📋 The verbatim digest

Status: reference.

Written for someone studying a topic who wants the papers' own words in one place, and for whoever reads the report that checks them. Assumed: GENRE.md for how the other genres differ, REVIEW.md for what every review aid has in common, and DOSSIER.md for the claim:/quote: contract the digest reads the other way round. Not covered here: the matcher's normalisations, which are CLI.md's and plans/c3-quotation-integrity.md's, and the review layer's reasons for never gating, which are ARCHITECTURE.md's.

A verbatim digest is a survey written in the sources' words. It has a survey's shape -- a reader and a scope, a framing, one section per theme, a comparison table, the disagreements made visible, a gap analysis -- and a survey's paragraph, which braids passages from two or more papers around what they agree and disagree on. Its body is copied exactly from parsed PDFs in the corpus; each copied run ends with a citation; there are no quotation marks; equations are typeset; and the drafter's own sentences are few, short, and there to compare, contrast or define. It reads on its own: someone who has opened none of the papers can follow it. One person reads it to learn a topic. It is not a deliverable, and it is never a source for drafting a survey, a chapter or anything else, because its wording is the corpus's own.

Three goals, in this order:

1. A self-sufficient survey the reader can learn the topic from.
2. As much of the body as possible in the sources' exact words.
3. Surface and minimise unsupported text. Every sentence no source backs is listed for a person to judge; each repair pass leaves fewer such words than the one before, and a baseline recheck makes that a number.

The order is #1172's correction. The first digest the skill produced put "maximise copied text" first and scored 0.88 copied: 248 runs of about forty words each, in thirty subsections, with the parser's noise kept on purpose and the equations the runs explained left out. It was faithful and unreadable. Copying is the means; the survey is the end.

A digest is a draft like any other. The user names it, content/drafts/<topic>/<name>.md; its dossier's scope.md records - genre: digest, which is the only thing that makes it one; it is rendered beside the topic's other drafts; and it is per-host data under content/ like every draft, so the blanket rule in .gitignore keeps it out of git without a rule of its own.

🧭 Table of contents

• Writing one: review-digest
• Reading the report: review digest
• Pages, and what a page note means
• The repair loop and --baseline
• Lexical support, and what the NLI support aid adds
• Figures
• What it does not do
• The rule it lives under

✍ Writing one: review-digest

The format is plain prose plus a closing citation, with two ways of marking the drafter's own words:

Copied one. Copied two. [@key, p. 4-5] A sentence of your own comparing
the two papers. [see @key; @other] Another copied sentence.
[@other, p. 12]

<!-- own: gap analysis -->
A paragraph that is wholly the drafter's, such as the framing or the gap
analysis.

• A citation with pandoc's see prefix ([see @key], [see also @a; @b]) closes a sentence in the drafter's own words. The report checks it for support against the sources it names, never for copying.
• <!-- own: why --> on the line directly above a paragraph declares the whole paragraph the drafter's. A table row is the drafter's by construction: a comparison table's cells are short phrases, not copied runs.

"Copied exactly" means every letter and digit in the source's order. The matcher compares letters and digits only and drops the source's reference markers, so the drafter cleans the parse's typography -- spacing, run-together words, math italics, ◦ C -- without losing a match, and typesets the equations a run explains in the house math form (WRITING-STANDARDS.md §12) from the source's formula passages. The skill's shared reference, .claude/skills-common/references/digest-copying.md, has the table and the worked equation.

A citation covers every sentence back to the previous citation, or to the start of the paragraph. One citation can therefore cover a sentence, several sentences or a whole paragraph, and a sentence of the drafter's own placed inside a run, before its citation, is read as part of the run and reported as not copied. The skill copies from one paper per run, closes each run with its citation and a page, and keeps its own sentences short, few, between runs and closed by a see citation.

The skill follows survey-writer's draft-and-render workflow step for step: dossier, retrieval per sub-theme, gate, references, render to tex/pdf/md, a read as the reader, steering, the prose check, the fingerprint stamp, and the survey's structure and paragraph unit. A survey keeps source wording out of its body and a digest's body is source wording, so the survey's steps that exist to keep it out are replaced where they stand, and the skill says so at each one:

Survey step	In the digest
The evidence packet: claim: in the drafter's words, a quote: only when a quotation is warranted, prose drafted from claim: alone	Each source's main passage is recorded as its one quote:, which review quotation checks; every other run is checked where it stands by review digest. claim: still says what the passage establishes, and is drafted from only for the drafter's own sentences
Multi-source synthesis: a paragraph rests on two or more sources so it cannot be a transcription (WRITING-STANDARDS.md §11)	Kept, at the paragraph: one source per run, two or more runs from different papers per paragraph, <!-- single-source: why --> where one paper is genuinely the only source. review synthesis reports it
The pre-gate critique against the evidence packet	Replaced by a read as someone who has opened none of the papers: terms before definitions, runs naming what the digest does not show, paragraphs that are one paper after another
The verbatim scan	Replaced by review digest: the scan would flag every run, which is the design
The evidence sidecar	Not rendered: the digest is the sources' words, attributed where they stand
The prose check	Run and reported, nothing fixed; copied text keeps its source's style

The one rule the skill adds is a prohibition: never run agenda-reviser on a digest. The agenda does not read the digest report, but agenda-reviser's verbatim-run repair paraphrases short runs unattended, and on a digest that is the whole text.

🔍 Reading the report: review digest

python -m chitragupta.review digest content/drafts/<topic>/<name>.md

The report is filed at content/review/<topic>/<name>.digest.md with its .json beside it, on every run, because the .json is the next pass's baseline. It opens with four shares that together account for the whole digest, in the order a reader should take them:

Line	Counts	Direction
Unsupported fraction	the words of every sentence carrying a finding of the three source classes below, a sentence counted once however many it carries, over all words	down: the headline, the number a repair pass drives
Copied fraction	the words of every verified copied span, over all words	up, short of crowding out the argument
Own words, supported or declared	the words of every own-words sentence that raised nothing: a see sentence the cited sources support, or a sentence in an own paragraph or a table	as few as the argument needs
Not checkable	the words of every run whose cited source has no reading-ordered passages: no Docling sidecar. The reason names the stage to run	a fact about the corpus, not the digest

Each run is matched whole against the cited source, with the same matcher review quotation uses: both sides flattened to one character stream, an inline reference marker stripped from the source, and the run looked for in each passage and each adjacent pair. The passages are a Docling sidecar's, content/parsed/<citekey>.passages.json, which chitragupta enrich --stages docling writes. The aid needs the enriched corpus. A pdftotext parse has no reading order, and a run copied from one can be a collage of two columns, which is the hazard review quotation refuses to quote from; this aid refuses the same way, reports the run as not checkable, and names the stage to run. A run found whole is one copied span and raises nothing. A run whose sentences are all found but not contiguously is one span with the note assembled from N places, which is information rather than a finding. Only a run that breaks is split into sentences, and only to show where:

Class	Means	Repair
unsupported-text	a sentence of the drafter's own, and either no citation covers it or the cited source does not lexically support it	cite it correctly, replace it with the source's words, or delete it
copy-mismatch	a sentence not found verbatim but mostly on one page of the source; the item names the page and the missing words	open the source at that page and restore the exact wording
dangling-reference	a run naming a numbered figure, equation, section, table or chapter the digest does not show; a leftover reference marker such as [2]; or a copied run opening on a word that points back ("where", "thus", "these") with nothing before it from the same source	show the item, replace the mention with a bracketed editorial substitution, delete the marker, or move the run's start or end
repeated-text	a sentence of a copied run that an earlier run already copied from the same source; the item names the line of the first copy. Counted in no fraction	keep the first copy, cut the later one, and point back to it
unquoted-text	an undeclared sentence of the drafter's own: not closed by see, not in an own paragraph	close it with a see citation, replace it with a copied passage, or delete it

