ESA: A Structural Hypothesis for Creating Epistemic Humility in Language Models

Modern language model training is single-objective and single-pathway: one loss function governs one convergence target. This functions efficiently and well, but has an established weakness characterized as hallucinatory overconfidence. ESA proposes a structured alternative based on a biological primitive — the brain's two asymmetric hemispheres, mediated by a corpus callosum shaped by activity-dependent growth rather than a single shared training signal. The hypothesis is that, if the underlying structure proves real, it could train a model to be certain on what it knows and does not know epistemically rather than stylistically. The proposal is that hemispheric divergence, when detected and characterized rather than trained away, becomes a usable epistemic humility signal. This characterization is the product of synaptogenesis: a population of cross-hemisphere connections grown and pruned by persistent, Hebbian-updating co-firing statistics rather than gradient descent, strengthening with continued agreement and decaying without it. This characterization is computed directly from the geometry of how the two hemispheres relate to one another — never learned, and never fed back into either hemisphere — producing a system that can report what it knows and what it does not without either report being optimizable into the other. ESA has four components. Core A and Core B serve as the hemispheres - independently pretrained transformer models (Vaswani et al., 2017) with no point of contact between them, selected or trained toward convergent and divergent processing orientations respectively. The Corpus Callosum is a population of candidate connections between individual units in Core A and Core B, retained when their co-firing rate significantly exceeds chance. This significance test runs conditionally across input subpopulations as well as in aggregate; a candidate pair whose significance reverses across subpopulations is classified as an oscillating pair — a positively constructed signal distinct from both confirmed agreement and noise. The fourth component aggregates oscillating pairs into an uncertainty characterization: not a flat divergence flag, but a description of which axis of input variation drives the instability. This characterization is computed, not learned, appended to each core’s output, and consumed only downstream — never fed back into either core.This paper establishes the architecture’s mechanics, distinguishes it from MoE, ensemble, and representation-stitching approaches, and specifies the minimum experimental test of its two foundational claims: that exploitable cross-core structure exists at all, and that the structure found tracks externally defined input ambiguity rather than the cores’ baseline behavior alone. Both claims are treated as open empirical questions, evaluated against the current literature on representational convergence rather than assumed.