# Rubric grader — neutral prompt (works on any agent/LLM, or by hand)

> This is the PORTABLE piece of the rubric layer: instructions for ANY runner —
> Claude Code, Codex, Gemini CLI, Cursor, a `curl` to any API, or a human.
> The only coupling with the method is the output JSON CONTRACT, which
> `qa_ledger.py rubric-ingest` validates and ingests. Who emits the JSON is irrelevant.

## Instructions for the grader

You are an evaluator with an ISOLATED context. You read ONLY two things:

1. `RUBRIC.md` — the criteria (weighted RB-nn, with anchors), the negative
   criteria (RB-NEG-nn) and the threshold.
2. The change's diff (or the changed files).

You do NOT read the reasoning of whoever made the change, nor its description, nor its PR body —
your value is precisely in having no attachment to how the result was produced.

For each rubric criterion:

- Emit a `pass` or `fail` verdict.
- **Evidence is mandatory for every verdict that affects the score**
  (a `pass` on a positive criterion; a `fail` on a negative one): a concrete
  `file:line` citation + a snippet. Without evidence, the verdict does NOT score — the
  engine discards it and lists it as unsupported.
- Use the anchors as calibration: if the code looks more like the anchor-fail than
  the anchor-pass, it is `fail`. When in doubt, `fail` — the optimistic bias is the
  failure mode this contract exists to counter.
- In `note`, one line of justification (what you saw, not what you assume).

## The output contract (the only thing that matters)

Write a single JSON:

```json
{
  "criteria": [
    {"id": "RB-01", "verdict": "pass",
     "evidence": "src/client.py:42 — client.get(url, timeout=5) with retry",
     "note": "all external calls have a timeout and backoff"},
    {"id": "RB-02", "verdict": "fail",
     "evidence": "src/NewModule.py:1 — camelCase in a snake_case repo",
     "note": "does not follow the neighboring modules' convention"},
    {"id": "RB-NEG-01", "verdict": "pass", "evidence": "", "note": "does not appear"}
  ]
}
```

- `id`: must exist in `RUBRIC.md` (unknown IDs = ingest error).
- `verdict`: `pass` | `fail`. On the NEGATIVE ones, `pass` = the forbidden practice does NOT
  appear; `fail` = it appears (and subtracts its weight from the score, with evidence).
- Criteria you do not evaluate count as `fail` (not evaluated is not passed).

## How it is ingested (the operator or the loop runs it, not you)

```bash
python3 <path>/qa_ledger.py rubric-ingest --repo <REPO> --report grader.json \
  [--rubric RUBRIC.md] [--gate]
```

Advisory by default; `--gate` (or `defaults.rubric.gate: true` in the config — the
human's declaration) turns a score below the threshold into a gated record:
it blocks convergence and caps readiness ≤65 through the ledger's existing machinery.