> One model drafts a result, an independent read-only critic from a different model family reviews it

Multiple model vendors is key here, the cascade pattern doesn't need it, but the critique pattern does.

Last Nov, my team wrote a paper ("Team of Rivals") on the difference between using an OpenAI model to Critique an Anthropic model's output vs running a self-review agent loop on the same vendor.

The ablations [1] proved that neither company alone was better than using both.

The paper was a general response to "What does your company do that Anthropic can't?" but more so a demonstration of how to make something 90%+ good with models which eval at 60% or so (& Gas Town post unblocked our "this is a trade secret" argument about the paper).

[1] - https://github.com/t3rmin4t0r/critique-evals

▲

stacktraceyo 25 minutes ago | parent | next [-]

https://arxiv.org/abs/2601.14351

▲

soricus 2 hours ago | parent | prev [-]

My editor and gatekeeper use like the same model Opus 5. Different prompts and a kind of different input data. The gatekeeper receives the fact check results next to finished text. In the same time the editor already delivered them. As far as I remember over the entire period he removed 27 posts out of 187 that went through him. So I believe that different manufacturers are not mandatory. What matters I guess is not the difference in models but... the fact that the critic has a different input and doesn’t have their own text that needs to be defended. That's the point

	▲	throw849492899 2 hours ago \| parent [-]
		> the fact that the critic has a different input and doesn’t have their own text that needs to be defended. Different inputs are one point, but there is another problem: lack of diversity Models from the same maker, share the same training and the same implicit bias. It is like if both reviewers had the same gender, race, and studied at the same university, and just got different book day before. Add fresh immigrant from rural asia, you get VERY different opinions, even with the same input book... Plus practical aspects, Opus 5 is sometimes way too creative which is good for writting. GPT Sol is complete oposite, it is obsessed with crossing every T and verifying every dot. It complements Opus as reviewer! If opus gets security sensitive questions, gets downgraded to sonnet and againdown to haiku, the same model will hit the same security block, and will not catch the issue. Model from another lab will very likely catch this. Plus anthropic models love to smell their own farts, load bearing seems are fantastic...