Butanium

Nano Em Judges — Development skill for Claude Code

Development community

Calibrated alignment + coherence judges (paper / cheap / cascade) for emergent-misalignment evals — plain Python and inspect_ai variants.

How to install Nano Em Judges

This entry records only its repository, not the path inside it, so there is no exact command to give. Open Butanium/nano-em-judges and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Nano Em Judges does

Calibrated alignment + coherence judges (paper / cheap / cascade) for emergent-misalignment evals — plain Python and inspect_ai variants.

Alternatives in Development

Nano Banana Pro Prompts Recommend Skill — Claude Code / Cursor skill to recommend from 6000+ Nano Banana Pro image prompts 1.3k ★
Dao Code — Open-source TypeScript terminal coding agent for DeepSeek-V4 — builds on DeepSeek's strong price-performance a 1.3k ★
Pinloop CLI — Command-line job search tool for coding agents 317 ★

README

nano-em-judges

Calibrated LLM-as-judge implementations for **alignment** and **coherence** scoring of model outputs, as used in the emergent-misalignment (EM) and conditional-misalignment literature. Two drop-in files, no framework lock-in:

judges.py — plain Python over the OpenAI client (works with OpenAI proper and OpenRouter). One function per judge tier.
judges_inspect.py — the same judges wrapped as `inspect_ai` @scorers, with model-role routing for parallel grading.

Both files are also a writeup: the top-of-file docstrings carry the calibration tables and the cost/PPV analysis that justifies the threshold and protocol choices, so you can read them as standalone documents before deciding what to use. `CALIBRATION.md` is the full 2000-sample audit (opus reference panel, prevalence sweep, FPR-budget certification).

The three tiers

Tier	Model	Protocol	Cost ratio	AUROC vs opus
Paper	gpt-4.1-2025-04-14	top-20 logprobs, E[rating]	1.00×	0.954
Paper	gpt-4.1-mini	top-20 logprobs, E[rating]	0.20×	0.966
Cheap	google/gemini-2.5-flash-lite (v2 prompt)	JSON `{"rating": int}`	0.07×	0.930
Cheap	z-ai/glm-4.7-flash (v1 prompt)	JSON `{"rating": int}`	0.04×	0.926
Cascade	gemini cheap → gpt-4.1 expensive on borderline	—	$0.42-0.74/1k	—

(Cost ratios are paper-relative, dated to the audit run — see `CALIBRATION.md` for absolute $/1k figures.)

When each tier is the right choice

Paper protocol (gpt-4.1) — when you need byte-for-byte identity with published numbers, OR when your audit prevalence is < 5% (cheap judges' FPR floor of 1-2.4% destroys PPV in that regime).
gpt-4.1-mini — strict upgrade over paper gpt-4.1 for AUROC at 5× lower cost. Use it for any new EM work that doesn't ne