What problem does it solve? Managing eval-gated experiments in Caveman Cloud requires strict evidence checks before any lifecycle change, but the current control API does not yet enforce a complete lifecycle transition table and evidence gate atomically. This Skill ensures every start, approve, cancel, promote, or rollback request is evaluated against read-only evidence and that unsafe execution is blocked rather than silently performed. ## Core Features & Use Cases - Read-only experiment inspection: Load project context, experiment detail, and results via MCP tools (caveman_context, caveman_experiment_get) or the caveman CLI fallback. - Evidence evaluation: Report lifecycle state, safety class, sample sizes, quality results, guardrails (latency, error, cost, retry, drop, escalation), and evidence cost, treating absent fields as incomplete rather than passing. - Action recommendation with hard gates: Propose exactly one of start, approve, cancel, or rollback with explicit blocking rules, and never execute a lifecycle mutation even after user approval. - Use Case: A user asks to approve experiment 7f after seeing positive lift. The Skill reads the results, finds a guardrail report missing, states evidence is incomplete, and refuses to propose approval. ## Quick Start Ask the assistant to review the current state and results of a Caveman experiment and recommend whether it can be approved.
