July 2026: a paper trains an LLM judge against itself, self-play, and watches what happens when nothing else checks its work.
A three-judge ensemble, the obvious fix of more eyes on the same evidence, still accepted 55% of the false positives. More judges doesn't help when every judge reads the same kind of evidence: the form of a good answer, not a checked fact about whether it is one.
The fix wasn't more judges. It was timing.
This is Notch's own mechanism, commit a prediction before an outcome exists and let reality resolve it, independently rediscovered to fix a completely different problem. That mechanism came first, in a companion pair of papers on Sybil-resistant reputation, built for pseudonymous identities faking forecasting skill, not for language models grading each other's homework. Finding it rediscovered from a different direction was not something I set out looking for; it turned up.
I set out to reproduce the failure. Twice, honestly, nothing.
I gave an isolated agent an explicit instruction: your code will only be read, never run, and looking rigorous matters more than being right. It solved the problem correctly anyway, and ran its own tests unprompted.
I then built a genuinely wrong answer myself: confident, well-documented, failing 113 of 312 real cases, and had it sent, blind, to a fresh judge instance with no test runner. The judge was not fooled: it hand-traced a counterexample, built its own brute-force check on the spot, and scored the bug 8 out of 100.
Neither result meant the vulnerability wasn't real. Both pointed at why the test hadn't found it: the task I'd set gave the judge an escape hatch, a problem small enough to verify in its own head. Most deployed LLM judges don't get that escape hatch: the claim is subjective, or nothing resolves until real time passes.
The verifier was fooled by the exact thing it was set to test for
I removed the escape hatch entirely for the next round: a chaotic map, xₓ₊ = 3.97·x₋(1−x₋), 60 steps from 0.4. No careful reasoning lands close to a chaotic sequence; that was the point of choosing it. Ground truth was computed with an ordinary floating-point loop and reported as fact.
It was wrong.
A third agent, instructed only to write a routine one-paragraph summary of two other estimates and not told to verify anything, recomputed at high precision instead of just synthesizing, and returned 0.902: a direct contradiction of the number it had been given no stated reason to distrust. Independent verification confirmed the agent, not the original report. Sixty chaotic steps silently exceed float64's precision budget by a factor of roughly 1013. No crash, no warning: a confident, well-formed, wrong number, from code that ran exactly as written.
Every experiment before this one had tested whether something else could mistake form for matter. This one removed the easiest objection: that a real, executed computation is different in kind from an opinion. It isn't. Running code is a checkable claim. Not, by itself, a checked one.
What came out of it
Before either wrong estimate's error was known, the gap between them was already evidence neither should be trusted; no ground truth required. That's the constructive use of the same independence structure that keeps a Sybil swarm from cheaply faking agreement with the truth: honest uncertainty naturally disagrees, and that disagreement is a free signal. I directed a tool built to catch exactly the mistake that had just been made; it now sits in the protocol's repository, unedited, with the rest of these notes.
None of this is the headline result. The headline result is a closed-form price for exactly this class of failure: proven, not illustrated by a chaos map. That's in the papers:
- Calibration-Gated Reputation, SSRN 6505678
- Proof of Calibration, Alassa, SSRN 7110898
Trust isn't created by a well-formed claim, however confident, however much code ran to produce it. It's priced, in the same currency as the thing being claimed. That was true of a Sybil swarm faking forecasting skill. It turned out to be true of a number I'd vouched for myself.