Skip to content

Verbatim scan: content/drafts/dt-overview/trust-chapter.tex

Review aid, not a gate. This report is evidence for a human judgement, never a verdict. A driver may read it back; no draft is blocked by what it says. See SOUL.md, and docs/ARCHITECTURE.md's "Layer 4: the review layer".

  • Draft: content/drafts/dt-overview/trust-chapter.tex
  • Command: python -m chitragupta.review verbatim scan content/drafts/dt-overview/trust-chapter.tex --min-run 8 --gap 1 --write
  • chitragupta 6.166.1
  • Allowlist: none configured (content/verbatim_allowlist.toml not found)

How to read this

Every run of at least --min-run words this draft shares with any parsed source in the corpus, cited or not. Sharing wording is not by itself misconduct -- a defined term, a standard's name and a correctly quoted sentence all show up here -- so each finding is a place to look, not a charge.

Two flags narrow the reading:

  • UNCITED SOURCE -- the paragraph the run sits in does not cite the source it matched. That is the finding overlap structurally cannot make, and the one most worth reading first.
  • quoted -- the run touches quote delimiters, so it is most likely a deliberate quotation. A run is usually wider than the quotation inside it -- it can open in the draft's own framing prose -- so this reads as overlap, not containment.

Each finding names its tier: exact is a verbatim run; skip-gram is a tolerant stemmed-subsequence match that also catches a passage with a handful of words substituted. A skip-gram finding's word count is matched words / span: how many words the tier actually matched, out of the raw width of text those matches span -- the two can differ a lot, since a skip-gram window can stretch across stopwords and opposite-family words that were not themselves matched.

embedding is the third tier: a sentence-level alignment between a section of this draft and the sources its dossier records that section as written from. It matches meaning rather than wording, so it is the only tier that can see a genuine restatement -- and the only one whose findings are not reproducible from the draft and the corpus alone, since the vectors change with [enrich].embedding_model. Its score is that alignment's strength, not a probability and not comparable to anything the other two report; a passage a deterministic tier already flagged is left to that tier rather than reported twice.

Where a finding shows two blocks, Draft: and Source:, the two are not the same wording -- had they been, the exact tier would have caught it and the tier that did report it would have stood aside. Read them against each other, using the markup:

  • Bold on either side -- a word that side has and the other does not: substituted, or added by the draft.
  • Italic, source side only -- a word the source has that the draft dropped.
  • Unmarked -- the wording the two actually share. That is the overlap, and it is left bare so it is the one thing on the line the eye finds without looking.

The draft side is normalized (lowercased, punctuation dropped) because that is the stream the tiers compare; the source side is its own text as parsed. So do not read a difference in capitals or apostrophes as a finding -- the comparison behind the markup is run on normalized words on both sides, and marks none of that.

An exact finding shows one unmarked block instead of two: the tier only fires where the wording is identical, so there is no second side to show and nothing to mark.

An embedding finding's word count says aligned, not matched: it is the width of the aligned sentences, not a count of words the two sides share. A long aligned count with a low score is a weak alignment over a long sentence, not a long lift.

Findings below are grouped most-damning-first: long runs, then short ones, then quoted runs -- but a quoted run only drops into the last group when it also cites the source it matched. A quoted run from an uncited source is still grouped by length (on matched words, not raw span), not buried under quoted.

The allowlist bullet above names a per-host, gitignored file (content/verbatim_allowlist.toml, see docs/PLAGIARISM.md) of boilerplate this host's owner has decided never to flag -- a run is only dropped when what's left after discounting the allowlisted text would no longer clear --min-run on its own, so a real lift that merely contains a defined term still shows up below.

A clean run is not a clean bill of health. This draft has been checked against all three tiers, but they do not cover the same ground: the two deterministic ones see wording, and the embedding tier sees meaning only within each section's own recorded sources. Reuse from a source a section's dossier does not record is outside what any of the three can find by restatement alone. See docs/PLAGIARISM.md.

Findings

No verbatim run of 8 words or more was found anywhere in the draft.