原文の説明責任 / Source accountability and validation

source-accountability.json records lexical occurrences found directly in the retained source. It binds the original bytes, decoded UTF-8 text and coverage matrix by hash. Each occurrence retains its exact spelling, decoded byte span, source region and recognized syntax families. Editorial legend examples remain visible as front matter. Unrecognized markers and lossy decoding remain explicit.

This scanner runs independently of the parser. Its stage identity binds the scanner executable and matrix; changes confined to the parser stage do not invalidate its evidence. Changes to shared scanner dependencies can invalidate it. The native command is ab-source-inventory --source source.txt --matrix matrix.toml --output-json accountability.json. The publication wrapper supplies AB_SOURCE_INVENTORY_BIN and AB_AOZORA_SYNTAX_MATRIX.

Lexical recognition does not establish a correct interpretation. Interpreter claims describe what the parser recognized; missing claims remain unaccounted. Neither source preservation nor an empty problem list proves semantic coverage. Source conformance tests exercise expected values and annotation placement for specific examples. TEI profile validation checks the exported XML against the profile. These are separate kinds of evidence.

A report about one edition does not compare its exported values independently against every source construction. Conformance tests support the examples they exercise; structural validation and interpreter claims are not equivalent to a per-edition source/export fidelity certificate.

interpretation-coverage.json joins this lexical evidence to explicit native interpretation facts. Family counts distinguish interpreter_claimed from unaccounted and from outside_claim_vocabulary; a claimed occurrence has at least one explicit compatible aspect, not necessarily every aspect. Each claim must match the exact span and a compatible family of an independently scanned occurrence or nested component. Nested markers retain their own component spans. Source apparatus is kept separate. Lossy decoding prevents positive claim accounting. Reported interpretation problems retain their original influence bounds alongside positive claims. A positive claim is not evidence that an overlapping or document-wide problem has been resolved.

The report states the counts and lists the exceptions. families and interpretation_problems cover every occurrence; occurrence_count and unclassified_occurrence_count give the totals; unaccounted_occurrences holds, whole, each occurrence with an unaccounted family or no recognized family at all. A claimed occurrence is not listed, because the correspondence is already recorded from the other direction: the TEI element the claim produced carries that occurrence's span in its source attribute. Nothing else the build writes records which occurrences went unaccounted, so those are kept in full.

A family can only be claimed on a marker kind some interpretation fact admits. The scanner recognizes kinds no fact kind mentions, among them the two forms of 〔…〕 and the terminal provenance line, and a family the scanner assigns to one of those is unreachable rather than missed. Those are counted under outside_claim_vocabulary and are not listed as exceptions, because what explains them is the marker kind and not the individual occurrence. Reading unaccounted as a proportion of occurrence_count therefore measures interpretation coverage and not the distance between two vocabularies.

Review exports

From the repository root, set these variables to absolute paths. The corpus must be a clean Git checkout and the export directory must not exist:

nix run .#soranoha-kernel -- build \
  --root "$BUILD_STORE" \
  --aozora-root "$CORPUS_CHECKOUT" \
  --assets-root "$PWD/soranoha" \
  --out "$EXPORT_DIR"

Each work directory contains tei.xml, plain.txt, text.md, tei-validation.json, plaintext-projection.json, markdown-projection.json, source-accountability.json and interpretation-coverage.json. A top-level build.json records artifact hashes and build results. Omitting --out retains outputs in the computation store. Building exports does not publish them or accept an assessment.

TEI retains ruby, gaiji declarations, headings, layout and source apparatus. Supplied kunten uses note type="kunten": subtype="return-mark" carries subscript rendition and subtype="okurigana" carries superscript rendition. Annotation text omits syntax parentheses; source spans retain exact notation. These notes preserve their insertion position and enclosing context without inferring a host character or spoken reading. The principal-body text view excludes notes, including kunten, while TEI retains their supplied text. Resolved gaiji contain Unicode text directly in g, including inside ruby bases; ref and charDecl retain source mappings. Plaintext contains visible body text. Paragraph-leading indentation is encoded as CSS text-indent; heading and source-note continuation indentation uses padding-inline-start. Structural XML is indented for inspection while mixed content retains its lexical whitespace.

Complete iteration notation ／＼ and ／″＼ retains its source spelling in choice/orig; choice/reg contains the whole Unicode mark 〱 or 〲. The body view, plaintext and Markdown select that realization without expanding the repeated reading. A mark inside ruby reading remains part of the ruby annotation. Malformed notation and literal Unicode upper/lower halves remain literal text; iteration marks are distinct from gaiji.

Supplied diacritic annotations apply to native-resolved target text. Their choice/orig retains the source spelling and choice/reg contains the composed Unicode text selected by the body view, plaintext and Markdown. An annotation addressing an interior letter retains its original marker separately as source apparatus. This is transcription of supplied marks, not an editorial correction or an inferred transliteration. Unresolved selectors retain explicit uncertainty.

