humanize-korean Eval
====================

Status: review
Score: 69/100
Risk: medium
Decision: manual_review
Policy: review
Reason: Test manually in an isolated workspace and compare against safer alternatives.

Install:
npx skills add bam-bam-2/solo-skills --skill humanize-korean

Required checks:
- PASS Task fit: Task wording matches this skill metadata.
- PASS Install path: Install handoff is available.
- PASS Install command safety: standard package or runtime install path
- WARN Trust score: Potentially useful, but at least one trust signal needs human inspection.
- WARN Audit score: Needs review
- WARN Agent safety gate: Sparse or mixed signals. Useful for discovery, but not for autonomous installation.
- PASS License clarity: MIT
- WARN Permission surface: shell or command execution

Warnings:
- Trust score: Potentially useful, but at least one trust signal needs human inspection.
- Audit score: Needs review
- Agent safety gate: Sparse or mixed signals. Useful for discovery, but not for autonomous installation.
- Permission surface: shell or command execution
- High-risk permission hints: Shell or command execution
- SKILL.md opens with a lengthy version changelog (v1.5 변경 고지 with prior v1.2–v1.4 failures) instead of a quick-start; new users must read through historical rollout notes before reaching usage instructions.
- Workflow inconsistency: the fast-mode Phase 2 states the monolith agent '메모리에서 패턴 탐지 + 윤문 + 자체검증' and does not mention running metrics.py, but references/metrics.py documents itself as an 'external pre-processor for the monolith fast path' with a CLI designed to run BEFORE the monolith. It is unclear whether metrics.py is actually invoked in the default path.
- The six referenced agents (humanize-monolith, ai-tell-detector, korean-style-rewriter, content-fidelity-auditor, naturalness-reviewer, taxonomist) are not defined in the provided excerpt — if their definitions (SKILL/agent files) are missing from the repository, neither fast nor strict mode can execute.
- The 2–3 minute wall-clock target for 5,000 characters is asserted but no benchmark data, config details, or model assumptions are provided, making the performance claim unverifiable.
- The trigger list includes 'AI detector bypass 한글' — a dual-use capability that could facilitate academic or editorial deception. The skill has no responsible-use note clarifying that input should be original or properly licensed, nor a reference to applicable policies.
- Quality score needs review
- Stars/forks activity: 128 stars, 22 forks; issue activity unavailable in current metadata

Validation plan:
1. Inspect repository, README/SKILL.md, license, and recent commits before production use.
2. Install in an isolated workspace or sandbox with no production secrets available.
3. Run the smallest representative task and record files touched, commands run, network access, and outputs.
4. Compare the selected skill against at least one alternative when the eval status is review or failed.
5. Promote only after the agent reports a successful verification result and unresolved warnings are accepted.

Do not use when:
- teams that need a vendor-supported SLA
- production agents without a repository review
- SKILL.md opens with a lengthy version changelog (v1.5 변경 고지 with prior v1.2–v1.4 failures) instead of a quick-start; new users must read through historical rollout notes before reaching usage instructions.
- High-risk permission hints: Shell or command execution
- Workflow inconsistency: the fast-mode Phase 2 states the monolith agent '메모리에서 패턴 탐지 + 윤문 + 자체검증' and does not mention running metrics.py, but references/metrics.py documents itself as an 'external pre-processor for the monolith fast path' with a CLI designed to run BEFORE the monolith. It is unclear whether metrics.py is actually invoked in the default path.
- The six referenced agents (humanize-monolith, ai-tell-detector, korean-style-rewriter, content-fidelity-auditor, naturalness-reviewer, taxonomist) are not defined in the provided excerpt — if their definitions (SKILL/agent files) are missing from the repository, neither fast nor strict mode can execute.
- The 2–3 minute wall-clock target for 5,000 characters is asserted but no benchmark data, config details, or model assumptions are provided, making the performance claim unverifiable.
- The trigger list includes 'AI detector bypass 한글' — a dual-use capability that could facilitate academic or editorial deception. The skill has no responsible-use note clarifying that input should be original or properly licensed, nor a reference to applicable policies.

URLs:
- Skill: https://www.openagentskill.com/skills/bam-bam-2-humanize-korean
- Audit: https://www.openagentskill.com/skills/bam-bam-2-humanize-korean/audit
- JSON: https://www.openagentskill.com/api/agent/evals?slug=bam-bam-2-humanize-korean