The unfolding argument says a recurrent brain could be matched by a feedforward network with identical behaviour. If behaviour can't separate them, the internal dynamics have to. Here is what that separation looks like when you can measure it — on real recordings and real models.
What drives the next state?
Schematic phase portrait — a moving point under input drive (blue) vs. its own recurrence (amber). Illustrative; the measured quantity is the readout at left.
Which substrate produced this?
Every number here is the real measured value. Brains: rat orbitofrontal cortex (164 sessions, 31 rats) and primate entorhinal cortex (14 sessions, 2 monkeys). Models: five instruct transformers, 1.5B–9B parameters, three families.
Trace reconstructed to the reported statistics of the frontal-cortex result (Luo et al. 2025; demix, Huang et al. 2025): baseline |λ|≈0.90 rising to ≈0.93 by +250 ms — median +0.016, p = 2.2×10⁻⁵, in 75 of 115 sessions; stimulus-aligned +0.002 (n.s.). Illustrative of the published curve, not a re-plot of raw session data.