The unit of coverage in a harness like this is the pairing, and pairings do not grow the way the model list does. Going from three models to four takes the distinct cross-model pairs from three to six [1]. What the run actually needed was one specific cell in that matrix, and the field test behind AdversarialDebate v0.2.2 [16] had no way to fill it without a fourth lab in a fourth region.
A three-PR run went first, to confirm the pipeline worked [3]. It did, but a smoke test only proves that much: it exercises the code paths for whichever pairings you configured, and each of those pairings returns a number that reads like a finding regardless. The author, writing on dev.to, says that stopping at the small corpus would have produced a clean story and the wrong one [19].
The waste is worth doing by hand. By the time Mistral joined, the first three models had each reviewed about 148 PRs, while Mistral finished about 70 to 73 [12]. Mistral pairs were then limited to the 70 PRs all four models shared [13]. That leaves roughly 78 passes per model outside the intersection, about 234 across three models, which brackets the 228 stranded passes reported [2]. At the stated $0.15, each stranded single-pass review cost around $0.0007 [3]. Cheapest design review I have been shown a receipt for.
Treat the winning pair's average as a claim about a corpus rather than a ranking of models. For it to mean that DeepSeek plus Mistral beats GPT plus Gemini, every pair has to be scored on the same 70 shared PRs, under the same rubric, with the difficulty spread of those 70 resembling the rest of the 148. The post states the intersection restriction for the Mistral pairs [13]. If the older pairs kept their full history in the comparison, then a 70-PR sample is being read against a 148-PR one.
The behavioural numbers need the same handling. About 34 concessions per shared PR [4] is a lot of ground for one pair to give in a review. The cascade count over those same 70 PRs works out to 63% [5], close enough to the reported in-pair capitulation rate [9] that the denominator is probably debates or PRs rather than turns. Pin that down before comparing the figure to any other system, because a per-turn rate and a per-debate rate of the same value describe very different failures.
What I would change is the order, not the decision. Enumerate the corners the thesis needs first, homogeneous through strong diversity [7], then pick models to hit them, then freeze the corpus. The author's own summary is that three models proved the system ran and four made pairing strategy learnable [18], and the wider spectrum also promoted the same-model control from curiosity to evidence that weak diversity is its own failure mode [17]. That is a statement about design coverage. Design coverage is cheaper to fix on a whiteboard than at pass 148.
