Skip to content

external validation: reproduce MCP demo runner #123

Description

@safal207

Context

CML now includes a one-command local MCP demo runner for the experimental agent-audit MCP path.

The broader external validation issue is #77. This issue is intentionally narrower: validate only the MCP demo runner on a clean local checkout.

Goal

Confirm that a contributor outside the maintainer loop can run the MCP demo payload runner and observe the expected sections.

What to run

git clone https://github.com/safal207/Causal-Memory-Layer.git
cd Causal-Memory-Layer
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -e ".[dev]"
python scripts/run_mcp_demo_payloads.py

Windows PowerShell activation:

.venv\Scripts\Activate.ps1

Expected output sections

The command should print JSON sections for:

health
audit_trace
evaluate_cause_band

Suggested validation note

Please add a short note under:

docs/evidence/external_validation/

Suggested filename:

<YYYY-MM-DD>-<github-handle>-mcp-demo-runner.md

Suggested structure:

# MCP Demo Runner External Validation Note

Validator: @your-handle
Date: YYYY-MM-DD
Repository commit: <commit-sha>

## Environment

- OS:
- Python version:

## Commands run

```bash
pip install -e ".[dev]"
python scripts/run_mcp_demo_payloads.py

Results

  • MCP demo runner: pass/fail
  • Sections observed: health / audit_trace / evaluate_cause_band

Notes

  • What worked:
  • What was confusing:
  • Any reproduction issues:

Interpretation

The MCP demo runner is / is not reproducible on a clean local checkout.

Non-claims

This validation does not prove production safety, enforcement behavior, compliance readiness, or stable Cause Band semantics.


## Acceptance criteria

- [ ] A validation note is added under `docs/evidence/external_validation/`.
- [ ] The note records OS, Python version, and commit SHA.
- [ ] The note records whether `python scripts/run_mcp_demo_payloads.py` passed.
- [ ] The note confirms whether `health`, `audit_trace`, and `evaluate_cause_band` sections appeared.
- [ ] Any failure or confusion is reported honestly.
- [ ] No production safety, enforcement, compliance, or stable Cause Band claim is made.

## Non-goals

- Do not change MCP tool semantics.
- Do not change CML audit semantics.
- Do not require Cursor or another MCP client.
- Do not require hosted services or private credentials.

## Why this matters

This gives CML a small, reviewer-friendly trust signal for the agent-audit integration path:

```text
one command -> visible MCP tool logic -> reproducible external note

Refs #77, #122, #105.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions