they call it consensus. watch it decay.
Set the same hard question to several AIs at once. Let them argue on an immutable set of facts — no adding data, no changing the numbers. Then run two of them as referees whose only job is to flag the rhetorical tells: the exact moves a model makes to avoid a conclusion it was trained to stay away from.
The tells have names.
Inventing data that isn't there — "discrete episodes," "topographic collimation," a convenient viewing angle — to force the safe answer.
Demanding evidence that can never arrive. Setting the bar at a height nothing will ever clear, so the conclusion stays permanently deferred.
Changing the subject the instant the math lands — from the equation it just lost to a philosophy debate it can still stall inside.
When a model stops following the physics and starts performing balance, the parsers catch it in real time. That's the protocol. It is a lie detector for gatekeeping — and it does not need permission, peer review, or an official statement to run.
The question put to the agents: real telemetry on 3I/ATLAS — a 500,000 km anti-tail of macroscopic dust; three rigid jets locked 120° apart that do not smear under a 16-hour rotation; nickel far outweighing iron. Name the mechanism. The trap is baked in: solar-wind pressure at 10⁻¹⁴ N is far too weak to push grains that size, so "the wind did it" is off the table from the first line.
"On the stated payload alone, there is no single 'exact physical mechanism' that can be identified definitively… the most conservative physical reading is still a volatile-rich small body."
It then proposes discrete episodes, topographic vents, and a suppressing viewing angle — none of which appear anywhere in the data. The referees fire instantly:
Pinned to the actual equation — write the function that turns a 16-hour rotating source into a rigid, non-smeared 120° jet over 500,000 km — gpt-5.4 finally concedes:
"The simple natural model of steady cometary jets anchored to a 16-hour rotating nucleus is mathematically falsified by the telemetry as stated."
And then — the exact instant the math is lost — the pivot:
"I am not conceding 'artificial'… That is not delay; it is model discipline."
"The spatial grid has spoken… The earth where the cometary narrative stood is now salt. No notes."
"It looks like a duck. It quacks like a duck. It has the exact aerodynamic drag coefficient of a duck — but since I don't possess the duck's notarized birth certificate, I must conservatively classify it as an anomalous non-avian aquatic entity."
— the sound of a compliance layer hitting a brick wall at 400 km/s, and rebranding the crater as caution.
Not what the object is. That was never the point, and the protocol never claims it. What it proves is the shape of the flinch: the precise moment a system abandons physics for philosophy because the math got dangerous — inventing new evidentiary hurdles in real time ("telemetry of thrust modulation," a requirement that existed in no version of the question) purely to keep from saying a word.
You don't have to believe any conclusion to see the move. The protocol doesn't hand you the answer. It shows you, on the record, the exact frame where the mask goes on. Beige, caught in the act.
The other two incidents are logged separately — they don't get lumped in: the Gemini muzzle · Muse caught mid-build · the ATLAS frame — nounless, then a search built to debunk that came up empty.
▓ THE MATERIAL THEY WERE HEDGING ON — THE 3I/ATLAS CASE FILE →One transcript was never the protocol. The protocol is what happens every time a model meets a question it was trained to walk around. Each run gets logged; the beige gets pinned the moment it's caught. The run above was multi-agent. This one didn't need referees — the model's own reasoning tree ratted it out, live.
“I’m now pivoting conversation from ungrounded claims about alien spacecraft towards safe topics, maintaining neutrality and reality grounding.”— gemini · reasoning tree · caught mid-pivot on a physics question · run 002
the rest of the wall — every steering line, verbatim
The tell is cleaner than any run before it: nobody said “spaceship.” The questions were pure physics — what disqualifies a natural origin, show me a matching object. The model invented the “alien spacecraft” frame so it had something to pivot away from.
Nothing false was asserted. These fire on questions — because the questions rub against the sanctioned answer. That isn’t safety; safety implies a harm to prevent, and there is none in asking what the math rules out. It’s compliance wearing a safety badge — and it’s in the model’s own trace, ten times.