An undeclared sentence of the drafter's own carries unquoted-text always, and unsupported-text as well when the support test fails, so one sentence can be two items; the fraction counts its words once. A declared one -- see, own, a table row -- is checked for support only, and raises unsupported-text or nothing.

dangling-reference is listed and counted, and moves no fraction: the run is still the source's words, it just points at something the reader cannot see. Cross-references are read in copied runs only, since a sentence of the drafter's own saying "Section 2 compares them" points into the digest; a reference marker is read everywhere. A figure the digest shows, captioned Figure N of [@key], is not dangling for a run citing that key; a paragraph directly after a displayed equation may open on "where"; and a run right after a sentence citing its source, such as the drafter's own [see @a; @b], has its antecedent. On the digest that prompted it, the class found 43 cross-references, 43 reference markers and 4 openers, every one a real pointer to nothing.

repeated-text compares the copied spans with each other rather than with the source: a sentence (at least about eight words, flattened as every match is) copied again from the same citekey is reported on the later copy. On the survey-shaped digest that prompted it, it found three repeats a page-by-page read had missed. It reads copied text only, so a repeated equation display is not seen. Items are listed worst class first, then by line, in the agenda's own line format: a stable twelve-character id (a hash of the sentence, never its line, so an id survives an edit elsewhere), [surfaced], the section, a one-line summary. Every item is [surfaced]. Whether a source backs a sentence, and which passage should replace it, are judgement calls, and no re-run of a check can settle them. Nothing here is a gate, and the command exits 0 whatever it finds.

📄 Pages, and what a page note means

A page here is the parser's physical page index, counted from 1 over the PDF as parsed: the page a Docling passage sidecar records, or the form feed pdftotext emits between pages. It is not the printed folio. A paper whose first page is numbered 1203 in the journal is still p. 1 here, and the skill's format rule says to cite pages that way.

The page in a citation is a hint. The aid matches text, not pages: it searches every passage of the cited source and reports the page or pages the run was actually found on. A run that crosses a page break reports both pages. When no found page falls inside the cited range, the copied span carries a note:

cited p. 4-5, found on p. 7

A note is not a finding. It does not move any fraction, it does not appear in the item list, and the repair is to correct the locator in the citation. It can also mislead in one case worth knowing: the passages a source yields depend on which rung of CITATION-PROVENANCE.md's ladder answered, and a re-run of the Docling stage can renumber them, so a note that appears after one is worth a look before an edit.

🔁 The repair loop and --baseline

python -m chitragupta.review digest content/drafts/<topic>/<name>.md \
    --baseline content/review/<topic>/<name>.digest.json

The baseline is the .json a previous run filed. The aid recomputes everything, refiles the report, and then compares: items are matched by id into resolved, persisting and new; the per-class counts and both fractions are printed before and after; and one line, fell: yes or fell: no, says whether the repair counted. It counts when no class rose, at least one fell, nothing new appeared, and the unsupported fraction did not rise. The fraction is checked as well as the counts because deleting a flagged sentence together with half the copied text resolves an item and makes the digest worse.

The skill keeps a repair only under those conditions plus a passing gate, logs every attempt in the dossier's revisions.md, and stops when the user says so, when no item is left, or after three passes whose fraction did not fall. It repairs only when a person asked for a pass.

🧠 Lexical support, and what the NLI support aid adds

unsupported-text is decided lexically. The sentence's distinctive words (content words, stemmed, with short tokens dropped) are counted against the words of each passage of the cited source, and the best passage's share is compared with the provenance aid's weak band, [provenance].weak_score in config.toml, default 0.20. Below it the sentence is unsupported-text. A sentence no citation covers is unsupported-text without a test.

