What problem does it solve? Managing eval-gated experiments in Caveman Cloud requires strict evidence review before any lifecycle change, but the current control API cannot atomically enforce lifecycle transitions and evidence gates. This Skill ensures every start, approve, cancel, promote, or rollback request is handled as a read-only, fail-closed review so unsafe or premature mutations are never executed. ## Core Features & Use Cases - Evidence-Based Review: Reads experiment state, results, sample sizes, quality scores, and guardrails (latency, error, cost, retry, drop, escalation) via MCP tools or the caveman CLI before recommending anything. - Fail-Closed Safety Gates: Blocks approval when results are pending, guardrails are missing, or evidence reports a breach, and treats unknown states and server errors as failures with exact cave_snake_code reporting. - Mutation Blocking: Refuses to execute lifecycle commands even after user approval, since agent-generatable action strings are not proof of human intent and the server does not yet enforce transitions atomically. - Use Case: A user asks to approve experiment 7f after seeing promising lift. The Skill reads the experiment and results, finds a guardrail report missing, states evidence is incomplete, and blocks the approval instead of promoting unverified changes. ## Quick Start Ask the agent to review the current state and results of a Caveman experiment and recommend whether it can be started, approved, cancelled, or rolled back.
