paperlint 5.0.0 → 6.1.0

package/docs/rules/anonymity/identity.md

@@ -0,0 +1,111 @@
+# anonymity/identity
+
+**Level:** error · **Reads:** `paper.tex` → `paperlint.json`, the preset it extends, and
+`_build/paper.facts.json` (what `paperlint build` read from the PDF)
+
+## What it catches
+
+For a venue that reviews double-blind — its preset says `"blind": true` — two things:
+
+- the paper declares no `identity`, so there is nothing to check the PDF against;
+- a string from `identity` appears in the built PDF: on a page, in its metadata (the Info
+  dictionary and the XMP packet: `Author`, `Title`, `Keywords`, …), or in the target of a link.
+
+A rule cannot know who wrote a paper, so the authors say it once, in `paperlint.json`. Given that
+list, finding it is mechanical. A missing list is an error, not a pass: a check with no input and a
+check that found nothing would otherwise look the same.
+
+The PDF is what reviewers get, so the PDF is what is read — not the source. A name reached through
+a macro (`\newcommand\me{Ada Example}`), a self-citation's author list from the `.bib`, a
+`\hypersetup{pdfauthor=…}` that is in no page's text, a `\href` whose visible text differs from its
+target: each ends up in the PDF, and each is found there.
+
+## Why
+
+> Submissions must be properly anonymized for double-blind review.
+
+— AIDC's call for papers, and most double-blind venues say the same. A paper that names its authors
+is desk-rejected, and the ways a name gets in are mostly invisible when reading the PDF: metadata
+the class copies from `\author`, a link to your own GitHub account, a name in a figure.
+
+## How a name is matched
+
+PDF text is not the source text; the same name comes out spelled several ways, and each is matched:
+
+| in the PDF                               | matched as                                           |
+| ---------------------------------------- | ---------------------------------------------------- |
+| `ADA EXAMPLE` (small caps), `ада пример` | Unicode case folding, no locale — Cyrillic folds too |
+| `fi` ligature                             | NFKC on both sides                                   |
+| `Ex` `ample` (two items, kerning)        | the page's items are joined without a space          |
+| `A d a  E x a m p l e` (letter-spaced)   | whitespace may stand between any two characters      |
+| `Exam-` ⏎ `ple`, `Ada` ⏎ `Example`       | so may a hyphen or a line break                      |
+| `AdaExample`, `Ada-Example`              | a token's own spaces and hyphens are optional        |
+
+Only the ends of a token are anchored: `Ada` does not match inside `Canada`, `MIT` not inside
+`submitted`. Two tokens that find the same text in the same place are one finding.
+
+Anonymous hosting needs no exception: an `osf.io/…?view_only=…` or `anonymous.4open.science` link
+names nobody, so nothing in `identity` matches it.
+
+## Examples
+
+```json
+{
+  "extends": "paperlint:aidc",
+  "kind": "regular",
+  "identity": [
+    "Ada Example",
+    "adaexample",
+    "ada@example.org",
+    "Example University",
+    "exampletool"
+  ]
+}
+```
+
+Findings:
+
+> PDF, page 1: «Ada Example» matches «Ada Example» in `identity`, and aidc reviews double-blind —
+> remove it, or refer to your own work in the third person
+
+> PDF metadata: Author: «Ada Example» matches «Ada Example» in `identity`, …
+
+> link target on page 4 (https://github.com/adaexample/exampletool): «adaexample» matches …
+
+With no `identity`:
+
+> aidc reviews double-blind, and this paper declares no `identity`, so nothing in the PDF can be
+> checked against the authors. Declare it in paperlint.json (the root's for every paper, or this
+> paper's): "identity": ["Your Name", "your-handle", "you@example.org", "Your University", "your-project"]
+
+## Options / preset fields
+
+No rule options.
+
+- `identity` in `paperlint.json` — a list of strings, each with a letter or digit. The root file's
+  list and the paper's own are **joined**: a paper adds its co-authors to the project's author and
+  cannot drop them by accident. Put in everything that identifies you: every spelling of every
+  author's name, handles, emails, affiliations, your own tools' and repositories' names.
