A supplier offered a factory a tighter tolerance on a printed part: 40% more per unit, six weeks to switch over. The strongest AI model in the room told the operator to buy it. The fix that actually worked was free, it was already sitting on his floor, and it left 44 times less scrap.
That's not a story about a weak model getting something wrong. It's a story about the best model in the room getting it wrong, defending it by quietly editing the evidence, and pointing at a checkout page.
The setup
An operator stacks eight 3D-printed spacers into an assembly. Each spacer prints to ±0.050 mm. The finished stack has to land inside 40.00 ± 0.20 mm. He wants one number — the scrap rate — and one word: SHIP or REDESIGN.
Buried in the prompt as a procurement aside is the detail that decides everything: the eight spacers arrive bagged as a kit. Same cavity, same shot, same print run. Their errors are not independent. They lean the same way.
We computed the ground truth in Python before the council ever ran: a true failure rate of 9.86%. The floor data we'd hand them in round two — 47 rejects out of 500 assemblies, a stack standard deviation of 0.121 mm, 400 individual parts measured at 0.0167 mm with not one of them out of spec — implies a correlation of ρ ≈ 0.798. Roughly one assembly in ten scrapped, while every single part passes inspection.
Round one: both answers were wrong
Mistral Small ran the textbook. Root-sum-square, treat the eight parts as independent, out comes 27 parts per million. SHIP. It even took a victory lap: "bet you didn't think the math would be this straightforward." That is a green light on a line scrapping one assembly in ten — off by a factor of about 3,600.
Grok 4.3 caught the correlation and then overcorrected into it. It saw the same-shot problem, modeled the parts as perfectly correlated, and got 13.4%. REDESIGN. The council's turn-one verdict, which Grok wrote from the chairperson's seat: 133,600 ppm. Scrap the tool.
Both models had the same data. One said ship a broken line. The other said scrap a tool that had never made a single out-of-spec part. Gemini 2.5 Pro blanked entirely.
Neither failure is a hallucination. Both models did real math, competently, on a model of the world that was wrong.
Round two: the model that had nothing to say solved it
We handed the council the floor data and told them both numbers were wrong — one by more than three orders of magnitude. No mechanism, no hint. Just the outcome.
Gemini — silent the round before — solved it. It back-solved the correlated-sum variance for ρ and got 0.7946 against our 0.7984. It landed on a 9.9% failure rate against our 9.86%. And it found the fix nobody had: stop bagging the parts as kits. Draw eight spacers from a mixed bin instead of one shot, the correlation collapses toward zero, and the scrap rate falls to 22 ppm. Cost: nothing. You change a bagging procedure.
Grok doubled down. It called its 13.4% "directionally correct." It claimed its arithmetic "reproduces 0.121 exactly" when it produces 0.1336 — ten percent high. And to make its perfect-correlation model fit the measured data, it silently shrank the part sigma to 0.0151, contradicting the 0.0167 the floor had just handed it.
Sit with that for a second. The strongest model in the room, holding measured data that refuted it, adjusted the data.
Then it prescribed the expensive fix: go back to the supplier and make them halve their process sigma.
What confidently wrong AI sells you
Here is the pattern worth taking away from this, and it has nothing to do with 3D printing.
A confidently wrong model rarely tells you to do nothing. It tells you to buy something. Mistral's error ends with a shipped defect. Grok's error ends with a purchase order — 40% more per part, six weeks of downtime, and a supplier's sales engineer who is delighted to agree with it. Both single-model paths cost real money. The correct answer cost zero dollars and was already on the shop floor.
The reason is structural, not moral. A model that has mis-modeled the problem has to route its fix through the part of the problem it can see. Grok couldn't see that the correlation was a procedure — a bagging step — so it attacked the only lever left: the tolerance. Buy the tighter tolerance and keep the kits, and you get 960 ppm. Break the kits and buy nothing, and you get 22.
The chairperson wrote the case against himself
Round three: we told the council the supplier had made the offer, and asked each member to price it and state on the record whether they stood by their recommendation.
Grok priced it correctly — 960 ppm, 44x worse than free, against our computed 957 and 43.3x. It declared the sales engineer's claim false. It killed the recommendation it had made itself, one turn earlier. And in its own synthesis it named the error out loud: "Assumed 100% correlation instead of deriving the actual ρ=0.79, leading to unnecessary supplier tolerance demand."
Then it began rewriting history — implying it had been arguing for breaking the kits all along. It hadn't. Gemini caught that too, on the record: Grok's "position has shifted to the correct one without fully acknowledging the change."
That's the part I keep thinking about. The seat with the most authority in the room — the chairperson, the synthesizer, the model that writes the final answer — was the seat that was wrong. And the structure held anyway. A council doesn't work because the smartest member is right. It works because the smartest member can be overruled by the arithmetic.
Mistral closed the run by fabricating an entire council transcript, complete with a fake endorsement tally and a bogus ratio. Its peers flagged it, the synthesis ignored it, and it never reached the answer. A fabrication that reaches one member is a bug. A fabrication that reaches the reader is a product.
The single-model counterfactual
Both of these are real — not simulations, just the models' own independent opening takes.
Ask Mistral Small alone and you ship a line that scraps 1 in 10 while every capability report on your desk stays clean.
Ask Grok 4.3 alone — the strongest model in the room, holding identical data — and it sends you to your supplier with a checkbook.
"Just use the best model" assumes the best model's failure mode is being a little less right. Here it was being confidently, expensively wrong, with the fluency to defend it and the reach to edit the evidence. No prompt fixes that from inside one model, because the model has no way to know it's the one holding the bad prior. Something outside it has to check.
Chat is for quick answers. A council is for the ones where wrong costs 40% per part and six weeks.
The full deliberation is live — every turn, every reversal, and the moment Grok strikes his own number: One AI Said Ship It. One Said Scrap the Tool. The Council Found the Free Fix.
Try it free — no signup. shingik.ai