No accuracy number here: Krishnan's councils kept 22% to 37% of one-model ideas

The band this experiment actually produces is 22 to 37 percent ^[13], and it counts idea survival, not correctness. The unit is a card: each answer was broken into small items using Sonnet, a mechanism, an observation, a metric, a failure mode, then cards meaning the same thing were clustered and labelled single-model if they appeared in one solo answer, shared if they appeared in more ^[9]. Two judges rated the solo-derived clusters blind to which model wrote them and to whether any council kept them ^[9]. Useful, non-obvious, worth keeping is the whole rubric ^[9]. There is no key, so there is no chance line to sit above or below ^[16].

The consensus tilt is where the arithmetic gets slippery. Krishnan puts peer review's preference for shared ideas at an 11 point uplift and calls it a 50% relative lift ^[7]. Eleven over twenty-two is fifty percent; eleven over twenty-four is nearer forty-six ^[14], and both single-model figures appear in his own write-up ^[4]. He also says the denominator for shared ideas is small ^[8], which puts the load on the thinnest count in the study.

On the four-agent framing: the only configuration with four models in it is the blend, three answers handed to a fourth model to write the final version ^[2]. That fourth model is a bottleneck, not a debater. What it reliably improved was prose, which Krishnan describes as calmer, more complete, less jagged ^[10]. The losses read the other way: scent cartridges salvaged from retail and traded as status symbols in a squatted mall, or the argument that logged-but-deprioritised risks are worse than unknown ones because they manufacture a false sense of control ^[17].