科研 Skill

版本校勘与谱系重建

textual-criticism-collation

Collates manuscript witnesses, builds a critical apparatus, and reasons toward a stemma codicum. Use when editing a text transmitted in disagreeing copies.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Textual Criticism and Manuscript Collation

Produces a critical edition workflow: align multiple witnesses (manuscripts, early prints, versions) of the same text, record every point of disagreement as a variant reading, and reason about how the witnesses relate to each other (stemma codicum) to recover the most likely original or earliest recoverable text. Classical philology, biblical/patristic studies, medieval literature, and any field editing a text that survives in more than one copy.

When to use

  • A text survives in two or more manuscripts, early editions, or translations that disagree, and the user wants a critical edition, an apparatus criticus, or a reasoned judgement about which reading is original.
  • Someone asks "which manuscript is more reliable here?", "build a stemma for these witnesses", "what are the variant readings at this line?", or "collate these two editions".
  • Preparing a base text (constitutio textus) for close reading, translation, or further argument when the received text is contested.

NOT for: OCR/digitisation of a single physical document (use historical-ocr); encoding a text in TEI once the edition is settled (use tei-xml, which can represent the apparatus this skill produces); single-author stylistic comparison (use stylometry-authorship-attribution).

Prerequisites

  • Transcriptions of every witness to be collated, as clean plain text, ideally segmented the same way (verse/line/chapter numbers, or a shared reference system) so that corresponding passages can be aligned.
  • Basic familiarity with the language and script of the witnesses — collation software proposes alignments; a human (or an agent working under review) must judge whether a proposed variant is a genuine textual difference or an orthographic/spelling non-issue.
  • pip install collatex (Python wrapper for the CollateX engine) or a local Java runtime for the standalone CollateX tool (https://collatex.net/), which this skill treats as the reference collation engine. pip install python-Levenshtein is useful for lightweight fuzzy alignment when CollateX is unavailable.

Step 1 — Prepare witnesses

Segment every witness into comparable units (typically one unit per verse, line, or sentence, matched by a shared numbering scheme). Record for each witness: siglum (a short conventional label, e.g. A, B, P45), date or date range, and provenance. Keep orthographic normalisation decisions explicit and consistent — e.g. whether u/v or medieval abbreviations are expanded — and log the normalisation rule once rather than deciding ad hoc per word.

Step 2 — Align and collate

Run automatic collation per unit, across all witnesses, to find candidate variant locations:

from collatex import Collation

collation = Collation()
collation.add_plain_witness("A", "the quikke and the dede")
collation.add_plain_witness("B", "the quick and the ded")
collation.add_plain_witness("C", "the quicke and dede")

table = collation.collate(near_match=True, segmentation=False)
print(table)  # aligned columns, one row per witness

For texts in non-Latin scripts or with heavy abbreviation, pre-normalise spelling variants that are not meaningful (e.g. known orthographic doublets) before collation, or the apparatus will drown genuine variants in spelling noise — see "Quality checks" below.

Step 3 — Classify each variant

For every point of disagreement the collation surfaces, classify it:

TypeMeaningExample
orthographicspelling/abbreviation only, not a textual variantquikke / quick
substantivea real difference in wording, word order, or presence/absenceomission of a clause
conjecturalan editor's proposed emendation, not attested in any witness[cruces: reading uncertain]

Only substantive variants belong in the apparatus criticus; orthographic ones are normalised silently (with the normalisation rule documented once) unless the edition's stated policy is to preserve all spelling.

Step 4 — Build the apparatus

Record each substantive variant unit as:

<lemma reading> ] <variant reading> <sigla> : <variant reading 2> <sigla>

e.g. quick ] quikke A : quicke C reads as "the base text's 'quick' is read as 'quikke' in witness A and 'quicke' in witness C". Keep the lemma (chosen base reading) and its justification separate from the raw variant list — which reading is adopted is an editorial judgement, the apparatus itself is a record of fact.

