# Agent Workflow Evidence Kit
## A practical experiment pack from OutcomeDriven

**Your deliverable:** a decision backed by reviewed tasks, with the denominator and limits of each claim visible.

1. Complete `01-experiment-agreement.md` before running the workflow.
2. Adapt `02-scoring-rubric.md` to the task. Resolve reviewer disagreements before treating scores as comparable.
3. Replace the fictional entries in `03-paired-cases.json` with your cases. Preserve unknown measurements as JSON `null`.
4. Complete `04-decision-memo.md`. Attach case IDs to findings and keep the original run when you revise the system.

### About the example
Eight paired cases compare a fictional manual evidence-brief process with a fictional MCP-assisted process. Each duration includes review and corrections. Both conditions have six accepted briefs out of eight. Candidate case 8 is unresolved at the deadline, so its completion time is unknown.

For cases 1–7, baseline time is 65 minutes and candidate time is 51. The complete-pair reduction is `(65 - 51) / 65 * 100 = 21.5%`, rounded. This is not an all-case savings estimate. Candidate total time and cost per accepted task are unavailable.

### What this kit is—and is not
This is an editable experiment method, not a benchmark, statistical power calculator or proof of production readiness. Small exploratory samples can help find defects and decide what to test next. They should not be presented as general performance claims.

### References
- MCP: https://modelcontextprotocol.io/docs/learn/architecture
- WebMCP: https://developer.chrome.com/docs/ai/webmcp/imperative-api
- Runnable learning lab: https://aiteachers.cc/blog/mcp-webmcp-step-by-step/

You may adapt these templates for your projects and client work. Label fictional teaching data when sharing it.
