Carve safety-adjacent task elements into isolated subagents with adversarial verification.
npx skills add https://github.com/AlexZio00/sovereign-skills --skill clean-room
Please help me install this Agent Skill. Skill: clean-room Source: https://github.com/AlexZio00/sovereign-skills/tree/main/clean-room Command: npx skills add https://github.com/AlexZio00/sovereign-skills --skill clean-room
This Skill ensures that tasks involving risky or safety-adjacent materials are executed by isolated subagents, maintaining the highest standards of isolation and safety.
/clean-room "carve this risky part out from the task and execute it in a clean-room subagent."
High-intent search queries and answers about installing and using this skill.
To isolate risky task elements for safe execution, carve out safety-adjacent components and execute them within isolated subagents, preventing contamination while performing adversarial verification to ensure no risks are overlooked.
Adversarial verification for safety-adjacent tasks is an analytical pass that scrutinizes isolated subagent output to ensure no risky content was incorrectly included or excluded, verifying overall task safety.
Track scope decisions when carving out risky task components by maintaining a descope ledger, which provides a detailed report of every carving decision, classification, and outcome to maintain accountability.
Yes, you can use subagent risk isolation for privacy and security sensitive tasks. It ensures that tasks involving sensitive areas are executed safely without compromise by isolating the risky elements during processing.
You need to descope risky elements from a task when the work involves privacy, security, or other safety-adjacent materials. Descoping executes these specific elements in a clean-room environment to prevent potential contamination.
LayerDynamicsReset AI operational scope to the original user request when task drift occurs.
alexfalkowskiImplement verified reliability-gap ledger entries with evidence-based validation.
pavanpaikExecute user-approved trades on Alpaca after human confirmation with audit logging.
kyo0221Dispatch isolated subagents for multi-task implementation plans with mandatory two-stage reviews.
