Fix Loop

BEHAVIOR-1K ASPIRE Fix Loop

nvlabs 4e84ffb 7 files · 12.7 KB Updated

File contents

BEHAVIOR-1K ASPIRE Fix Loop

Use this skill for:

Follow the protocol and run BEHAVIOR-1K Soda Can ASPIRE experiments.
Follow the protocol and run BEHAVIOR-1K Radio ASPIRE experiments.

From aspire/sim, read:

  1. ../../AGENTS.md
  2. CLAUDE.md
  3. .claude/behavior/CLAUDE.md
  4. .claude/behavior/api-reference.md
  5. .claude/behavior/skills/system-pipeline.md
  6. .claude/behavior/fix-loop/INSTRUCTIONS.md
  7. .claude/behavior/fix-loop/clean-task-slate.md

Then follow .claude/behavior/fix-loop/main-agent-prompt.md.

The protocol is fixed:

  1. Preflight and wait for approval.
  2. Learn a campaign-owned skill library on seeds 26-35.
  3. Freeze skills and the experimental contract—not a policy.
  4. Run seeds 1-25 with a fresh context and empty policy per seed, allowing debugging only within that seed.
  5. Preserve and report all outcomes.

Every trial uses replay mode, which calls env.step directly without an internal model call. Do not delete the shared built-in multi-turn engine; other experiments use it. Do not launch before preflight approval.

nvlabs/aspire/tree/main/aspire/sim/.claude/behavior/fix-loop commit 4e84ffbe55

Frequently asked questions

npx skillmds@latest add nvlabs/fix-loop