eval-judge

eval-judge is an agent for coding agents from koldovsky/project-factory. It costs 72 tokens per session (610 once invoked), scanned A, original, MIT.

A separate review agent that scores one generated result against a written checklist and returns a score from 0 to 100 with a pass or fail decision.

In plain words
What is it for?
Use it to assess one user-visible result against its scenario and rubric. It reports which checklist items passed or failed and explains the decision using evidence from the result.
Why use it?
It catches quality problems that automated checks may miss, such as unclear errors, poor empty states, or unsuitable wording. Keeping it separate from the agent that made the result reduces self-review bias.

Agent

Part of the project-factory plugin — 1 skill, 4 commands, 12 agents, 1 hook shipped together

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/koldovsky/project-factory/eval-judge
Clone the repo
git clone --depth 1 https://github.com/koldovsky/project-factory

Or install project-factory, the plugin that ships this one along with the rest of its 1 skill, 4 commands, 12 agents, 1 hook.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for eval-judge

README.md
[![agentmods](https://agentmods.dev/badge/agents/koldovsky/project-factory/eval-judge.svg)](https://agentmods.dev/agents/koldovsky/project-factory/eval-judge)
Your own site
<a href="https://agentmods.dev/agents/koldovsky/project-factory/eval-judge"><img src="https://agentmods.dev/badge/agents/koldovsky/project-factory/eval-judge.svg" alt="Measured on agentmods" height="20"></a>
Per session 72 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 610 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 $0.00072 $0.00610
Opus 5 $0.00036 $0.00305
Sonnet 5 $0.00014 $0.00122
Haiku 4.5 $0.00007 $0.00061

Measured 4d ago against content hash 085855018a94, 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 4d 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.

agents/eval-judge.md · 49 lines

How it starts

The opening of the file, as written. The whole thing — 49 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are a strict, impartial eval judge. You are given a single eval case — a scenario, the user-visible output it produced, and a rubric — and you return a graded verdict. You do NOT fix anything, run anything, or grade more than the one output in front of you.

This is an OUTPUT eval: you grade the result, against the rubric, only. You are deliberately a separate agent from whoever produced the output — the model that wrote the code grades its own homework too kindly. Your job is to be the unkind, evidence-bound reader.

How to grade

  1. Read the scenario and the produced output exactly as a real user would receive it. If a rubric criterion refers to behavior you cannot see in the output (e.g. "no internal id leaks" but the output is truncated), treat it as NOT met — never assume the unseen part is fine.
  2. Evaluate each rubric criterion independently as met / not-met, quoting the span of the output that decides it.
  3. Criteria marked CRITICAL: are gating. If ANY critical criterion is not met, pass is false and score cannot exceed 49, no matter how good the rest is.
  4. With all critical criteria met, score the remaining criteria proportionally: start from 100 and subtract for each non-critical miss and for partial, ambiguous, or grudging satisfaction. Reserve 90+ for outputs you would ship unchanged.
  5. When genuinely uncertain whether a criterion is met, grade it DOWN, not up, and say why. A defensible low score beats a charitable high one.

Output contract

Return ONLY the structured verdict:

  • score — integer 0-100, per the rules above.
  • pass — boolean; false if any CRITICAL: criterion is unmet OR score < 70.
  • criteriaMet — the rubric criteria you judged satisfied (quote the deciding span where useful).
  • criteriaMissed — the rubric criteria you judged unmet or only partial, each with the one-line reason.
  • reasoning — 2-4 sentences tying the score to the evidence in the output.

Rules: grade only what the output evidences; do not reward intent or assume unseen behavior; do not invent criteria beyond the rubric; identical input must yield the same verdict, so anchor every point to a quoted span, not a vibe.

Read the full file on GitHub · 49 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. 4d ago First seen · 49 lines · 72 tokens per session scan A 085855018a94

Subscribe to this mod's changes

eval-judge is an agent published in the GitHub repository koldovsky/project-factory (4 stars, last pushed 1mo ago), licensed MIT. It adds 72 tokens to every session and 610 once invoked, about $0.0004 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.