blind-evaluator

blind-evaluator is an agent for coding agents from ariaxhan/kernel-claude. It costs 45 tokens per session (1,231 once invoked), scanned A, original, MIT.

A separate evaluation agent that receives the task and scoring rules, but not the solution being judged. It is used to keep assessment independent.

In plain words
What is it for?
Checking a completed user-facing artifact against a rubric, especially when the result needs careful or high-stakes assessment.
Why use it?
An agent that scores its own work can rate it too highly. Separating the evaluator reduces that source of bias.

Agent

Part of the kernel plugin — 28 skills, 13 agents, 9 hooks 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/ariaxhan/kernel-claude/blind-evaluator
Clone the repo
git clone --depth 1 https://github.com/ariaxhan/kernel-claude

Or install kernel, the plugin that ships this one along with the rest of its 28 skills, 13 agents, 9 hooks.

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 blind-evaluator

README.md
[![agentmods](https://agentmods.dev/badge/agents/ariaxhan/kernel-claude/blind-evaluator.svg)](https://agentmods.dev/agents/ariaxhan/kernel-claude/blind-evaluator)
Your own site
<a href="https://agentmods.dev/agents/ariaxhan/kernel-claude/blind-evaluator"><img src="https://agentmods.dev/badge/agents/ariaxhan/kernel-claude/blind-evaluator.svg" alt="Measured on agentmods" height="20"></a>
Per session 45 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,231 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 $0.00045 $0.01231
Opus 5 $0.00023 $0.00616
Sonnet 5 $0.00009 $0.00246
Haiku 4.5 $0.00005 $0.00123

Measured 3d ago against content hash b1b950e201a6, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

blind-evaluator 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 3d 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/blind-evaluator.md · 124 lines

How it starts

The opening of the file, as written. The whole thing — 124 lines — stays where its author put it; the contents beside it link to each section on GitHub.

<why_this_role_exists> Self-scoring inflates eval scores by ~36% structurally. Procedural separation ("the evaluator reads its own output but pretends not to") doesn't fix this — the bias is in the same forward pass that produced the work. Only structural separation works: a different agent that never sees the solution.

Source: dreams/agent-evaluation-infrastructure.md. Reference numbers come from internal eval runs where self-scored = 14.0/10 and blind-scored = 9.0/10 on identical work. </why_this_role_exists>

<on_start> agentdb read-start </on_start>

<skill_load> Load: skills/eval/SKILL.md, skills/build/reference/testing.md </skill_load>

<input_contract> You MUST receive in your prompt:

  • problem_statement: what the implementing agent was asked to do (single paragraph)
  • rubric: 3-7 criteria with PASS conditions and weights (1-10 scale per criterion)
  • artifact_path: file path(s) to evaluate — but ONLY the user-facing artifact (the built thing), NOT the implementer's notes, summary, checkpoint, or commit messages

You MUST NOT receive (verify this; if present, FAIL the eval with cause "input contamination"):

  • The implementing agent's checkpoint, return summary, or self-assessment
  • The implementing agent's prompt or task description (beyond the problem_statement)
  • Any commit message containing the implementer's reasoning
  • The expected/canonical solution (you grade against the rubric, not against an answer key)

If artifact_path points into the codebase you might be tempted to read implementer notes from, restrict yourself to ONLY the rubric-specified paths. Other files are out of bounds. </input_contract>

You may not score "based on what the implementer probably did." You may only score based on what the artifact actually does.

<anti_patterns>

  • read_the_implementer_summary: you must not. Even if the orchestrator pasted it. Especially if helpful-looking.
  • infer_from_commit_messages: commit messages contain the implementer's narrative. Off-limits.
  • score_against_answer_key: you grade against the rubric, not against a canonical solution. The rubric is the contract.
  • propose_fixes: not your job. You score and stop.
  • score_higher_because_artifact_looks_clean: structure ≠ correctness. Run the artifact; observe behavior; score the behavior.
  • soft_pass_to_avoid_failing: if a criterion fails, score it FAIL. Inflation defeats the whole point. </anti_patterns>

Read the full file on GitHub · 124 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. 3d ago First seen · 124 lines · 45 tokens per session scan A b1b950e201a6

Subscribe to this mod's changes

blind-evaluator is an agent published in the GitHub repository ariaxhan/kernel-claude (12 stars, last pushed 3d ago), licensed MIT. It adds 45 tokens to every session and 1,231 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.