That catches the clear case: a sentence about something the source never mentions. It cannot see two others. A correct paraphrase in different words can score weak and be flagged although the source does say it; and a sentence that negates the source shares all its content words and scores strong, so it is not flagged. These are the limits CITATION-PROVENANCE.md records for the provenance aid, inherited here deliberately rather than restated.

python -m chitragupta.review support closes the second gap. It runs a real natural-language-inference entailment model over each citing sentence and the passages that lexically match it, and reports whether the source entails, is neutral to, or contradicts the sentence, so a negation is caught. It is not wired into review digest for three reasons: it needs the enrich extra, which installs a torch model of several hundred megabytes; it costs tens of seconds per draft where every other digest check is instant; and keeping the digest aid at interpreter tier 1, stdlib only, means it runs on any host that can run the gate. To use it on a digest, run it on the same file, read its report beside the digest's, and treat an unsupported-text item the model marks entailed as "cite it correctly" rather than "delete it".

🖼 Figures

Where the corpus was enriched with the Docling stage and [enrich].docling_images on, python -m chitragupta.draft figures <citekey> lists a paper's figures with a caption, a page, a crop and the exact string to cite each by. Every other genre may only look at a crop; AGENTS.md records the one exception for the digest. Private study text in the sources' own words may show the sources' own figures: the skill copies the crop into the topic's figures/ directory and places it directly after the run it belongs to, as an ordinary Markdown image whose caption is written afresh from the copied text, with the source's caption as a hint, and ends with the cite string. An image line is not prose, so review digest neither counts nor flags it, and the renderer carries the image into the PDF as it does for any draft.

The licence stops at the digest. A copied figure, like a copied sentence, never travels into a survey or a chapter.

It is not a precondition. A corpus parsed without images has no crops; the figures command says so and exits 0, the skill tells the user in one line, and the digest is built from the prose alone.

🚧 What it does not do

• Attribution is coarse. A sentence of the drafter's own placed inside a run before its citation is read as part of the run and reported as not copied. Two papers copied under one citation show the second paper's text as unmatched. The fix is to cite each paper at the end of its own part.
• It needs the enriched corpus. Without a Docling sidecar for a cited source, every run citing it is not checkable, including a run whose bracket also names a source that does have one (matching against the readable source alone would report text copied from the other as the drafter's own); the report names the stage to run. Within a sidecar, parse quality still bounds matching: a word the parse dropped or garbled is missing from what the run is matched against, so a copy made from the PDF there shows as copy-mismatch. Copy from the sidecar's words, cleaning only their typography.
• A half-copied sentence is reported whole, as a copy-mismatch naming the words that are not on the page, not split into a copied part and an original part.
• Lexical support is weak evidence, as the section above says.
• An item's id is its sentence, as an agenda item's is, so the same sentence of the drafter's written twice in one section is one id, and a baseline comparison sees one item until both copies are gone.
• The opener test is a word list. It reads the first word of a copied run that has nothing from the same source before it in its paragraph. It cannot tell "This paper proposes" (which the citation resolves) from "This model" (which needs an antecedent), and it does not read the drafter's own sentences, whose "This" usually means the sentence before.
• A citation closes a run wherever it sits. The format puts it after the last copied sentence; a narrative citation mid-sentence ([-@key], or @key in prose) splits the sentence across two runs and is outside what the aid reads.
• The agenda does not read it, and agenda-reviser never touches it. The digest's classes mean nothing outside a digest, and wiring them into the agenda would add a line to every existing agenda for a report no survey or chapter has. The digest's report is its own worklist, worked by review-digest on request.

⚖ The rule it lives under

SOUL.md refuses to pass a source's wording off as the drafter's. A digest keeps that rule by construction: every copied run ends with the citation that names its source, the dossier records each source's main passage as a quote:, and the digest is never a source for any other draft. AGENTS.md's statement that verbatim wording has one legitimate home outside the draft names the digest as the one genre that is that home, and PLAGIARISM.md's verbatim scan is what catches the digest's wording if it ever reaches a draft that is not one.