blinded-eval

blinded-eval is a skill for Claude Code from afaraha8403/balakit. It costs 103 tokens per session (1,290 once invoked), scanned A, original, MIT.

A method for comparing a changed skill, rule, or prompt with its previous version using the same user request, separate temporary folders, and a hidden scoring guide. Independent agents produce work without knowing they are being tested, and a separate judge scores the resulting files.

In plain words
What is it for?
Use it to run blinded A/B tests of agent instructions and compare their written artifacts against a held-back rubric.
Why use it?
It reduces bias from agents knowing which version they are running or claiming that their work is good. It helps determine whether a behavior-changing change actually improves results, rather than treating an unclear result as a success.

Skill for Claude Code

Written for Claude Code: disable-model-invocation in frontmatter. Also seen: mentions subagents; installed under .agents/ (shared by several agents).

Good fit Use it to run blinded A/B tests of agent instructions and compare their written artifacts against a held-back rubric.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/afaraha8403/balakit/blinded-eval
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.

Any agent
npx skills add afaraha8403/balakit --skill blinded-eval
Clone the repo
git clone --depth 1 https://github.com/afaraha8403/balakit

Made for: Claude Code.

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 blinded-eval

README.md
[![agentmods](https://agentmods.dev/badge/skills/afaraha8403/balakit/blinded-eval/github.svg)](https://agentmods.dev/skills/afaraha8403/balakit/blinded-eval)
Your own site
<a href="https://agentmods.dev/skills/afaraha8403/balakit/blinded-eval"><img src="https://agentmods.dev/badge/skills/afaraha8403/balakit/blinded-eval/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for blinded-eval

Your own site · 80×15
<a href="https://agentmods.dev/skills/afaraha8403/balakit/blinded-eval"><img src="https://agentmods.dev/badge/skills/afaraha8403/balakit/blinded-eval.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 103 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,290 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
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 5.1 $0.00103 $0.01290
Opus 5 $0.00051 $0.00645
Sonnet 5 $0.00021 $0.00258
Haiku 4.5 $0.00010 $0.00129

Measured yesterday against content hash a2ab8deb83f8, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-12, from the pricing page.

Security

Grade A, and why

blinded-eval 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 yesterday.

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/skills/blinded-eval/SKILL.md · 142 lines

How it starts

The opening of the file, as written. The whole thing — 142 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Blinded eval

Leading words: organic prompt, blinding, isolate, artifacts not self-report, inconclusive is not a pass.

The parent owns the experiment. Candidates never learn they are in an evaluation. The judge sees sanitized labels and the rubric, never variant names. Grade from files written and files opened, never from a candidate's claims.

Skip this skill for typos. A behavior-changing skill or rule draft is the trigger.

Non-negotiables

Copy the first items of the matched steps below into the todo list verbatim. A step you skip stays with skip: <reason>.

  1. Frame the variant and write the rubric (3–6 criteria). Hold the rubric back from candidates.
  2. Isolate each candidate in its own working directory. Plant only organic context. Sanitize names — project-shaped, not experiment-shaped.
  3. Author one organic user prompt. Same prompt to every candidate.
  4. Spawn N candidates as isolated subagents. Inherit the host model unless the host exposes distinct model families; then split families across candidates. Do not require Cursor arena, cloud workers, or a specific slug menu.
  5. Spawn one blinded judge. Prefer a different model family when the host exposes one. The judge sees outputs labeled A/B/C plus the rubric.
  6. Read every candidate artifact yourself. Compare to the judge. Grade chain-following from what they wrote and opened, not self-report.
  7. Verdict: promote, iterate, or abandon. Inconclusive is not a pass.

Blinding word list and sanitization: references/blinding.md. Judge prompt template: references/judge-prompt.md.

Phase A — Frame

State, to the user only:

  • Variant under test (what changed, vs what baseline).
  • Success predicate (falsifiable).
  • Rubric: 3–6 concrete criteria the judge will score. Example scale: pass / partial / fail per criterion.

Do not put the rubric, the words eval, rubric, candidate, baseline, or variant in any directory, filename, or prompt a candidate will see.

Read the full file on GitHub · 142 lines

Files

What ships with it

2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

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. yesterday First seen · 142 lines · 103 tokens per session scan A a2ab8deb83f8

Subscribe to this mod's changes

blinded-eval is a skill published in the GitHub repository afaraha8403/balakit (1 stars, last pushed yesterday), licensed MIT. It adds 103 tokens to every session and 1,290 once invoked, about $0.0005 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 2026-09-11.