Step 5 — Reason about relationships (stemma codicum)

Variant distribution is evidence about how witnesses are related:

  • Shared errors are diagnostic; shared correct readings are not — two witnesses independently "fixing" an obvious error is common; two witnesses sharing the same nonsensical or demonstrably wrong reading is strong evidence they share an ancestor (a shared exemplar) that already had that error.
  • Build a matrix of witness-by-variant-unit and look for witnesses that consistently share errors — these cluster into families.
  • Draw a stemma (a family tree) with the extant witnesses as leaves and hypothetical lost ancestors (archetypes, denoted α, β, ...) as internal nodes, each split justified by at least one shared diagnostic error.
  • Contamination (a scribe consulting more than one exemplar) breaks a clean tree — flag witnesses whose error pattern does not fit any single branch cleanly, rather than forcing a tree that does not fit the data.
  • For large witness sets, phylogenetic software built for biological trees (e.g. PAUP*, or scipy.cluster.hierarchy on a 0/1 variant-presence matrix as a rough first pass) can suggest groupings, but the diagnostic judgement about which variants are genealogically significant stays with the editor — software clusters by similarity, not by inferred error direction.

Step 6 — Choose the base text and report

  • State the editorial method explicitly: "best manuscript" (follow one witness, apparatus records deviations) or "eclectic" (choose the most likely original reading unit by unit, drawing on the stemma).
  • For every lemma chosen over a competing variant, give the reason: stemmatic priority, lectio difficilior (the harder reading is more likely original, since scribes simplify more often than they complicate), internal sense, or editorial conjecture — and label conjectures as such.
  • Produce: apparatus.md (the variant list), stemma.md or a tree diagram (witness relationships and the evidence for each split), and edition.md (the constituted text with inline apparatus references).

Quality checks

  • Every substantive variant in the apparatus must be traceable to a specific witness and location (siglum + unit number) — no "some manuscripts read..." without naming which ones.
  • Orthographic-only "variants" must not appear in the substantive apparatus; spot-check a sample of automatically flagged differences to confirm the orthographic/substantive split was applied consistently.
  • Every stemma split must cite at least one shared error, not merely a shared correct or plausible reading, as its evidence.
  • Re-run the collation after any normalisation change and diff the variant list — normalisation choices should shrink noise, not silently delete real variants.

Limitations

  • Automatic collation proposes alignments; it does not know which variant is "original" or classify orthographic vs. substantive differences — that judgement requires philological expertise in the specific language, script, and textual tradition, and this skill should be used under a domain expert's review, not as an unsupervised judge.
  • Stemmatics assumes a reasonably tree-like transmission; heavily contaminated traditions (common in liturgical and legal texts that scribes actively cross-checked) resist a clean stemma, and forcing one misrepresents the evidence.
  • CollateX and similar tools work on segmented plain text; manuscripts with significant physical damage, marginalia, or uncertain readings need that uncertainty recorded (e.g. [...] for lacunae) before collation, or gaps will silently misalign the comparison.
  • This skill does not perform palaeographic dating of the witnesses themselves — see paleography-manuscript-transcription for script-based dating and transcription of the physical manuscripts.

示例提示词

帮我比对这几个抄本的异文并推断它们之间的传抄谱系关系。

所属插件

人文学科面向历史、哲学、语言学、考古与艺术研究:TEI/OCR 与数字档案、论证分析、语料库与语音学、放射性碳测年、乐谱与音频分析,以及人文论证写作与审稿。

安装该插件会一并获得同领域的 62 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

版本校勘与谱系重建能做什么?

Collates manuscript witnesses, builds a critical apparatus, and reasons toward a stemma codicum. Use when editing a text transmitted in disagreeing copies。

如何安装版本校勘与谱系重建?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「版本校勘与谱系重建」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

版本校勘与谱系重建需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

版本校勘与谱系重建经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。