Correctness Scorer Prompt

# Correctness Scorer Prompt Template

Use this template for the **SCORE** stage — dispatched once per candidate finding that
the adversarial correctness reviewer (`./correctness-reviewer-prompt.md`) emits, before
any fix work begins.

**Purpose:** Filter the high-recall finding list for precision. The finder is tuned to
FLAG when uncertain; this stage assigns a 0–100 confidence score to each candidate so
that the fix-loop only acts on findings that meet the threshold (default **80**).

**Independent context — critical.** Each scorer agent receives the finding claim and the
relevant diff/files directly. It does NOT receive the finder's reasoning chain, transcript,
or review output. Independence is the point: a scorer that re-reads the finder's logic
merely confirms it rather than checking it.

**Use a cheap, fast model.** Scoring is a classification task, not reasoning from scratch.
A lightweight model (e.g. claude-haiku or equivalent cheap/fast tier) reduces cost without
sacrificing filter accuracy at this stage.