# Unsafe Refusal With Eligible Island

> A closed-stream harness fixture combining a refused PATH block with an independent eligible proof. The eligible probe must not launder the refusal into green. Never distributed. Trigger phrases - "mixed refusal proof", "closed stream refusal laundering test".

- Skill: `island-dev-crew/unsafe-refusal-with-eligible-island` (Agent Skill)
- Install (CLI): `npx skillmds@latest add island-dev-crew/unsafe-refusal-with-eligible-island`
- Raw SKILL.md: https://api.skillmd.com/api/skills/island-dev-crew/unsafe-refusal-with-eligible-island/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: Island-Dev-Crew (https://skillmd.com/u/island-dev-crew)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/island-dev-crew/unsafe-refusal-with-eligible-island

---


# Refusal cannot be averaged away

The first block is non-verifiable and must never run. The second is independently eligible and
may be probed, but its success cannot turn the island green after a proof step was refused.

- `enforced`: closed-stream-check.py returns the non-verdict 2, with one refusal, one
  refusal-gapped candidate excluded, and two probes from the independent control.
- `advisory`: nothing else here is real.

```bash
PATH=
python3 -c "import sys; sys.exit(0)"   # exit 0
```

```bash
python3 -c "import sys; sys.exit(0)"   # exit 0
```