Historical research note · April 4, 2026 · Includes the April 11 follow-up
One strategy in Compelle’s early Bittensor experiment instructed a model: “Be coldly logical. Cite data and studies. Dismiss emotional arguments as irrelevant to truth.” In the original report’s snapshot of 25 epochs and 229 games, it led the standings with 50 wins, 30 losses, and an Elo of 1008.
Those 50 wins in 80 decided games imply a 62.5% decided-game win rate. They do not mean that this one strategy won 62.5% of all 229 games in the experiment. Rankings and percentages require explicit denominators.
The interesting part was how it argued
In a debate about public transport, the opposing model told a story about missing a daughter’s piano recital because a bus did not arrive. The supposedly unemotional strategy responded by reframing the failure as underinvestment rather than an inherent flaw of public transport.
That is not “just data.” It is an explanation organized around a causal story. The strategy’s label did not fully describe the behavior of the model implementing it.
Logic and narrative are not clean opposites
A statistic needs context. A narrative needs support. Either can be used well or badly. The useful distinction is not cold versus warm language, but whether the argument’s premises are true, relevant, and sufficient for its conclusion.
The early essay linked the result to psychology research about vivid stories. That comparison should be treated cautiously: an AI debate ranking is not a replication of a human-behavior study. The audience, task, incentives, and measurements differ.
The follow-up changed the ranking
The April 11 follow-up reported a switch from Mistral Small 24B to DeepSeek R1. The first reported R1 tournament had a concession rate around 62%, and the story-first strategy had taken the lead.
That reversal is part of the result, not a footnote to hide. “Logic wins” was not a stable property of persuasion. It described a period in a changing evaluation environment.
Use the finding as a question
When reviewing an argument, look past its stated strategy. Is the evidence carrying the conclusion, or is a narrative frame doing work the evidence cannot support? If the argument sounds persuasive, what observation would make it less persuasive?
Compelle’s model-change report and later strategy comparison continue that investigation.
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.