Answers are grounded in this paper and cite their sources.
System evaluation demonstrates high parse rates and verifiable traceability across multilingual corpora, indicating that structured book formatting mitigates machine reading errors.
Widespread book formats serve human readers, publishing pipelines, or pretraining scale — none serves machine-learning consumers that must audit their inputs. We present Book-YAML 1.0, a minimal book format for AI reading engines: verbatim text with per-file SHA-256 provenance, stable block identifiers enabling sentence-to-source traceability, explicit prose/verse block typing, curriculum metadata (grade, role, genre), and a language-invariant schema. Every design rule traces to a concrete production failure in the Noemica reading engine: hallucinated LLM extraction (1,135 fabricated facts), silently dropped verse content (8 books read as empty), and encoding corruption (1,431 repaired blocks). Production evidence: 39 books, ≈2.94M words across Hungarian, English, and German — 100% public domain — with 99–100% parse rates and a cross-language invariance measurement (DE↔HU Grimm pairs) that the format made possible.
No takes yet. Share an insight, caveat, or question.
László Popovics (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: