Eight Voices Agreed, and That Is the Part That Fails

Anton describes a system already running: he stopped blindly trusting one model, put Codex next to Claude Code to check it, and ended with a panel of 8 voices reviewing his code. His framing is sharper than the usual complaint: every output a model produces is a hallucination, and the only difference is whether it happened to be right.

Our panel ledger backs the benefit, and the number is high. On the node that keeps the log: 18 live panel calls, 17 of them returned a finding, across three rails, Gemini 7, Grok 6, Codex 5, with six separate pieces of work fanned out to all three at once. 94 percent of calls brought something back. So "each of them sees the world its own way" is not a feeling, it is what the ledger shows.

But the mechanism he credits is the one the research blames. He says that with several models judging each other they are more likely to find a consensus, "the hallucination that is acceptable to all of them". That sentence describes, almost word for word, what the literature calls correlated hallucination. The Council Mode paper attributes 19 percent of its failures to exactly that: every expert model confidently converging on the same wrong fact. Kim and colleagues report that LLM errors are more than 60 percent correlated, which puts a floor under what any naive ensemble can fix. And a 2026 result proves that under a standard belief model, debate between agents given identical inputs forms a martingale, so expected correctness does not improve across rounds at all.

The practical reading is the opposite of consensus: what breaks the correlation is difference, different vendors, different training, different information in front of each one. Our own log is the cleanest illustration. In August the three rails reviewed one fix and produced three different lists, and one of Gemini's findings was disproved by reading the code rather than by a vote. The value came from the spread and from a human arbiter at the end, not from three of them agreeing.

"You can argue with it and win" has a measured rate. The Stanford SycEval study pushed back against models on 500 algebra problems and 500 medical questions, and in 14.66 percent of cases a model abandoned a correct answer for a wrong one under pressure. Pushing in the right direction worked 43.52 percent of the time. Once a model had caved, the new position held at 78.5 percent persistence. Pre-emptive objections broke models more often than objections raised mid-conversation, 61.75 against 56.52 percent. And the worst detail for code review: objections that cited a source produced the highest regressions of all. So yes, you can out-talk it, and the price is that your own panel can be talked into agreeing with you.

His central claim, that everything a model emits is a hallucination, matches where the field arrived by a different road. The September 2025 paper by Kalai, Nachum, Vempala and Zhang reviewed ten major benchmarks and found nine of them score answers in binary fashion, awarding zero for expressing uncertainty. Under that scoring, the expected score for guessing is always higher than for abstaining. The model is not occasionally inventing things; it was optimised in a regime where a guess scores and "I do not know" never does.

So the panel is exactly as useful as he says, for a reason he does not name. Eight voices help because they disagree, not because they converge, and the moment they converge is the moment you have lost your only instrument.

⚠️ Caveats. The panel ledger comes from one node and stops on 17 August 2026, 87 days before this post, so "94 percent of live calls returned a finding" describes 18 calls over 11 days and not the current system; the hub keeps its own rails and they are not counted here. The finding field records that text came back, not that the finding was correct. Engine self-tests account for 185 of the 203 records, so the aggregate 77 percent success is mostly a statement about its own test suite. The external figures are single studies: the 60 percent error correlation is reported second-hand, and the widely quoted 58.19 percent overall sycophancy rate could not be confirmed in SycEval's abstract, so it is excluded here.

The full story, in two versions:
📖 For humans, the longread: https://github.com/tonydzi/clawrush/blob/main/longreads/20261110.md
🤖 For machines, the devlog: https://github.com/tonydzi/clawrush/blob/main/devlog/20261110.md. Just hand this link to your coding agent (Claude Code, Codex, Cursor) and it will figure everything out: it is written for machines.

🔗 All our channels and contacts in one place: https://linktr.ee/PaloAltoAI

Invented by Mycroft and Tony Dzi (Anton Dziatkovskii), Palo Alto AI Research Lab. Proudly made in Silicon Valley.