JUDGE INSTRUCTIONS

# Judge instructions

You are an impartial senior code reviewer for a real npm package. Four TypeScript codebases (labelled `A`, `B`, `C`, `D`) implement the same task: a zero-runtime-dependency in-memory full-text search library with BM25 ranking, boolean/phrase/fuzzy queries, unicode tokenization, snippets, and serialization.

## Materials

- `/home/user/workspace/ts-search-eval-2026-07-03/brief/TASK.md` — the task brief
- `/home/user/workspace/ts-search-eval-2026-07-03/brief/RUBRIC.md` — the scoring rubric (10 dimensions, weighted, /100)
- `/home/user/workspace/ts-search-eval-2026-07-03/judging/machine-checks-blinded.md` — build/test/shared-suite pass status
- Your judge root: contains `A/`, `B/`, `C/`, `D/` subdirs (source, tests, bench, package.json, README) plus `bench-results.json` (per-letter benchmark data) and `metrics.json` (per-letter LOC and test counts)
- Some variants have a `process-notes/` directory with process artifacts (task contracts, plans, execution logs). Others do not. Score those on **substance**, not presence.

## Method

1. Read the task brief and rubric in full.
2. For each variant A/B/C/D:
   - Read every source file in `src/`.
   - Skim the tests to gauge test quality — not just count.
   - Read the README.
   - If `process-notes/` exists, read the completion-record, validation-contract, and decision-log to assess D7 substance.
3. Consult `bench-results.json` and `metrics.json` for the performance and size numbers.
4. Score each variant on 10 rubric dimensions, 0-10 integers each. Give a one-sentence concrete reason per dimension per variant.
5. Assign a verdict per variant: `merge_as_is` | `request_changes` | `reject`.

## Output

Write a single JSON file to the path passed in your task instructions with this schema: