extract `evidence.zip`. Python 3 and its standard library are sufficient; no login, network access or installed packages are needed.

```sh
python3 -I reproduce-findings.py
python3 -I recompute.py
python3 -I author-side/verify.py
python3 -I recompute-full-review.py
```

All four commands are read-only by default and resolve inputs relative to their own directory. The first reconstructs the finding rows from the frozen public comments and verifies the published JSON and CSV exports. The second checks those rows against the scoring inputs and reproduces every scenario using exact fractions. The third checks the author-side export. The fourth joins every author record to its disclosed category assignment, checks that the original rubric is unchanged, and reproduces the expanded scenario plus the plan-excluded sensitivity. A mismatch stops with an error.

## Files

- `method.md`: full methodology, assumptions, worked example and interpretation limits.
- `assessment.json`: all 15 category anchors and weights, ordered category rules, severity deductions and 25 calculated group/scenario results.
- `categories.csv` and `sensitivities.csv`: inspectable score tables.
- `counts-public-source.json`: 114 PR records and 622 public inline comments, retrieved November 2, 2026.
- `counts-public-findings.json` / `.csv`: all 289 review roots, with inclusion flags, severity, category, public links, final author dispositions and explicit fix-commit evidence.
- `counts-public-233.json` / `.csv`: the 233 retained-correction threads on 70 merged PRs.
- `counts-summary.json`: denominators, severity/category totals, cohort reconciliation and unmerged exclusions.
- `counts-findings.md`: extraction definitions and the distinction between final disposition and retained correction.
- `counts-disputed-resolution.json`, `counts-disputed-commits.json` and the four `counts-public-code-*.rb` snapshots: public evidence for the earlier correction retained after its final re-raise was set aside.
- `counts-source-category-map.json` and `counts-source-comparison.json`: per-root assignments and their agreement with the independently reconstructed census.
- `counts-public-manifest.json`: file hashes and provenance.

## Cohort distinction

The calculation uses 233 retained corrections: 232 threads with a final fixed disposition plus one earlier correction that remained in the merged code after a final ruling declined further changes as already correct. This broader cohort reconciles with the case study's 196 non-Minor implementation findings as 196 + 24 Minor implementation findings + 12 design findings + 1 retained earlier correction = 233. The totals must not be added together.

The source corpus and scorecard cover the same recorded work. Of 103 merged PRs, 99 are implementation, two design and two human preparation/scope. The score calculation uses only retained corrections on merged PRs; six retained corrections on unmerged PRs are excluded. A root thread is not necessarily a unique defect. Recorded fix evidence is not an individual patch audit of every finding.

Evaluated repository revision: `d72e6ce1ea5bf9c0dae781ccb5b290781025304a`. Baseline: `13e46ed8ca21ea923c653ecb48c7485da4072ffb` in https://github.com/Homie-Improvement/miru.

For the expanded comparison, start with `full-review-method.md`, `full-review-assessment.json` and `author-category-assignments.json`. `assessment.json` and `method.md` preserve the original independent-review-only scenario. The two `full-review-*.csv` tables contain the expanded results; the original CSV tables remain unchanged.

The scripts reproduce this snapshot, not an empirical estimate of another system's performance. The rubric, deductions, path mapping and process assumptions remain judgments even when the arithmetic reproduces exactly. No private session transcripts, credentials or private assessment PDFs are included.

`counts-fetch-public.py` is an optional network refresh utility using the installed GitHub CLI. It replaces the frozen input; run it only in a separate copy if intentionally creating a new snapshot. `reproduce-findings.py --write` explicitly regenerates derived finding exports. Neither is needed to verify the published bundle.

The included Miru code snapshots are reproduced under the MIT license in `miru-LICENSE.txt`; public source and commit links accompany the evidence.