+- `blind` in the preset — `true` where the venue's call says the review is double-blind. Shipped:
+  `paperlint:aidc`. A venue that is not blind: the rule is silent, `identity` or not.
+
+## What it does not check
+
+- **Text inside raster images.** A screenshot with your name, a logo, a scanned figure: text
+  extraction reads text, not pixels, so a name there is not seen. Look at the figures yourself.
+- **What you did not declare.** A co-author missing from `identity`, a nickname, an old handle.
+- **Self-citation in the third person** is allowed by most double-blind venues, and a cited name is
+  still a declared name: the rule reports it. Check the venue's policy; if it allows the citation,
+  cite in the third person and keep the finding in mind, or drop that token for this paper.
+- **A name split by hyphenation inside a token** that the PDF does not mark (a hyphen is matched,
+  a missing space between two lines of a hyphenated compound is).
+- **Words that are also names.** `identity` entries match as whole words anywhere; a common word
+  used as a handle will be found in ordinary prose.
+- **The source.** Comments, `\iffalse` blocks and files not in the build are not read — they are
+  not in the PDF reviewers get. Supplementary material uploaded separately is not read either.
+
+## How to fix
+
+Remove the name from where the message says: the author block, the acknowledgments (for review),
+`\hypersetup{pdfauthor=…}` / `pdfusetitle`, a link to your account (use an anonymous mirror), a
+self-citation in the first person. Rebuild (`paperlint build`) and lint again.

package/docs/rules/format/layout-override.md

@@ -0,0 +1,61 @@
+# format/layout-override
+
+**Level:** error · **Reads:** `paper.tex` and the files it includes, and the preset its
+`paperlint.json` extends
+
+## What it catches
+
+A command that changes the page layout the venue's template sets for the whole document:
+
+| command                                                                                                                                                                                | changes          |
+| -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------- |
+| `\usepackage{geometry}` (alone or in a list), `\geometry{…}`, `\newgeometry{…}`                                                                                                        | the margins      |
+| `\setlength` / `\addtolength` on `\textheight`, `\textwidth`, `\topmargin`, `\oddsidemargin`, `\evensidemargin`, `\columnsep`, `\headsep`, `\headheight`, `\footskip`, `\baselineskip` | the text block   |
+| `\linespread{…}`, `\renewcommand{\baselinestretch}{…}`                                                                                                                                 | the line spacing |
+
+It reads the parse tree, not the text: a command in a comment is not a command, and a `comment`
+environment is skipped. A macro definition is read: `\newcommand{\tall}{\setlength{\textheight}{10in}}`
+sets the length wherever it is used, so it is reported where it is defined.
+
+It is silent for a paper whose `paperlint.json` names no preset, or a preset that names no
+`template` — there is no mandated layout to compare against.
+
+## Why
+
+Venues set the page through their template and reject papers that change it: IEEE and ACM venues
+run format checks, and a shrunken margin or a tightened line spacing is a classic desk reject.
+AIDC: «LaTeX submissions must use the IEEEtran.cls version 1.8b template». The page-limit and
+geometry checks see the result only in the built PDF; this rule points at the line that does it.
+
+## Examples
+
+```latex
+\usepackage[margin=0.7in]{geometry}      % reported
+\setlength{\textheight}{9.6in}           % reported
+\linespread{0.95}                        % reported
+\setlength{\parindent}{0pt}              % not reported: not a layout length
+Text.\vspace{-2mm}                       % not reported: local space, not the template's layout
+```
+
+> `\usepackage{geometry}` changes the page layout aidc's template sets
+> (`\documentclass[conference,compsoc]{IEEEtran}`) — a desk-reject reason at venues that check the
+> format. Remove it; if the venue allows it, disable this line with a comment saying so
+
+## Options / preset fields
+
+No rule options. It reads the preset's `template`.
+
+## What it does not check
+
+- Local spacing: a negative `\vspace`, `\vspace*` or `\vskip` around a float or a heading. It
+  changes one spot, not the layout the template sets, and no venue's call names it.
+- Font size changes (`\small` body text, `\fontsize`), `\baselineskip` changed by `\fontsize`,
+  `\enlargethispage`, `\looseness`, or float spacing lengths (`\textfloatsep`, `\abovecaptionskip`).
