One model sounds certain. Five models show you the risk.

The failure mode of a good model is not that it stalls or refuses. It is that it hands you a fluent, well-structured, completely wrong answer in the same register it uses for a correct one. You cannot read uncertainty off the surface, because the surface is always calm. The obvious fix, blend several models into one polished consensus voice, makes it worse: averaging smooths the exact split you needed to see back over into another confident paragraph. So the council does not blend. It keeps every answer separate and puts the seams on display. This is the mixture-of-agents shape, many proposers and one aggregator, tuned so the output is a comparison instead of a blend.

Here is the whole mechanism, live, and one run only teaches one shape, so the replica below carries three. The outlier: a cross-region latency question with a physically correct answer near 80 ms, where one proposer confidently reports a same-region number an order of magnitude too low. The split: a Postgres migration question where the room divides three against two and both camps are internally consistent, because the real fork is a hidden version assumption. And consensus: a question about webhooks and absent deploys (readers of the deploy radar post already know this one) where five models agree for the same reason stated five ways. Pick a scenario, press the button, and watch what the console shows me.

Plot that run and the argument makes itself. Four answers cluster between 70 and 85 ms, right where the physics of transatlantic fiber puts them. DeepSeek V3 sits alone at 8 ms, a cross-AZ number that would read as perfectly plausible in isolation. On a single-model day, one time in five, that 8 ms is the answer I would have gotten, delivered in prose exactly as confident as the correct one. The spread is what makes the error visible before it ships.

The panel is cross-lineage on purpose, because two models trained on the same data tend to be wrong in the same places. Claude Opus 4.8 sits next to Llama 3.3 70B from Meta, Qwen2.5 72B from Alibaba, and two from DeepSeek: V3 and the R1 reasoning model, which streams its visible thinking so I can watch it reason before it commits. Five makers, five sets of blind spots. They answer in parallel and never read each other, so nobody anchors on the loudest voice in the room.

Then a sixth model reconciles. I use Sonnet for the job, and the load-bearing detail is that it is not one of the five. A judge that also answered would be grading its own homework. Sonnet reads every labeled answer and writes one report with fixed sections: what they agree on, where they diverge (attributed by name, so I can see the outlier was DeepSeek and not Opus), which claims look like outright errors, and a single consolidated recommendation. Handing an answer to a different model to grade is the LLM-as-a-judge pattern; the council just runs it across a whole panel at once.

Asking five models a question is useful. Making them argue is where it gets sharp. The second mode, Review, takes an artifact instead of a question, a diff or an essay or a schema, and runs two rounds. In round one every model critiques it for defects only: find the passage, say what is wrong, rate the severity, propose no fixes. That produces a pile of findings, and some of them are lazy, and a few are hallucinated outright. Round two sorts the pile. Each model receives the other models' critiques with the names stripped to Reviewer A and Reviewer B, and its one instruction is to find the single weakest claim and knock it down. A real defect survives a peer trying to refute it. A style nitpick with no reproduction does not. The anonymizing is deliberate: a model that could see a critique came from Opus might defer to it, so the identities come off and every claim stands on its own legs. Then Sonnet reconciles both rounds into confirmed defects, contested defects, rebuttals that themselves look wrong, and a prioritized punch list.

This is not free, and I want to be exact about the cost. Every ask fans out to five paid models plus the synthesis, and a review runs both rounds. So the council is owner-only and noindex, with no public cost surface, and it is the wrong tool for most questions. If an answer is cheap to verify or cheap to be wrong about, ask one model and move on. I reach for the council when being wrong is expensive and I cannot tell from the prose alone: a security review, an architectural call I will have to live inside for a year, a claim I am about to publish. Picking which single model to trust for a given job is what the benchmarks page is for. The council is for the times I would rather not pick.

After enough runs the verdict patterns start to rhyme, and there are only four of them. A tight cluster is consensus: boring, and usually trustworthy. Two camps is a split, and a clean split almost always traces to an assumption nobody stated, so the useful move is finding the fork, never picking a side by vote. Four and a loner is the outlier, the shape the latency run showed. And the fourth shape looks exactly like the first: five answers in one tight cluster, sitting together in the wrong place.

That fourth shape is the honest limit, and it sits right at the center of the design: the council shows me where the risk is, not what is true. If all five models are confidently wrong in the same way, the report will call it consensus and hand me a clean, well-attributed lie. A diverse panel narrows that failure. It does not close it. Every stream still runs through the same red-lines guard as the rest of the site, and the final call stays mine. The council does not decide anything. It just makes sure I am deciding with the disagreement laid out in front of me, instead of buried under one smooth voice that sounded certain either way.