Counterfeit Coherence: The Two Failure Directions of Narrative Regularization

A convergent literature now reports that machine-generated narrative is structurally impoverished. Russell et al. (2026) find humanauthored stories occupy a mean rarity percentile of 0.71 in narrative feature space against 0.49 for five frontier models, with structureonly detection reaching 93% accuracy. Concurrent work characterizes post-training as narrative regularization — a suppression of thematic, affective, and stylistic variation — and offers this as a structural account of why aligned fiction feels mechanical. A parallel literature in epistemology reports the analogous failure in reasoning: coherence and grounding are separable, and fluent explanation may be independent of, or inversely related to, genuine understanding.Across both literatures, regularization appears exclusively as pathology. Constraint suppresses variation; therefore constraint is the problem. This paper argues that this framing is incomplete in a way that matters, and that correcting it yields a design framework rather than a diagnosis. Regularization has two failure directions, and the literature has measured only one. Under-regularized narrative does not become richer; it overfits — becoming maximally resonant for a single reader and impossible to hold in common. What regularization purchases, at the right weight, is shareability: the condition under which ten thousand readers enter the same world and can afterward speak of it as one thing.I name this third property Belonging, and set it alongside Aliveness (the generation of prediction error) and Coherence (the provision of a resolution path for it), grounding all three in predictive processing. Expressed as a regularized objective, the authored constitution of a world functions as the regularizer, and its weight λ becomes a design parameter with an optimum rather than a minimum. λ too high yields airlessness — the flattening the literature has already measured. λ too low yields fragmentation — a failure mode nobody has named because nobody was looking on that side of the curve.I further argue that the high-λ failure has been mischaracterized. It is not experienced as boring. It is experienced as competent. I name this counterfeit coherence — the feeling of pattern without the contact with reality that makes pattern true — and argue that its danger lies precisely in being asymptomatic. I propose narrative rarity percentile against a reference corpus as a candidate operationalization of Aliveness, sketch a two-instrument evaluation procedure, and state four falsifiable predictions, one of which the flattening literature has explicitly identified as an open gap.