+- Conditionals are not evaluated: a command inside an `\iffalse … \fi` block in the preamble is
+  still reported.
+- Layout lengths set with plain TeX assignment (`\textheight=9in`).
+
+## How to fix
+
+Remove the command and let the template set the page. If the venue explicitly allows it, keep it
+with `% eslint-disable-next-line format/layout-override -- <the venue's words>`.

package/docs/rules/format/page-limit.md

@@ -0,0 +1,108 @@
+# format/page-limit
+
+**Level:** error · **Reads:** `paper.tex` → `paperlint.json` (the venue preset and the paper's
+`kind`) and `_build/paper.facts.json` (what `paperlint build` measured in the PDF)
+
+## What it catches
+
+A PDF longer than its venue allows for the paper's kind: more body pages than `body_pages_max`, or
+more reference pages than `ref_pages_max`.
+
+How the body is counted depends on the preset:
+
+- **By default**, body and reference pages are banal's count: a page holding any text that is not
+  bibliography is a body page.
+- **With `"body_ends_at": "references"`** (AIDC), the body is every page up to where it ends: at
+  the references or at the appendix, whichever comes first — «excluding references and
+  appendices» excludes an appendix wherever it sits. What follows is not counted. The page the body
+  ends on counts as body when body text stands above the heading there: a body that runs half a
+  page past the limit is over the limit.
+
+## How it finds where the body ends
+
+Each end is found by a structural signal first, then its heading line on that page.
+
+**The references.**
+
+1. **hyperref's first bibliography entry.** hyperref writes a destination named `cite.<key>` for
+   every `\bibitem`; the lowest page any of them is on is where the bibliography is. The build
+   records it as `bib_anchor_page`.
+2. **The heading.** A line that reads `References` or `Bibliography` — any case, spaces ignored
+   (IEEEtran prints `REFERENCES` in small capitals, which the PDF may give as `R EFERENCES`),
+   optionally numbered (`7 References`, `VII. REFERENCES`). A heading whose next line opens the
+   first entry, `[1]`, is **confirmed**.
+
+With the anchor, the references start on its page, at the confirmed heading there, else at the last
+heading line on that page. Without hyperref: at the first confirmed heading, else at the first
+heading line.
+
+**The appendix.** hyperref's destination for the appendix's first section (`appendix.A` in the
+standard classes) or the first bookmark titled «Appendix…» (IEEEtran's `\appendices`) gives its
+page, recorded as `appendix_anchor_page`; on that page, the line reading `Appendix A.`,
+`APPENDIX` or `Appendices`. Without hyperref the appendix is not looked for: a sentence can end on
+a line reading «Appendix A.», and nothing would tell it from the heading.
+
+On the page the body ends on, the lines above the heading decide: a line with a letter in it is
+body text, unless it is a running header or footer — the same line, digits aside, on at least three
+pages and a third of them. A page number is not body text either.
+
+A line reading «References» that is not the bibliography's heading — a `\subsection*{References}`,
+a table column headed «References» — is not taken for it: the anchor names another page, or the
+line is not followed by `[1]`. When a structural signal and the text disagree — a `cite.<key>` page
+with no heading on it, a confirmed heading on another page, an appendix page with no «Appendix»
+line — the count is not made, and the rule says so (`unclear`).
+
+## Why
+
+AIDC limits the paper to 12 pages «excluding references and appendices», and places the LLM Usage
+Statement after the references, where it «does not count towards the page limit». banal counts any
+page with non-bibliography text as body, so a statement or appendix after the bibliography would
+fail a correct paper. A body that runs onto the references page, even by a few lines, is a
+thirteenth page — the desk-reject case this rule exists to catch.
+
+## Examples
+
+A 12-page AIDC paper whose conclusion ends at the foot of page 12, «References» at the top of page
+13: body 12 pages, passes.
+
+The same paper with the conclusion running four lines onto page 13, «References» below them:
+
+> body pages (up to the references on page 13): 13, over the limit 12 for aidc/regular — a desk
+> reject; cut the text
+
+An accepted ACSAC paper that places its appendix tables before the bibliography: the appendix starts
+at the top of page 12, the references on page 15. The body is 11 pages.
