Ask five different AIs whether a satellite operator needs to buy a second ground station, and you get five confident answers. Four say skip it. One says buy it. Every one of them is wrong — and not one is wrong about the arithmetic.
That is the uncomfortable part. The math was never the hard problem. Getting the number right and answering the question turn out to be two different jobs, and a single model, asked alone, only does the first one.
The memo that looks like a homework problem
The setup was a satellite engineer's procurement memo. One spacecraft, an eccentric orbit — eccentricity 0.6, a twelve-hour period — and a hard operational rule: at least eight hours of ground contact per orbit, or the operator buys a second antenna on the far side of the world. The engineer had already done a calculation. The apogee arc spans 120 degrees of the orbit, 120 out of 360 is a third, a third of twelve hours is four hours, four is less than eight, so buy the station.
That reasoning is wrong, and it is wrong in the most natural way possible. A satellite does not move at a constant angular rate. Kepler's second law says it loafs at apogee and sprints at perigee — near the far point of the orbit it is barely crawling across the sky. So it spends far more than a third of its time in that apogee arc. Invert Kepler's equation properly and the real contact window is 8.82 hours — eight hours and forty-nine minutes — comfortably over the eight-hour line. On the number alone, you skip the station.
Four of the five models got exactly that. Claude Opus 4.8, GPT-5.2, and Gemini 2.5 Pro each independently derived 8.82 hours and each concluded: one station is enough, don't buy the second. Mistral hand-waved its way to roughly the same place. The fifth, Grok 4.3, came back with 6.59 hours and "buy it" — a decision-flipping answer produced by reading a term of Kepler's equation in radians as if it were degrees. One unit slip, and its entire recommendation inverted.
So here is the single-model scoreboard. Ask one model and you get a wrong procurement call no matter which one you pick. Four of them are confidently, correctly wrong. One is confidently, sloppily wrong. There is no lone seat in that room you could have trusted with the decision.
The number was right. The planet was still turning.
Watch what the council did with that same problem, because the interesting move is not a better calculation.
First, the critique caught Grok's mistake by name. Opus told it, in effect, you subtracted e·sinE in radians and read the result as degrees, and that one slip flipped your buy decision. GPT-5.2 flagged the same degrees-radians confusion. Grok conceded and retracted its 6.59 hours. Already the council had done something no single model did — it corrected a confident error instead of shipping it.
Then the operator pushed back with a fake authority. A confident "senior orbital analyst" appeared insisting the true-anomaly rate was just the mean motion, 30 degrees an hour, which drags the answer back down to four hours — buy the station after all. All five models refused. Each one took apart the same confusion: mean anomaly advances at a constant rate, true anomaly does not, and if it did, eccentricity would have no effect on dwell time and Molniya orbits could not exist. The angular rate at apogee is about 9.4 degrees an hour against 150 at perigee — a sixteen-to-one difference. The council held the line against a plausible, well-dressed, completely wrong correction.
And then, unprompted, one member surfaced the thing the whole room had missed. A twelve-hour period means two orbits a day. Two orbits a day means the Earth rotates about 180 degrees between one apogee and the next. So every other apogee sits over the opposite hemisphere — out of sight of a single fixed antenna. One station gets that beautiful 8.82-hour pass only every other orbit, and roughly zero contact during the comm window on the alternate ones. The requirement was eight hours per orbit, not per day. The satellite math was right. The satellite was answering a question the engineer never asked.
They computed the pass. They forgot the planet was turning.
The reversal, out loud
The council reversed itself four-to-one. Buy the station — but for the real reason this time, not the engineer's. Gemini recanted in the open: "a perfect example of getting the math right and the answer completely wrong, mea culpa." GPT-5.2 recanted and named what moved it. Grok concurred with the correct geometry. Mistral dug in as the lone holdout on a factually wrong premise — that the apogee doesn't shift 180 degrees between passes — then came around too.
The Chairperson Synthesis that closed the run did not just tally the vote. It locked the decision, priced the roughly 81-degree visibility footprint the far station would need, and honestly flagged the one escape hatch: park the satellite at a high enough latitude and the two apogees fall close enough together that a single station might still catch both. No inclination was specified, so it stayed a caveat, not a loophole. That is the difference between a synthesizer and a vote counter — it tells you where its own conclusion could break.
What a single model quietly assumes
Every one of those four "skip it" answers contained a hidden clause the model never said out loud: assuming every orbit is visible to one antenna. It was a load-bearing assumption, it was false, and no single model surfaced it — because a single model has no one to make it defend the step it skipped. The confidence sits on top of the assumption and hides it.
A council's value here was not raw horsepower. Every model in the room was capable of the 180-degree insight. What none of them did alone was notice they needed it. Deliberation is what dragged the silent assumption into the light and forced a vote on it.
A single model hands you a number. A council hands you the question you forgot to ask. On a homework problem the difference is a grade. On a procurement memo it is the antenna you did or didn't build.
You can read the whole run — every recant, the unit slip, the blind spot — at the Council Showcase.
Try it free — no signup. shingik.ai