SayFable Books: 31 annotated novels in the .sayfable format (books-v1)

SayFable Books is a corpus of 31 public-domain English novels with full-book structural annotation, distributed in the .sayfable v1 package format. Each package contains the source text with character offsets and a multi-layer annotation: scenes (boundaries, label, mood tags from a closed vocabulary, and a passport of who / where / when / what it leads to / causal weight); speaker attribution at both paragraph level and individual quotation level, with character offsets; a character registry (aliases, gender, age band, narrative importance) and relations between characters with polarity and turning points; chapter narrators and a book-level narrator profile; chapter summaries and narrative weights; and a book card (plot type, narrative arc, protagonist, fortune curve). Character cards cite paragraph addresses rather than paraphrasing the text. Corpus totals: 52,850 paragraphs, 5,698 scenes, 1,667 characters, and 14,513 quotations with character offsets, across 31 novels (10.6 MB). Every annotation layer carries per-unit provenance: which pipeline pass produced it, which model, the prompt hash, the status of each unit, whether the response was truncated, and the monetary cost. A package validator is included; it refuses non-conforming packages rather than warning about them. EVALUATION. Speaker attribution was compared against the human-annotated Project Dialogism Novel Corpus (Vishnubhotla et al., LREC 2022) on the seven overlapping novels: 95.8% agreement over 7,013 quotations. Caveats are stated openly. Agreement is computed over the 81.1% of PDNC quotations that could be located verbatim in our editions, since editions differ; whether the remaining 19% form a random or a systematically harder subset has not yet been assessed. The benchmark novels are skewed toward dialogue-transparent nineteenth-century fiction, and the hardest of them scored lowest at 91.6%. A single annotation model was used, with no cross-model comparison. RIGHTS. All source texts are in the public domain in the United States. Each package carries per-jurisdiction rights fields, including EU and Spanish expiry years, because a pre-1931 publication date is not by itself evidence of public-domain status outside the United States. LICENSING is three-layered: the source texts are in the public domain; the annotation layer is CC BY 4.0; the accompanying code (benchmark harness and package validator) is Apache 2.0. Repository, format specification and reproduction scripts: https://github.com/kubataba/sayfable-books