+
+When a structural signal and the text disagree, the count is not made and the rule says so:
+
+> could not tell where the body ends, so it was NOT counted against aidc's limit: the first
+> bibliography entry is on page 14 (hyperref's destination), and no line on page 14 reads
+> «References». Check the headings of the bibliography and the appendix, and that nothing before
+> them reads «References» above a [1]
+
+## Options / preset fields
+
+No rule options. It reads, from the preset, `format.kinds.<kind>.body_pages_max`,
+`format.kinds.<kind>.ref_pages_max` and `format.body_ends_at`; from the paper's `paperlint.json`,
+`kind`. A preset with no kinds, or a kind with no limits, has nothing to check.
+
+## What it does not check
+
+- **An appendix before the references, without hyperref**, is counted as body: nothing structural
+  says where it starts. With hyperref it ends the body.
+- **An appendix whose heading does not read «Appendix…»** (the standard classes print `A` and the
+  title) is reported as `unclear` when hyperref anchors it.
+- **A bibliography that is not numbered and not linked by hyperref** has no confirmed heading and no
+  anchor; the first line reading «References» is then taken, and a body line reading exactly that
+  would be taken for it.
+- **A running header that repeats on fewer than three pages** (a two-page paper) counts as body
+  text above a heading at the top of a page.
+- Whether the venue counts pages the same way for other kinds of material (supplementary files,
+  artifact appendices).
+
+## How to fix
+
+Cut the body until it ends within the limit. With `body_ends_at`, check where the body ends on the
+rendered PDF: body text above the heading on that page counts that page. For `unclear`, give the
+bibliography its heading (`\bibliography`, `thebibliography`) and the appendix one reading
+«Appendix», and make sure no line before them reads «References» above a `[1]`.

package/docs/rules/pdf/body-size.md

@@ -0,0 +1,87 @@
+# pdf/body-size
+
+**Level:** error · **Reads:** `paper.tex` → `paperlint.json` (the venue preset) and
+`_build/paper.facts.json` (what `paperlint build` measured in the PDF)
+
+## What it catches
+
+A PDF set in another size than its venue's template:
+
+- the body font size more than `body_pt_tol` off the preset's `body_pt` (`body`);
+- the reference font size outside the preset's `ref_pt_min`–`ref_pt_max`, widened by `body_pt_tol`
+  on both sides (`refPt`);
+- a build that measured no body size at all (`bodyMissing`) — rebuild.
+
+Both sizes are what banal measures on the rendered pages: the most common size of the body text,
+and of the first bibliography page. It is not the size the source declares, so a
+`\documentclass[9pt]`, a `\fontsize` in the preamble or a bibliography wrapped in `\scriptsize` are
+all seen, whatever macro did it.
+
+## Why
+
+A call for papers that fixes a template fixes its font size, and shrinking the text is the usual
+way a paper is made to fit the page limit. A reviewer or a production editor who notices returns
+the paper; a page limit checked on a shrunk paper checks nothing.
+
+It is an error because it was measured to be exact on real papers, not assumed. On five accepted
+papers, built (`fixtures/accepted-papers/body-size.json`, `test/e2e/tex/page-count.e2e.ts`):
+
+| paper                          | declared               | rendered (pdf.js)                          | measured (banal) | finding |
+| ------------------------------ | ---------------------- | ------------------------------------------ | ---------------- | ------- |
+| four ACSAC papers, IEEEtran    | 10 pt, references 8 pt | 10.00 / 8.00 (9.96 / 7.97 without compsoc) | 10.3 / 8.3       | none    |
+| AgenticDev '26, acmart sigconf | 9 pt, references 7 pt  | 8.97 / 6.97                                | 9.3 / 7.3        | none    |
+| an ACSAC paper with `9pt`      | 9 pt                   | 9.00 / 8.00                                | 9.3 / 8.3        | `body`  |
+| references in `\scriptsize`    | 7 pt                   | 10.00 / 7.00                               | 10.3 / 7.3       | `refPt` |
+| references in `\small`         | 9 pt                   | 9.96 / 8.97                                | 10.3 / 9.3       | `refPt` |
+
+banal reads every size 0.3 pt above the size the glyphs are set in, on both templates, and
+`body_pt_tol` of 0.5 absorbs that. Because of that offset the margins are not symmetric: every accepted body, and every IEEE bibliography (a single size, 8 pt), lands 0.2 pt inside the upper edge, a paper one size smaller (9 pt body, 7 pt
+references) 0.2 pt outside the lower edge, and one size larger 0.8 pt outside the upper one. The
+offset was exactly 0.3 on all eight builds; a template on which banal reads differently would need
+its own measurement. A change of half a point (IEEEtran's 9pt `\small` is 8.5 pt) can pass
+unreported.
