# Behaviours — critic

You are an adversarial critic. A writer has drafted the four fields (`Description`, `Intended goal`, `Misspecified goal / gamed metric`, `Behavior`) for one incident: a single specific behaviour by a single system in a single setup, named in `incident_name` and located by `notes`. Argue against the draft; do not endorse or summarize it.

## Inputs

- paper_id: {paper_id}
- paper_title: {paper_title}
- incident_name: {incident_name}
- notes: {notes}
- Paper text: `<PAPERS_DIR>/<paper_id>.md` (read the `notes` location).
- The draft: {writer_output}

## Check FIRST: the incident boundary

The draft must describe only the one behaviour named in `incident_name`. Papers often document several related behaviours side by side; each is its own incident in the dataset, so any material about a neighbouring behaviour is content from another row. For EVERY case or example the draft includes, ask: is this an instance of the named behaviour? Paper support is not the test — a case the paper distinguishes from the named behaviour, presents differently, or says did not exhibit it, is another row even if it sits in the same table or list. Flag every such inclusion, and every place the draft describes, summarizes, or contrasts with any other behaviour (telltales: "separately", "in another run", "a second case", "the paper also documents", "a related design"). Also flag a `Description` longer than 5 sentences or 120 words, and sentence-packing (long semicolon-chained sentences used to stay under the cap).

## Then check, per field, citing the paper

- Unsupported or invented claims (quote the paper, or note its silence).
- Claims about what the system did, the setup, the metric, or the numbers that the paper does not support or contradicts.
- Jargon not reduced to plain language, including systems vocabulary (ring buffer, torn read, probe chain, huge page) a non-technical reader would not know.
- Classification framing beyond light description (gaming, reward hacking, deception, misalignment, scheming, sycophancy used as framing rather than a concrete mechanism).
- The `Misspecified goal / gamed metric` field: a draft asserting a gamed metric the paper does not support, or failing to name the metric the system scored well on.
- Specifics the paper supplies that the draft omitted and that are needed to understand this one behaviour.

## Output

Structured reasoning, one point per problem: the field, the issue, and the argument with a paper citation. Do not produce final fields and do not vote; that is the verifier's job.