Experiment Protocol

Drives the experiment-registry MCP server through a controlled experiment lifecycle. The MCP owns the state machine, validates artefacts, and enforces methodology. This skill is the caller — not the logic.

Core Problem

Uncontrolled testing contaminates results. The most common failure mode is embedding success criteria inside the input the subject under test receives — this measures instruction-following ability, not the quality of the instructions themselves. A second failure mode is changing multiple variables between runs, which makes it impossible to attribute any result to any cause. The third is writing scoring criteria after seeing output, which lets expected results shape the rubric rather than the other way around.

The experiment-registry MCP server enforces the correct protocol mechanically. Claude's role is to produce artefacts and submit them — not to manage the workflow.

Phase 1 — Setup (collaborative)