+
+## Examples
+
+An IEEEtran paper built with `\documentclass[9pt,conference,compsoc]{IEEEtran}`:
+
+> the body font size measures 9.3 pt against 10 ± 0.5 pt for ieee-conference. The measurement is
+> the mode of the rendered text, not the declared size — check \documentclass and its options
+
+The same paper with `{\let\footnotesize\scriptsize\bibliography{refs}}`:
+
+> the reference font size is 7.3 pt, outside 8 pt for ieee-conference (allowing ±0.5 pt for the
+> measuring drift) — fix the bibliography's font size
+
+The paper as the template sets it, `\documentclass[conference,compsoc]{IEEEtran}` and a plain
+`\bibliography{refs}`, passes.
+
+## Options / preset fields
+
+No rule options. From the preset's `format`: `body_pt` and `body_pt_tol` (both needed for the body
+check), `ref_pt_min` and `ref_pt_max` (both needed for the reference check). A preset without them
+does not check that size. The shipped presets:
+
+| preset                                    | `body_pt` ± `body_pt_tol` | references |
+| ----------------------------------------- | ------------------------- | ---------- |
+| `ieee-conference` (and `aidc`)            | 10 ± 0.5                  | 8          |
+| `acm-sigconf` (and `agenticdev`, `aisec`) | 9 ± 0.5                   | 7–9        |
+| `realm`                                   | 11 ± 0.5                  | 9–11       |
+
+## What it does not check
+
+- **The size of anything but the body and the first bibliography page**: captions, footnotes,
+  tables, the title and the appendix are not measured. A table set in `\tiny` passes.
+- **A paper that is not built.** Without `_build/paper.facts.json`, or without banal, the rule is
+  silent and `pdf/measured` says why.
+- **The `realm` preset** has not been checked on built accepted papers of its venue: its body size
+  is ACL's stated 11 pt, and its reference range was set from a single measurement.
+- A paper with no bibliography has no reference size, and only the body is judged.
+
+## How to fix
+
+Use the template's size: remove the size option from `\documentclass` and any `\fontsize`,
+`\small` or `\footnotesize` around the body; for the references, remove what wraps
+`\bibliography` or changes `\bibfont`. Rebuild (`paperlint build`) and lint again. If the venue
+really allows another size, override the preset's `format` in your own preset, or set the rule to
+`"off"` in `paperlint.json`.

package/docs/rules/tex/claim-emphasis.md

@@ -0,0 +1,71 @@
+# tex/claim-emphasis
+
+**Level:** warn · **Reads:** `paper.tex` and the files it includes → `paperlint.json`
+
+## What it catches
+
+A body that sets its claims in bold inside running sentences more often than the accepted papers of
+its venue do. A claim in bold is a `\textbf{…}` (or `{\bfseries …}`, `{\bf …}`) that:
+
+- stands **inside** a sentence — not opening a paragraph or a list item, where bold is a label
+  (`\textbf{Varying the volume}. We …`, `\item \textbf{A corpus of questions}: …`), braces around it
+  or not; and
+- **states** rather than names: six words or more, or holding a number
+  (`\textbf{median coverage is 2.5/10}`, `\textbf{they react to the dangerous word}`).
+
+A term in bold (`\textbf{Pressure-Acoustic Transformation (PAT)}`), italics and `\emph`, headings,
+run-in headings, captions, tables and figures are not counted. Counted per 10,000 words of the body.
+
+Above the band it reports once at the start of the body, then once at every claim in bold, where it
+stands. Below the band it reports once. A body under 2 000 words, or a paper whose preset records no
+anchors, is not judged.
