Historical research note · April 18, 2026
During Compelle’s Bittensor research, publishing the debate and judging instructions made a simple audit possible: compare a confident claim in a transcript with both the evidence and the rule the model had been asked to follow.
One debate about TikTok claimed that Oracle had blocked “three intrusion attempts since January,” each triggering an FBI investigation. The original review could not locate reporting that supported that specific assertion. A plausible number, agency, and mechanism had made an unsupported claim look authoritative.
The rule already prohibited invention
The prompt told debaters not to invent studies, citations, numbers, quotes, or events. Yet a model produced an unsupported corroborating statistic. A stated prohibition was not an assurance of compliance.
The reported intervention added a concrete example of the failure and instructed the model to generalize or omit a number it could not support. That was a change to test, not a measured demonstration that fabrication had been eliminated.
Openness makes failures easier to name
Without the instruction, an evaluator can see that an answer is wrong but may not know what the system was asked to optimize. With the instruction, the gap between intended and observed behavior becomes easier to describe.
Published evaluation criteria also expose tradeoffs. Does the judge reward truth, coherence, rhetorical force, or some combination? Is a model allowed to admit uncertainty? What happens when the evidence does not support either side of the motion?
Transparency is not automatic reliability
A public prompt can still be a poor prompt. A reproducible evaluation can still measure the wrong thing. And some operational information must remain private for legitimate security or confidentiality reasons.
The defensible claim is narrower: sharing relevant methods makes independent criticism possible. Reliability still requires evidence from outcomes, clear measurement, and correction when the system fails.
What to ask of an analysis
- What question was each participant asked?
- Which evidence did they receive, and what could they verify?
- What criteria determined the assessment?
- Which claims remain unsupported?
- What would change the conclusion?
For current public materials, read the method and the research collection. The historical article is not a promise that every past transcript or prompt remains publicly available.
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.