loop-experiment

# /loop-experiment. Disciplined autoresearch-style optimization Lifts the `setup.md` pattern from `zkarimi22/autoresearch-anything` (public repo; pattern lifted, not installed) and grafts it onto the stack's discipline. ## Phase 0. Fitness-Function Gate (HARD STOP if fail) Before scaffolding ANY loop experiment, screen the task against 2 conditions. If either fails, refuse to set up the experiment. Push back to the user. 1. **Stable measurable fitness function exists**: a function that maps (input, output) → number. Examples: engagement count, qualified-lead yield, classification accuracy, test pass rate, token count, latency. NOT examples: "the operator's taste", "feels right", "looks better". 2. **Cost-per-iteration cheap enough for 50-700 attempts**: each loop iteration must be runnable in seconds-to-minutes, not hours-to-days. A once-daily content draft is borderline. Most subjective tasks fail this gate. Example screen: - ✅ a content pipeline (engagement = likes+replies+bookmarks; iter = 1 draft/day, batch on 30+ days) - ✅ lead-gen prompts (qualified-lead yield per query; iter = batch query) - ✅ a sub-agent's scope classifier (HANDLE/ESCALATE accuracy; iter = prompt rerun) - ✅ skill prompt optimization (`skill_failure_tracker.py` count; iter = tracked over usage) - ❌ sourcing, intel, subjective outreach drafts (no stable fitness. the operator's taste IS the metric) If the user requests a loop experiment for a non-fit task: refuse politely, explain the gate, suggest...