Current status
- independently runnable
- 26 standalone tests passing
- callback-based experiment runner
- advisory MDC classifications
- JSON-compatible reporting
- Markdown reporting
- two demonstration experiments
ABOUT JUST A MARBLE
Just a Marble helps experimental systems ask whether testing the same axis again is likely to produce useful new information, or simply consume more time and tokens.
Purpose
Modern AI systems can generate, vary, and rerun experiments almost endlessly.
That ability is useful, but it creates another problem: knowing when another round of testing is worth the cost.
A harness can automate repetition. Just a Marble is intended to help evaluate the value of that repetition.
It does not decide whether an experiment is true, important, or complete. It examines a narrower question:
Capability
Just a Marble runs controlled variations across a defined experimental axis and compares the resulting metric structure.
It returns an advisory classification that can help a person, script, or agent decide whether to:
Boundary
Just a Marble does not establish:
Its output depends on the quality of the experiment, the chosen axis, the selected metric, and the supplied runs.
It is a witness, not a judge.
Current release
Alpha 1 is a standalone, single-axis Python reference implementation.
Separation
Just a Marble began as a small diagnostic instrument inside a broader private research system.
It was separated into its own project to test whether the instrument could stand independently, remain understandable outside its original environment, and eventually become useful to researchers, developers, game studios, and AI agents.
The broader research system remains private and separate from this release.
Name
Sometimes the experimental axis matters.
Sometimes the metric matters.
Sometimes the apparent distinction disappears when the right comparison is made.
And sometimes it is just a marble.
Author
Just a Marble is created by Stephen Montgomery.