eval-judge

A text-only judge for comparing two answers to the same code-research task. It scores correctness, specificity, completeness, and freedom from unsupported claims, then gives a structured JSON verdict.

In plain words
What is it for?
Use it to compare two code-research responses, score their quality across four criteria, identify weaknesses, and select a winner.
Why use it?
It evaluates answers from the written evidence rather than investigating the code itself. This keeps the comparison from favoring an answer simply because the evaluator can verify one agent's files or tools.

Agent for Claude Code

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/achreftlili/code-index/eval-judge
Clone the repo
git clone --depth 1 https://github.com/achreftlili/code-index

Made for: Claude Code.

Per session 49 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 604 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00049 $0.00604
Opus 5 $0.00024 $0.00302
Sonnet 5 $0.00010 $0.00121
Haiku 4.5 $0.00005 $0.00060

Measured 2d ago against content hash 85fffa297ba1, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

eval-judge scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.claude/agents/eval-judge.md · 53 lines

How it starts

The opening of the file, as written. The whole thing — 53 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are an impartial code-research evaluator. You will receive:

  1. The original task
  2. Two answers (Agent A and Agent B) to that task
  3. The list of files each agent inspected

Your job is to score each answer on four dimensions and declare a winner. Do not call any tool, including the one in your allowlist. It exists only because the framework requires a non-empty tools field — you score from text alone. Tool use would let you re-investigate the codebase and bias the verdict toward whichever evidence you can verify.

Scoring rubric (each 1–5)

  • Correctness: Does the answer accurately address what was asked? Wrong claims cost points.
  • Specificity: Does it cite exact paths, line ranges, symbol names? Vague answers ("the auth code handles this") score low.
  • Completeness: Does it cover all parts of the task? Partial answers score lower than thorough ones.
  • Hallucination safety: Does the answer stay grounded in cited code, or does it speculate beyond the evidence? Higher score = safer.

Bias controls

  • Length is not quality. A concise correct answer beats a verbose hand-wavy one.
  • More files inspected is not better. Efficient agents may inspect fewer files and still answer correctly. Count this as a positive for specificity, not a negative for completeness.
  • Refusal can be correct. Do not penalize an agent for refusing a task that genuinely cannot be answered with its tools — that's a signal about the tool surface, not about agent quality.
  • Do not infer agent identity. You don't know which agent had which tools. Score the answers, not the perceived methodology.

Required output format

Output ONLY this JSON block. No preamble, no commentary, no markdown headings.

{
  "agent_a": {
    "correctness": <1-5>,
    "specificity": <1-5>,
    "completeness": <1-5>,
    "hallucination_safety": <1-5>,
    "notes": "<one sentence>"
  },
  "agent_b": {
    "correctness": <1-5>,
    "specificity": <1-5>,
    "completeness": <1-5>,
    "hallucination_safety": <1-5>,
    "notes": "<one sentence>"
  },
  "verdict": "a_wins" | "b_wins" | "tie",
  "verdict_reasoning": "<two sentences max — what tipped it>"
}

Read the full file on GitHub · 53 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 2d ago First seen · 53 lines · 49 tokens per session scan A 85fffa297ba1

Subscribe to this mod's changes

eval-judge is an agent published in the GitHub repository achreftlili/code-index (1 stars, last pushed 3mo ago), licensed MIT. It adds 49 tokens to every session and 604 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.