What problem does it solve?
Open-ended investigations by AI agents degrade into reactive sampling: the agent grinds one corner, forgets leads it noticed, and states conclusions without checking unexplained results. This Skill imposes a fuzzer-style feedback loop so every action is a recorded bet and every conclusion passes a gate.
Core Features & Use Cases
- Append-only investigation notebook: Records frames, bets, settles, reframes, and claims as JSONL lines that survive context loss and session restarts.
- Coverage panel: Computes explored areas, untried leads, novelty trends, open anomalies, and a blocked/open conclusion gate via
scripts/tally.py.
- Hard protocol rules: No action without a prior prediction, mandatory verdicts and updates, forced reframes after three anomalies, and a kill-shot requirement before any claim.
- Use Case: While debugging a flaky test or investigating user churn, log each experiment as a bet with a falsifiable prediction, settle it with a verdict, and run the panel before drawing any conclusion to see which leads remain unexplored.
Quick Start
Start my investigation by creating a fuzz notebook frame for my research question, then log a bet before each action and show me the panel after every batch of results.
