Imagining without Checking

Does a language-model agent check an equivalence before extending it?

Zain Ali (Alexein AI)

2026 · Written and audited; prepared for an AAAI-27 workshop, Montréal, February 2027

Hassabis has proposed testing general intelligence by giving an AI system a 1911 knowledge cutoff and seeing whether it finds general relativity. That test cannot be run, because every model has read the answer. We build a runnable analogue of one step, Einstein’s 1907 equivalence argument, and ask a narrower question: does an agent check the premise of an equivalence before extending it? In procedurally generated universes with alien vocabulary, a textbook theory is exact on every experiment the agent can run, and the question asks for a clock-rate ratio in a regime no experiment reaches. The two universes of a test pair have byte-identical texts; in one, the equivalence fails for one kind of matter, which one available experiment reveals, so no policy that skips it can exceed balanced accuracy (BA) 1/2. In a preregistered study, GPT-6 Luna used its whole query budget in the median episode but ran that experiment in only 2 of 140 episodes, extended the equivalence in every answer (BA 0.479 [0.450, 0.500]), and afterwards asserted that all kinds respond alike in all 138 episodes that gave a verdict. Asked directly, in an extension added after these results and locked before its runs, it ran the experiment in 140 of 140 episodes and gave the right verdict in 139 (BA 0.993 [0.979, 1.000]). It can run the check when told which comparison to make, but almost never makes it unprompted; whether checking would correct its answers is untested. The setup is recognisably relativistic, so the equivalence may be recalled rather than invented. A 3.3M-parameter PPO agent never made a correct jump after its first evaluation. The instrument, the preregistration with its 29 amendments, every transcript and the scripts that re-derive every number are released with the paper.