LLM-as-a-Verifier for DeepSeek Harness

Key features

• 1–20 instead of 1–5 — a finer scale separates close candidates better.
• compare can return a tie on an exactly zero margin
• Grading goes through the harness's own LLM service (ctx.llm) — same provider routing and auth as everything else, no direct API calls.
• Several repetitions (default 4) — averaging repeated grades reduces noise.
• Several criteria — small focused questions beat one big vague one.