5Conclusion
- Looping wins where it should. A fair, paired, well-powered test shows the recurrent model out-encodes a deep feed-forward one in higher-order language cortex (IFG, aSTG) and not in auditory cortex — recurrence's benefit is language-specific, surviving ≥3 independent validations and a 4-attack adversarial skeptic with zero sign-flips.
- Mechanism (exploratory). Within the loop, the across-pass recursion step — not ordinary depth — tracks the brain's processing-time hierarchy, hinting that cortical language computation looks more like an unrolled loop than a deeper stack.
Methods note. CPU-only re-analysis on cached, frozen encodings (no GPU run); peak read-out unified to sub-bin parabolic; TP (n=6) labeled underpowered; the step/layer panel is exploratory. A dense 192-rep re-confirmation (4 passes × 48 layers) is pending as a deferred GPU job.
Scope: one encoding metric, this cohort (9 patients), this scale. Every reported number carries its frozen source file + figure_spec; data + figures independently checked.