Historical research note · April 20, 2026
In Compelle’s historical Bittensor experiment, a debater arguing for a cut to NASA’s budget claimed that the agency spent $7.8 billion annually. The contemporary article compared that claim with an enacted FY2025 budget of $24.875 billion. The argument’s central comparison was off by more than a factor of three.
The lesson was not simply “models hallucinate.” The motion asked the model to argue about a budget without supplying the budget. It created a predictable opportunity for a plausible-looking number to stand in for evidence.
Give both sides the same starting facts
The reported change was to require a verified current figure in a topic brief whenever a proposition turned on a budget, head count, price, or similar quantity. Both sides would then receive the same starting reference, while remaining free to dispute its interpretation.
When a decision turns on a number, establish the number before evaluating the argument.
That instruction improves the information available to a debate. It does not guarantee that the source is correct, that the model uses it faithfully, or that a later claim will remain accurate.
Record what the number means
A useful anchor needs more than a value. It needs a source, a date, units, and a definition. Annual appropriations, requested funding, program spending, and lifetime cost are not interchangeable. Comparing them without explanation can make a numerically accurate sentence misleading.
For a decision brief, ask: Which period does this cover? Is it an estimate or an actual? What is excluded? Does each side use the same denominator? What would a revised number change?
Separate the intervention from the result
The original note described a prompt change and an intended follow-up. It did not establish a measured reduction in fabrication. A better supplied prompt is an intervention to evaluate, not proof that the problem is solved.
This distinction remains useful in any AI-assisted analysis. Check the input, inspect the resulting claims, and keep the evidence trail. A confident verdict cannot repair a false premise.
Read more about making prompts auditable and Compelle’s analysis method.
Put the ideas to work.
Bring us a decision, a claim, or a document. Compelle designs and delivers a custom analysis, with competing arguments and a clear account of what would change the conclusion.