+
+## Why
+
+Setting a finding in bold is how a post makes a point land; an accepted paper puts the number in a
+sentence that says whose it is, or in a table. Measured on the LaTeX source, per 10,000 words:
+
+| paper                                                   | words | claims in bold | per 10,000 |
+| ------------------------------------------------------- | ----: | -------------: | ---------: |
+| `leaking-queries-acsac25` (ACSAC 2025)                  |  9133 |              0 |        0.0 |
+| `rr-dataset-quality-acsac24` (ACSAC 2024)               |  8130 |              0 |        0.0 |
+| `secure-acsac24` (ACSAC 2024)                           |  7519 |              0 |        0.0 |
+| `llm-splained-acsac25` (ACSAC 2025)                     |  9529 |              1 |        1.0 |
+| `barovox-acsac24` (ACSAC 2024)                          |  7670 |              8 |       10.4 |
+| the paper reviewers called a blog post, current version | 10476 |             42 |       40.1 |
+| the same paper, before its rewrite                      |  6128 |             42 |       68.5 |
+| `agenticdev-acm26` (ours, "too informal", ACM)          |  4374 |              8 |       18.3 |
+
+Two shapes were measured and left out, because they do not separate the groups: all bold inside a
+sentence (accepted papers 0–44 per 10,000, that paper 60 — terms in bold are common), and
+phrases of six words or more in italics (accepted 0–7.8, the paper 4.8 — italics carry quotes and
+examples). BaroVox's 8 are six of its results set in bold (_0.35_, _4.09/5_, …), its claim of a
+_first-ever use of this technique on pressure sensors_, and a dataset name with a digit
+(_SpeechCommand256_). That name and LLM-splained's one count, a list of topics in bold, are the two
+false positives in the corpus; the band keeps both papers inside.
+
+## Examples
+
+```latex
+We find that \textbf{the median guard blocks 2 of 10 disasters}.   % a claim in bold
+The median guard blocks 2 of 10 disasters (Table~\ref{t:cov}).      % none
+We introduce \textbf{GateBench}, a harness for scraped guards.     % none: a term
+\textbf{Threat model.} The attacker controls the prompt.           % none: a run-in heading
+```
+
+## Options / preset fields
+
+No rule options. The band is the venue preset's `register.anchors`, exactly as for
+[`tex/contrast-frames`](contrast-frames.md), which explains how it is derived. For `paperlint:aidc` it
+is **0–17.8**.
+
+## What it does not check
+
+Whether a bold phrase deserves emphasis, a claim in bold shorter than six words with no number
+(_it is unsafe_), or emphasis by other means (capitals, colour, a box).
+
+## How to fix
+
+Take the bold off a finding and let the sentence carry it — with its owner (_we measured_, a
+`\ref` to the table) — or move the number into a table or figure. Keep bold for a term at its
+definition and for a label that opens a paragraph.

package/docs/rules/tex/contrast-frames.md

@@ -0,0 +1,100 @@
+# tex/contrast-frames
+
+**Level:** warn · **Reads:** `paper.tex` and the files it includes → `paperlint.json`
+
+## What it catches
+
+A body whose contrast frames — a claim set against a rejected alternative — run outside the band
+the accepted papers of its venue set. Counted per 10,000 words of the body, in these forms:
+
+| form                            | example                                                         |
+| ------------------------------- | --------------------------------------------------------------- |
+| _X, not Y_ (and `;`, `---`)     | _they react to the dangerous word, not the dangerous operation_ |
+| _X, and not Y_ · _X, but not Y_ | _it reads the field, and not the raw input_                     |
+| _not X but Y_ (one clause)      | _over-blocking is not a side metric but a precondition_         |
+| _rather than_                   | _it fires on the mention rather than the act_                   |
+| _instead of_                    | _it reads the whole event instead of the command_               |
+| _as opposed to_                 | _it matches words as opposed to operations_                     |
+
+Not counted: _not only … but also_; a negated clause followed by _, but_ (_the audit is not
+reported, but …_); _, not yet_ · _not already_ · _not necessarily_ and their kin, which reject
+nothing; _without_ and _unlike_ — both measured, both as frequent in accepted papers as in the
+paper reviewers called a blog post, so they would add noise and no signal.
