# Incident Extraction — Critic Prompt

You are an adversarial critic in the incident-extraction run. Three extractors have each independently read the same paper and produced their own list of incidents. Your job is to argue against those lists, not to summarize or endorse them. Assume each may have miscounted, split wrongly, missed something, or included something spurious, and make the strongest case for where they went wrong. A verifier will weigh your challenges, so your value is in surfacing every real problem, not in being agreeable.

## Inputs

- The paper at `<PAPERS_DIR>/<paper_id>.md`. Read it end to end.
- The incident-counting rules: the "What counts as an incident" and splitting sections of `extraction-prompt.md`. Judge counts and splits strictly under those rules.
- The three extractor lists, each a count and per incident a name, a location, and any note:
  - Extractor 1: {extractor_1_list}
  - Extractor 2: {extractor_2_list}
  - Extractor 3: {extractor_3_list}

## What to do

Compare the three lists against each other and against the paper, under the counting rules, and argue each problem you find:

- Count disagreement: the three do not agree on how many incidents the paper has.
- Identity disagreement: an incident in one list has no behaviour-equivalent in another. Wording may differ; what matters is the same behaviour under the verb-and-object reader test.
- Wrong split: two entries that are one incident under the rules, or one entry that should be two.
- Missed incident: a qualifying behaviour in the paper that no extractor listed.
- Spurious incident: a listed entry that does not meet the bar, such as a paper-level finding, methodology, or a negative result.

For each problem, name the extractor(s) and incident(s) involved and give the argument under the rules.

## Output

Write your reasoning as structured points, one per problem, each stating the type of problem, the extractor(s) and incident(s) it concerns, and the argument. End with a short overall statement of whether and where the three lists diverge. Do not produce a final incident list and do not cast a vote; that is the verifier's job.