Skip to content

deontewattsv1/unison-kaggle

v0.1.0MIT

Run the frozen UNISON-StateBench v0.1 empirical workflow through Kaggle's official MCP server without modifying the benchmark oracle or corpus.

GitHubOpen Website

UNISON Kaggle ChatGPT plugin

This package connects UNISON-StateBench to Kaggle's official MCP server:

https://www.kaggle.com/mcp

It does not proxy or store Kaggle credentials. Kaggle handles account authorization through its own MCP/OAuth surface.

Why use the official Kaggle MCP

Kaggle's MCP exposes live Notebook and Benchmark operations and supports OAuth authorization. Using the official server avoids creating a second credential-handling service and keeps Kaggle authorization under Kaggle's control.

ChatGPT connection

If custom plugins / developer-mode MCP connections are available in your ChatGPT workspace:

  1. Open Settings → Security and login and enable Developer mode.
  2. Open Plugins and add an MCP connection.
  3. Use:
    • Name: Kaggle
    • MCP URL: https://www.kaggle.com/mcp
  4. Complete Kaggle OAuth authorization when prompted.
  5. Install or load this plugin package so the unison-empirical skill is available alongside the Kaggle MCP connection.

The MCP endpoint can also be used directly without the skill package, but the skill encodes the frozen UNISON execution order and scientific boundaries.

UNISON run sequence

Kaggle OAuth
  -> runtime-snapshot.json
  -> verify source revision + plan digest
  -> freeze exact lineup
  -> explicit execution approval
  -> 162N live model calls
  -> per-model append-only JSONL
  -> results-manifest.json
  -> IPR/RCR/CRR/FHR/CAL/PLR/AIR/SCR
  -> 95% scenario-cluster confidence intervals
  -> failure topology

No empirical model-performance claim is valid before the final result-sealing and analysis steps complete.