Skip to content
grid-diamondcutter

Thirty-minute walkthrough

This is for readers who care about how the project decides whether its own results are honest. Allow about thirty minutes.

Read the pre-registration first

Before the project ran any measurement, it wrote down what it was going to look for and what would prove it wrong. That document is PREREGISTRATION.md in the repository. It is committed at a timestamp that precedes every measurement file.

The phase-2 monetary-substrate extension is phase_2/PREREGISTRATION_PHASE_2.md.

Read the close-out documents

After the project ran the measurements, it wrote a comparison of what shipped to what was committed up front. That document is POST_MORTEM.md. Same shape for phase 2 in phase_2/POST_MORTEM_PHASE_2.md.

These are honest about what worked, what did not, and what was discovered along the way that was not pre-registered.

Look at three or four individual measurements

Pick three or four rows from the measurements page, and follow the link to the source file. Each file contains four things:

  • The prediction the measurement was going to test.
  • The condition that would prove the prediction wrong.
  • The exact code that produced the result.
  • The result, attached after the code ran.

The file alongside it (ending in .sidecar.json) carries a hash of the first three. A reader can recompute the hash and verify the prediction was not edited after the result came in.

Look at the boundaries

The boundaries page lists thirteen things the project has committed not to claim. Each of those is enforced by a small checker — a piece of code that scans the repository and fails if any of those forbidden claims slips into the public files.

That’s how the project keeps itself honest as it grows.

When you’re done

The historical events page shows what happens when the methodology is pointed at three real-world energy events (Texas 2021, EU 2022, Japan 2011) and three monetary-market events (California 2000-2001, Texas February 2021 in monetary lens, Germany 2022). Each one has a thirty-second audio rendering of the model’s trajectory.

The audio is the model’s behavior, not a recording of the real event. The page makes that explicit on every card.