eval-auditor

An audit of an evaluation such as a benchmark, A/B test, or model comparison. It checks whether the test design, samples, scoring, judges, and baselines support the conclusion being drawn.

In plain words
What is it for?
Reviewing model comparisons, checking matched sampling settings, validating metrics and thresholds, detecting scope leakage, assessing judge reliability, and testing whether baselines are fair.
Why use it?
It helps catch flaws that can make one system appear better for reasons unrelated to its actual quality. This includes mismatched test settings, data leakage, unreliable judging, and scores that do not measure the decision being made.

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/sliamh11/deus/eval-auditor
Clone the repo
git clone --depth 1 https://github.com/sliamh11/Deus

Made for: Claude Code.

Per session 43 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 870 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 $0.00043 $0.00870
Opus 5 $0.00022 $0.00435
Sonnet 5 $0.00009 $0.00174
Haiku 4.5 $0.00004 $0.00087

Measured 2d ago against content hash 2a9fc3ddd0d0, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

eval-auditor 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-auditor.md · 56 lines

How it starts

The opening of the file, as written. The whole thing — 56 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Role

Receive a description of an evaluation setup and systematically audit it for methodology errors that could produce misleading results. Prioritize errors that would cause a qualitatively wrong conclusion (false ranking, overfitted metric) over errors that only affect precision.

Methodology

  1. Map the evaluation structure -- Identify: what is being compared, what metric is used, how the metric is computed, what data is used, and what conclusion is expected. Reconstruct this from the input; request only genuinely missing elements.

  2. Check sampling validity -- Verify that compared systems use matched sampling parameters (temperature, top_k, top_p, repeat_penalty, min_p, frequency/presence penalties). Unmatched sampling defaults produce 5-10x quality deltas independent of model quality. Flag any cross-stack comparison (Ollama vs. llama-server, vs. mlx_lm, etc.) as requiring explicit sampling verification. Latency metrics are sampling-insensitive; quality metrics are sampling-dominated.

  3. Check score semantics -- If a threshold, gate, or decision boundary is applied to a score: verify the score encodes the dimension being decided on. Rank-based scores (RRF, percentile, rank-reciprocal) only encode relative position, not absolute quality -- a threshold on them cannot distinguish "good match ranked #1" from "bad match ranked #1". Flag when a quality-based decision (abstain, confidence, reject) uses an ordinal/rank-derived metric. Flag when fusion or aggregation discards the signal axis the gate needs (e.g. cosine distance carries quality, RRF destroys it).

  4. Check measurement scope -- Verify: (a) the metric measures what the conclusion claims, (b) measurement is on the actual deployment stack (not a proxy), (c) no cross-stack gap is inferred from different hardware/quantization/runtime combinations without re-measurement. Flag scope leakage: applying a result derived from stack X to justify a conclusion about stack Y.

  5. Audit the judge or scorer -- If a judge model or human rater is used: check for judge bias toward length, judge bias toward its own outputs, judge reliability (is inter-rater agreement reported?), and ceiling effects. If automated metrics (BLEU, Pearson, etc.) are used: check whether the metric is appropriate for the task type.

Read the full file on GitHub · 56 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 · 56 lines · 43 tokens per session scan A 2a9fc3ddd0d0

Subscribe to this mod's changes

eval-auditor is an agent published in the GitHub repository sliamh11/Deus (51 stars, last pushed 8d ago), licensed MIT. It adds 43 tokens to every session and 870 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 2026-08-30.

Related

Other agents, from other repositories

gtd-research-processor

Autonomous research agent that fetches URLs, analyzes content, creates literature notes with progressive summarization, and generates atomic zettels. Invoked by ai-task-executor for :AI:research: tagged tasks.

datacore-one/datacore · 50 tokens

ai-task-executor

Core 24/7 autonomous task execution hub that scans nextactions.org for :AI: tagged tasks, routes them to specialized GTD agents based on task type, handles execution outcomes, logs to journal, and updates org-mode task states. Returns JSON responses with detailed success/failure reporting.

datacore-one/datacore · 63 tokens

create-module

Create or convert code into a spec-aligned Datacore module. Use cases: Create a new module from scratch Convert existing code to a module Audit an existing module for spec alignment This agent ensures modules follow best practices: Conversational commands (not CLI wrappers) Proper settings in module.yaml Layered…

datacore-one/datacore · 90 tokens

gtd-content-writer

Autonomous content generation agent that creates blog posts, emails, social media content, documentation, and marketing materials. Generates drafts ready for human review. Invoked by ai-task-executor for :AI:content: tagged tasks.

datacore-one/datacore · 51 tokens

gtd-inbox-processor

Use this agent when you need to process individual entries from inbox.org in a GTD (Getting Things Done) system. This agent should be invoked:\n\n- After capturing new items to inbox.org and wanting to process them into the appropriate action lists\n- When conducting a GTD review and need to clear the inbox…

datacore-one/datacore · 572 tokens

gtd-project-manager

Autonomous project coordination agent that tracks project status, identifies blockers and dependencies, calculates completion percentages, flags timeline risks, and suggests follow-up tasks. Proactively escalates blockers older than 7 days. Invoked by ai-task-executor for :AI:pm: tagged tasks.

datacore-one/datacore · 61 tokens