An Ensemble Reduction Engine for Empirical Discovery of a Finite Structural Grammar of Written Prose

This deposit is a snapshot of the Reduction Engine project at version 1.4. The project investigates whether written prose rests on a small, finite catalog of structural templates discoverable by unsupervised methods. Across fifteen experimental iterations on ~1,600 documents in eleven written-prose genres and four languages, an ensemble of five independently-trained large language models is used to reduce each document to a structural skeleton via majority vote, and HDBSCAN discovers a finite catalog without being told how many clusters to expect. Headline findings: (i) five architecturally-distinct models converge on the same skeleton for a given document; (ii) skeletons survive translation across English, German, Chinese, and Japanese; (iii) HDBSCAN finds a finite catalog with zero cross-genre confusion across 156,000+ tested pairs; (iv) the catalog is reproducible (ARI 0.633 ± 0.008); (v) recursion stacks at least three levels of abstraction and terminates in a super-skeleton matching Aristotle's classical rhetorical disposition; (vi) a binary wire format achieves 3–7× compression over original prose; (vii) the codec output is modality-agnostic; (viii) typed-slot architecture beats open-vocabulary induction and pointer/copy-span on Minimum Description Length and LLM-judge fidelity; (ix) the 14-tag vocabulary empirically decomposes into ~4–6 higher-level semantic primitives; (x) the super-skeleton survives an adversarial-corpus test (liturgical Collects, recipes); and (xi) under a discrimination test, the super-skeleton decomposes into a ~5-element universal boundary-marking primitive plus 1–3 domain-specific extensions. Self-corrections: Iter 7 retracted an over-claim about hierarchical sub-genres after external review caught a tautology risk; Iter 12b reversed an internal reframe of Iter 9's transposition test; Iter 15 replaced the universal-7-element claim with a two-layer model (5-element universal + 1–3 goal-specific extensions). Scope: All tested documents are written prose. Non-prose modalities (painting, music, gesture, facial expression, ritual) are entirely untested and represent the project's primary future-work axis. Repository: https://github.com/svenschlegel/reduction