An emphasis selector may name the range to mark or the range to leave out. ［＃「自律・自由・人格・性格」の「・」を除く部分に傍点］ marks every part of the target the named run does not cover, once for every occurrence of that run rather than the first alone, and the cut-out characters remain unmarked body text in their source position. The claim covers the marker, not the ranges the marker points at. A run the target does not contain, or one that covers the whole target, leaves the marker unread rather than marking a range the source did not select.

A marker with two operands can also state a position rather than a range. ［＃「自」と「然」の間に白三角傍点］ places a mark at the junction of two adjacent characters and marks neither of them, so it publishes a supplied mark anchored between them rather than emphasis over a span. The mark's side and its anchor stay separate: every such mark sits on the right, and の間に says where along the line rather than which side of it. The note carries the source's own word for the mark, with the extent that word occupies in the marker. Operands the source names with anything between them, or in an order the text does not have, locate no junction and leave the marker unread.

Supplied block geometry retains indentation, continuation indentation, line width, and distance from the line end independently. Hanging indentation uses the continuation padding and the difference between first and continuation indents. TEI retains supplied line counts as rend="line-count(N)" and placement between the physical left and right page edges as rend="page-horizontal-center". Page placement and text alignment remain independent. Consecutive indentation instructions replace the preceding indentation scope. An indentation close that restates its measure, as ［＃１字下げここまで］ and ［＃１字下げ終わり］ do, ends the scope that measure opened. A restated measure that disagrees with the open leaves the pair unmatched rather than closing at a boundary the source did not name. A close that names scopes no block opened, such as ［＃字下げ、地付きここまで］ following line-scoped 次行は…で instructions, stays retained source rather than ending scopes that do not exist. Page placement ends before the next supplied page break, which remains in TEI. Frame styles and co-applied typography share the same formatting vocabulary on TEI div and hi scopes. Supplied solid and dashed rules retain their border style; framing alone does not establish an independent embedded text. A supplied table role is retained on the scope the source supplies it for, both on its own marker and as a clause of a compound indentation marker, alongside that marker's independent indentation, direction and frame. TEI records it as div type="table", and 図表 as type="figure-table", because that spelling names an exhibit without choosing between a diagram and a table. The enclosed lines keep their own spelling and spacing; no cell, row or column structure is derived from them. A supplied 罫無し is retained as rend="table-rules(none)", an explicit absence of table rules distinct from a marker that says nothing about them, and independent of any frame drawn around the scope. A base-edition table description remains source apparatus and is not executed as current formatting. A standalone table opener carries the axes it names beside the role in the same way: ここから表罫囲み an enclosure and ここから横組みの表 a writing direction, with no indentation invented for either. 罫仕切り is not read as a table: it states a reading convention over the lines that follow, rather than delimiting the partitions it names, and the rules that bound them are literal text in the source. A supplied quotation or letter role is read the same way when the source writes it ahead of the measure that scopes it, as ［＃ここから引用文、３字下げ］ and ［＃これより手紙文、１字下げ］ do. The role and the indentation are both retained on that one scope, and TEI records the role as div type="quotation" or div type="letter". The two stay apart because the source spells them apart: a letter is a document the text reproduces, not a passage it quotes. Nothing about the quoted passage's own source, or a letter's correspondents, is derived from the marker. A quotation's close is matched only against a scope carrying that role, so a closer never ends a plain indentation scope and thereby report its lines as quoted matter. Clauses the role reading does not cover stay retained source: a supplied vertical gap such as ３行アキ, and a line-subset exception such as はじめの「一」のみ２字下げ. A line-scoped foot measure can carry a qualitative type size beside it, as ［＃この行、下揃え、下から５字上げ、相対的に字が小さい］ and ［＃この行はポイントを下げて、地より２字上げ］ do. The measure and the size are retained on the same line, and no magnitude is derived for the size: the source states a direction and no degree. A marker that names a target for the measure, or that has lost its target and leaves a bare は, stays retained source, and so does one that changes treatment at its own position inside the line rather than stating the whole line's. Qualitative font comparisons preserve direction and any supplied やや or ひとまわり qualifier, without a numbered size level. TEI records these as font-size qualitative(smaller) or font-size qualitative(larger) and an optional qualifier(...); explicit bold weight and gothic typeface remain separate attributes. Bare 下付き preserves lowered placement without small type, while 指数 preserves the exponent role without identifying a mathematical base. TEI marks the original target glyphs with baseline-lowered or exponent; plaintext and Markdown retain those glyphs unchanged. Markdown uses inline HTML with inherited font size for the placement; explicit 下付き小文字 remains distinct. Presentation supplied on an indentation opener can end before the indentation. Alignment supplied within a horizontal-writing clause ends with that clause; an unclosed indentation remains explicit uncertainty. Relative placement keeps its source anchor and supplied character offset separately from absolute indentation. TEI links the placed scope to the anchor's source-span record with corresp, records placement-below, anchor-kind(text) or anchor-kind(horizontal-block), and any anchor-offset-chars(N) in rend, and preserves supplied writing direction and alignment. A missing or multiply matching literal anchor remains unresolved. A vertical label below an established horizontal block ends that horizontal presentation without ending independent surrounding indentation.

Validation provenance binds the ODD, RNG and Schematron hashes through schemas/tei-profile-generation.json. See TEI validation.