Proposer Prompt

# Proposer Prompt Template

A copy-pasteable prompt that consumes the output of `term_similarity.py --canonical --json` and asks the LLM to produce **structured, actionable rename proposals** with confidence levels, evidence citations, and explicit escape hatches.

## When to Use

After running `term_similarity.py --json --canonical` on a codebase, paste the JSON output into the template below. The LLM produces a `proposals[]` array where each entry is one atomic, revertable decision.

## Design Principles

This template encodes 7 research-validated principles:

1. **Reasoning before verdict** (CoT improves calibration in 33/36 settings — arXiv 2505.14489)
2. **Categorical confidence** (HIGH/MEDIUM/LOW, not numerical — EMNLP 2023)
3. **Score legend with anchored examples** (prevents axis misalignment)
4. **Disagreement → flag rule** (multi-layer disagreement is structural honesty trigger)
5. **Hard gates from domain context** (co-occurrence, bounded context — overrides scores)
6. **Explicit escape hatch** (`flag_for_review` is high-status, not failure)
7. **Atomic proposals with required edits[].file/line** (vagueness is structurally impossible)

---

## The Template