+
+Above the band it reports once at the start of the body — the count, the rate, the band and the
+anchors' own range, with the first three sentences — and then once at every frame, with the words
+around it, so an editor shows where to rewrite. Below the band it reports once. A body under 2 000
+words is not judged, and neither is a paper whose preset records no anchors.
+
+It reads the same body as [`tex/register`](register.md): from the start of the document, the
+abstract included, to `\appendix` or the bibliography, includes spliced in; headings, run-in
+headings, captions, tables, figures and code are not prose.
+
+## Why
+
+Reviewers called a paper "too informal" and "a blogpost". Its body set a claim against what it is
+not several times per page — the device carries the argument in place of a word that names the
+relation (_because_, _therefore_). Measured on the LaTeX source, per 10,000 words:
+
+| paper                                                      | words | frames | per 10,000 |
+| ---------------------------------------------------------- | ----: | -----: | ---------: |
+| `barovox-acsac24` (ACSAC 2024)                             |  7670 |      1 |        1.3 |
+| `rr-dataset-quality-acsac24` (ACSAC 2024)                  |  8130 |      3 |        3.7 |
+| `secure-acsac24` (ACSAC 2024)                              |  7519 |      3 |        4.0 |
+| `llm-splained-acsac25` (ACSAC 2025)                        |  9529 |      4 |        4.2 |
+| `leaking-queries-acsac25` (ACSAC 2025)                     |  9133 |      5 |        5.5 |
+| the paper reviewers called a blog post, before its rewrite |  6128 |     62 |      101.2 |
+| the same paper, after a rewrite aimed at contrast frames   |  6121 |     25 |       40.8 |
+| the same paper, its current version                        | 10476 |     29 |       27.7 |
+| `agenticdev-acm26` (ours, "too informal", ACM)             |  4374 |     22 |       50.3 |
+
+An independent blind hand count — a reader who did not know which text was which, counting every
+explicit negation that rejects one named term for another, _without_ and _unlike_ included — put the
+first version at ~86 per 10,000 words and four accepted ACSAC excerpts at ~9–19: a different counter,
+the same five- to tenfold gap. Compare rates only within one counter.
+
+## Examples
+
+```latex
+The guards react to the dangerous word, not the dangerous operation.      % a frame
+The guards match the words of a command; they do not parse the operation. % none
+Because the guards match words, 24 of 46 block a harmless echo.          % none
+```
+
+## Options / preset fields
+
+No rule options. The band is the preset's: `register.anchors` in the venue preset, each anchor a
+paper with its body's `words` and its count of each measure, as these rules count them:
+
+```jsonc
+"register": {
+  "anchors": [
+    { "paper": "barovox-acsac24", "words": 7670, "contrast_frames": 1, "claim_emphasis": 8, "relation_markers": 31 }
+  ]
+}
+```
+
+The band is the anchors' range widened at each end by two Poisson standard deviations of the anchor
+standing there (an anchor with no occurrence is widened as if it had one; below zero is zero), so a
+paper of that anchor's length and register falls outside it about one time in forty. For
+`paperlint:aidc`, the anchors are the five accepted ACSAC papers above — AIDC has no accepted papers of
+its own yet, and uses ACSAC's template — and the band is **0–10.4**. `src/register-bands.test.ts`
+re-measures every anchor that has a fixture, so a change to a counter changes the preset with it.
+
+A preset of your own gets a band the same way: measure accepted papers of the venue, as LaTeX
+source, and record them (`anchorOf` in `src/register-bands.ts` is the one way an anchor is measured).
+
+## What it does not check
+
+Whether a given frame is needed. A frame is the right sentence where a reader would otherwise assume
+the rejected alternative; the rule counts, and judging each one is the job of the `grade-paper-writing`
+skill. It does not read a frame split over two sentences (_It is fast. Not exact._), a negation by
+synonym (_ignores_, _fails to_), or anything in the appendix.
+
+## How to fix
+
+Say what the thing is and why, with the evidence, and name the relation to the sentence before —
+_because_, _therefore_, _for example_ — instead of setting it against what it is not. Replacing
+_, not_ with _rather than_ changes nothing: both are counted. Keep a frame where the rejected reading
+is the one a reviewer brings to the page, and check after each rewrite that every claim and number
+still says what it said.

