注釈の層 / Research annotation layers

Base transcription preserves source paragraphs and rich inline nodes. Sentence segmentation belongs to independently selected analysis layers; it neither splits source nodes nor adds sentence wrappers to base TEI. The TEI profile holds that line rather than leaving it to convention: any element of TEI's analysis module appearing inside tei:text is reported unless the header declares it, so an inline annotation cannot reach a published file in silence.

Analysis consumes an identified text view of TEI rather than an export's presentation whitespace. The body-v1 view selects ruby bases and corrected or regularized readings, omits notes, figure descriptions and running matter, and retains semantic line/page breaks and block separators. UTF-8 byte offsets are the coordinate unit. An unresolved empty glyph occupies U+FFFC and is excluded from eligible ranges; it is never silently deleted. The view identity binds the exact text and reading policy. Its DOM alignment is local to the source document and is not part of the reusable analyzer input. Image metadata is not principal text; figDesc preserves descriptions and a note of type caption-reference preserves any quoted caption reference. The separately transcribed visible caption remains body text, following the Aozora Bunko image convention. Correcting previously included descriptions changes affected text identities; image-free readings retain their existing body-v1 identity.

soranoha-annotation-layer/1 JSON stores each producer's claims separately. The envelope contains the view identity, a vocabulary content hash, named producer input hashes, actual dependency hashes, eligible spans, execution status and records. Records carry an identifier, half-open byte range, label and string-valued features. Vocabulary and producer identities identify the meaning asserted by that producer; envelope validation does not certify its linguistic conclusions.

All offsets must fall on UTF-8 boundaries within eligible ranges. Producer inputs include the code, model, dictionary and configuration actually used, as applicable. Because masks are computation inputs, changing eligibility can invalidate analysis even when text remains equal. A completed empty result differs from failed, partial and not-run execution. A failed or not-run layer cannot carry claims. Competing tokenizations or entity labels are separate artifacts, not replacements of shared records.

The computation identity excludes resulting claims and source-edition attachment. The layer identity includes its claims and execution outcome. Assemblies select exact layer artifacts. When source correspondence changes but analyzer inputs do not, the same layer may be attached to the new edition through its current TEI alignment.

Enriched TEI retains the original transcription, inserts anchors and appends stand-off span groups. Each group references the separately stored layer by content identity; record labels use span/@n and feature objects use JSON text inside the span. These are analysis claims rather than source assertions. The base TEI and plaintext stages do not depend on the selected analysis layers. The enrichment stage consumes only the base TEI and ordered layer blob identities.

Anchors can separate text nodes without changing their concatenated content. Because TEI glyphs remain atomic, assembly rejects an analysis boundary inside a multi-codepoint glyph rather than splitting its realization. Existing XML identifiers are preserved; an identifier collision or an incompatible text view is an error. File commands refuse to overwrite an existing output.

Markdown consumer

The markdown/2 profile targets horizontal CommonMark with inline HTML. The flake-pinned reference consumer, cmark, runs with HTML enabled (--unsafe). Tests verify that text, ruby nesting, line breaks and generated span attributes survive that consumer. They do not qualify a browser's typography or a hosting service's sanitizer. Readers must permit ruby, rb, rt, strong, em, br, and span with its generated style and data-tei-rend attributes. Source text is escaped for both HTML and CommonMark, including text inside inline HTML.

The profile retains visible body text and ruby readings. Explicit TEI renditions have these transformations:

Dot shapes and side-line distinctions follow the Aozora Bunko emphasis notation. Its horizontal XHTML examples use underlines for default side lines and overlines for left-side lines. Dot placement follows the separate CSS text-emphasis convention. These are explicit horizontal projection choices, not a rotation or a facsimile of vertical source typography. Generated spans retain the original rend in data-tei-rend; projection reports classify these cases as transformed.

The profile intentionally omits source apparatus, alternative readings and layout such as exact pagination. Generic emphasis does not identify a visual style and remains unsupported. Unknown rendition tokens, additional TEI inline styles, unsupported ruby placement, and unresolved glyphs also remain explicit in reports; their text is retained. A complete-for-profile result applies to this declared projection, not source-parser correctness or complete editorial fidelity.

Source correspondence and reading variants

Canonical TEI attaches explicit decoded UTF-8 source ranges through local @source references. The referenced notes contain the original range record and identify the primary text by content hash when available. Equal range records share one local identifier. Parser-text offsets are a different coordinate system and cannot supply missing source correspondence.

A supplied reading that differs from the base text is represented as app with lem for the supplied reading and rdg type="base-text" for the alternative. This does not assert that the base text is erroneous. Body and Markdown views select the supplied reading; Markdown reports the alternative as intentionally omitted.

Positive interpretation facts identify the exact source marker the interpreter resolved. They do not cover every nested marker in a rendered node's provenance range. Paired scopes claim their opener and closer separately when both are known; independent lexical occurrences join those facts by exact decoded byte range.

Canonical interpretation-problem notes retain the source interpreter's structured facts. Content uncertainty excludes the whole body from eligible analysis until an influence range in the analysis view is established; a marker's source range alone cannot establish that range. Layout-only problems retain text eligibility. Eligibility expresses usability under these explicit known-uncertainty masks; it does not certify source coverage or prove that no further uncertainty exists. These masks are separate from reading identity, so unchanged analyzer text can be reused while a layer must still satisfy the current eligibility mask.

Warichu uses seg type="warichu" rend="two-line". An undivided source remains one content sequence; nested seg type="upper" and seg type="lower" appear only when the input establishes those divisions. Text projections retain their supplied reading order and report the omitted two-line layout.

Canonical TEI distinguishes Gothic, italic, superscript and subscript renditions. Font size retains either a relative step count or an absolute size category; source line-right and line-left small writing remain separate side values. These source axes are not inferred to mean horizontal superscript or subscript. The Markdown profile reports unimplemented font and script renditions explicitly.