What problem does it solve? Long research sessions lose context across threads: goals, routes, experiment evidence, and review results disappear, making it impossible to resume work safely. This Skill runs exactly one visible, auditable worker iteration so progress is recorded and the next session can pick up where things stopped. ## Core Features & Use Cases - Bounded single-iteration execution: Chooses exactly one action (scout, idea, execute, debug, review, analysis, write, branch, rollback, stop, ask-human, or status-check) with a stated prediction, falsification condition, budget, and stop condition. - Receipt-based state management: Registers, finalizes, and verifies iterations via ds_lite_iteration.py, updates the research graph through ds_lite_state.py with expected revisions, and renders STATUS.md. - Evidence and safety gates: Enforces Evidence Pack and review requirements before analysis, preserves negative results, and blocks unauthorized GPU, long-run, or external-data actions. - Use Case: Resuming a multi-day ML experiment project, the worker reads the mission board, finds a blocked training run, performs one status check with evidence backup, records a frontier decision, finalizes a terminal receipt, and hands off the next single action. ## Quick Start Ask the agent to read the mission board, choose one bounded next action, record the frontier decision, update STATUS, and stop.