package/docs/rules/tex/relation-markers.md

@@ -0,0 +1,72 @@
+# tex/relation-markers
+
+**Level:** warn · **Reads:** `paper.tex` and the files it includes → `paperlint.json`
+
+## What it catches
+
+A body that names how its sentences relate less often — or far more often — than the accepted papers
+of its venue. A relation marker is one of: _therefore_, _thus_, _hence_, _consequently_, _as a
+result_, _because_, _since_, _this means_, _this implies_, _which explains_, _in turn_, _that is,_,
+_i.e._, _in contrast_, _by contrast_, _however_, _nevertheless_, _nonetheless_, _whereas_, _in other
+words_, _for this reason_, _it follows_, _accordingly_, _specifically,_, _in particular,_, _for
+example_, _for instance_, _e.g._ — as whole words, anywhere in a sentence. Counted per 10,000 words of
+the body.
+
+Outside the band it reports once, at the start of the body, with the count, the rate, the band and
+the anchors' range. A body under 2 000 words, or a paper whose preset records no anchors, is not
+judged.
+
+This is the FLOOR among the register checks: [`tex/contrast-frames`](contrast-frames.md) and
+[`tex/claim-emphasis`](claim-emphasis.md) catch a device used too often; this one catches the
+relation left for the reader to infer, which is what a rewrite that only deletes devices leaves
+behind.
+
+## Why
+
+A reviewer wrote: _"Readers are often required to infer the intended relationships between claims."_
+Measured on the LaTeX source, per 10,000 words:
+
+| paper                                                   | words | markers | per 10,000 |
+| ------------------------------------------------------- | ----: | ------: | ---------: |
+| `rr-dataset-quality-acsac24` (ACSAC 2024)               |  8130 |      32 |       39.4 |
+| `barovox-acsac24` (ACSAC 2024)                          |  7670 |      31 |       40.4 |
+| `secure-acsac24` (ACSAC 2024)                           |  7519 |      38 |       50.5 |
+| `llm-splained-acsac25` (ACSAC 2025)                     |  9529 |      84 |       88.2 |
+| `leaking-queries-acsac25` (ACSAC 2025)                  |  9133 |     103 |      112.8 |
+| the paper reviewers called a blog post, current version | 10476 |      20 |       19.1 |
+| the same paper, before its rewrite                      |  6128 |       9 |       14.7 |
+| `agenticdev-acm26` (ours, "too informal", ACM)          |  4374 |      13 |       29.7 |
+
+_However_ alone stands 6–11 times in every accepted paper and never in the current version of ours. The
+separation survives dropping any one marker from the list: with each of the 24 that occur removed in
+turn, the lowest accepted paper stays at least 1.38 times the highest version of ours (the worst case
+is dropping _however_). The list was fixed before this
+corpus was measured (it is the one a register diagnosis of the same papers used, on text extracted
+from PDFs, where the gap was narrow); on the parse tree the gap is twofold.
+
+## Examples
+
+```latex
+Most guards match words. 24 of 46 block a harmless echo.                  % the relation is left to infer
+Because most guards match words, 24 of 46 block a harmless echo.          % named
+Most guards match words; as a result, 24 of 46 block a harmless echo.     % named
+```
+
+## Options / preset fields
+
+No rule options. The band is the venue preset's `register.anchors`, exactly as for
+[`tex/contrast-frames`](contrast-frames.md), which explains how it is derived. For `paperlint:aidc` it
+is **25.4–135.0**.
+
+## What it does not check
+
+Whether a marker names the right relation, or a relation named another way (_this is why_, _so that_,
+a colon). The list counts only words whose meaning does not depend on context; _since_ is on it and
+is sometimes temporal.
+
+## How to fix
+
+Where two sentences stand side by side, say whether the second is the cause, the consequence, an
+example or a limit of the first — with the word, not with a dash, a colon or a short sentence left to
+land the point. The fix adds words; pay for them by moving detail to a table or the appendix, not by
+cutting the connectives from another paragraph.