What records are required to make a GEO claim testable?

In plain language

Saying, 'We improved,' is not enough. You need the earlier mark, the work performed, the later result from the same test and the method used to calculate the mark. Without them, no one can check the success again.

Why this matters

Without a chain of records, the reason, size and scope of the change cannot be verified. The result may come from a selected example, a changed model or a different question.

Do not confuse

• A screenshot alone is not a complete evidence package.
• A measured change and a causal claim are different.
• A change on one platform cannot be generalised to every system.

What should you do?

1. Give every claim a unique identifier.
2. Lock the baseline before making the change.
3. Record the change at file and version level.
4. Preserve test conditions and raw outputs.
5. Write the assessment rules in advance.
6. Explain uncertainty between the result and its possible cause.

How do you audit it?

• Can the claim be reproduced?
• Are the baseline and later measurement comparable?
• Were the question set and model conditions the same?
• Does the raw output produce the score in the report?
• Does the claim exceed the boundary supported by the evidence?