density.connective-overuse
Remove repeated transitions where the connection is already clear.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
Rules catalog
Every rule the engine ships, generated from the engine.
Builtin rules grouped by family. This page is rendered from a data file that a Go program writes by walking the real rule catalog, and CI fails when the committed file and the engine disagree.
Read a row as a proposal. Every builtin rule is experimental: the thresholds are
initial editorial policy, and no corpus precision has been established for any of
them. Enable what your project wants to argue for, and say why in the
configuration file.
severity is presentation. gate is policy: forbid fails a run on any match,
none leaves the decision to the score. group is the correlation group, which
shares a cap across related rules, so it is not the same thing as the identifier
prefix that groups this page. weight and cap are index points.
Inspect any of these from the CLI: unswell rules list, unswell rules show <id>, and unswell rules test to execute the catalog’s own examples.
density
1 rulefiller
16 rulesfiller.announced-importance
State the fact without announcing its importance.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
filler.document-justification
State the information directly; remove redundant document or outcome announcements.
A document subject followed by 'exists so/to/because', deliberate non-repetition of material, a statement that a document part earns its place, a page explaining what its fields mean, editorial ownership, document dedication or future-presentation clauses, or announcements that definitions or counts are repeated or omitted here. Ownership may use a pronoun only after an adjacent explicit document announcement in the same block. A generic possibility announcement may precede two complete conditional branches in the same block. Exact counts, concrete outcome kinds and scoped introductions remain excluded. Storage existence and ordinary scope exclusions do not match. Reports complete matched clauses, once per clause.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
filler.document-metadiscourse
Document self-description: check whether this announcement or navigation helps the reader.
Recognizes document subjects followed by communicative verbs, and 'in this document, we' announcements. Also recognizes bare invitations to see an anaphoric referent in action; named objects and conditions stay excluded. Embedded document subjects identifying themselves as a map of a relative topic receive a self-description explanation; literal maps remain excluded. Groups their complete clauses within eight sentences of uninterrupted prose; one clause is sufficient.
The document-metadiscourse frame is a bounded token/POS heuristic, not a dependency or semantic parse. Clauses are limited to 48 tokens and subjects to eight. Protected text is opaque, never a lexical cue. Recognition does not establish redundancy, authorship, or a need for revision. Technical definitions and useful navigation can match; defaults carry no score or gate. Precision is unqualified.
Off by default: On real pages two reviewers both accepted 60 of 133 findings and both rejected 31. They disagreed on the rest: a statement of a page's scope helps one reader and delays another.
filler.empty-transition
Check whether these repeated opening transitions add information.
Counts nonoverlapping configured phrases in bounded prose windows; one occurrence is allowed by default.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Only configured phrases are recognized; literal meanings and local exceptions require editorial review.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. Its list holds openings such as "with that being said".
filler.evaluative-closure
This clause announces information or endorses a result; state the information directly.
An anaphoric purpose tail ('which is the point of doing it', 'which is what doing it is for'), a bare whole-point/design declaration, an anaphoric assertion of value or verification, an information subject announcing cognitive worth, abstract feature-purpose or intentionality closures, output claiming to prove understanding, or an impersonal modal notice followed by a that-clause. Also recognizes cognitive notices modifying information nouns, gerund-anaphor purpose clefts, relative feature-purpose predicates and discourse subjects declaring importance or editorial integrity. Information judgments may join adjacent evaluative predicates and retain a quoted noun-phrase subject. Discourse subjects admit selecting importance and attention predicates; author notices and document-outcome endorsements qualify. Document endorsements admit helped/helps with a reader object or a bare/to-infinitive action. Gerund action subjects can close with abstract functional clefts. Bare noun/counts compounds remain excluded. Discourse subjects contrasting verification with assertion receive a local evidence-commentary explanation; a bare it requires an information antecedent in its sentence. Concrete quantities, checking actors and explicit methods remain excluded from that construction. Bare information-subject/abstract-value identities retain their written offline/online prefix. Anaphoric abstract recasting contrasts receive a constraint-focused explanation; concrete alternatives and conditions before or after the candidate remain excluded. Attribution is scoped through the candidate clause; later operational reporting does not erase a preceding judgment. Bare anaphoric question and worth predicates retain local spans before coordinated or causal explanations. Deictic conversion appraisals, demonstrative-plan completion metaphors and unqualified abstract value labels receive guidance that preserves operational facts and requires verification before weakening claims. Literal control loops, explicit criteria and operational consequences remain excluded from these constructions. Concrete component purposes and measured evaluations stay excluded; notice and cognitive-worth frames retain their attached conditions. Reports complete matched clauses, once per clause.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
filler.instruction-scaffolding
State the supported action directly; preserve its conditions and optionality.
An affirmative impersonal capability with a transitive configuration or presentation action, or a reader-intention clause introducing an instruction. Anaphoric gerund methods and nominalized methods are also recognized. Nested enables/allows plus an action nominalization and passive infinitive, and nominal actions performed through a gerund method, also qualify. Actor-to-action projection recognizes ability nouns, modal used-to/used-for actions, named method predicates, and nested intended-to-enable actions. Narrated prerequisites and repeated same-actor obligations can introduce an action. Standalone generic reader enablement requires an adjacent method; concrete capability explanations remain controls. Infinitive reader goals admit coordinated actions. Relative operation or actor antecedents can carry nested support; capable-of gerunds and purpose-to-steps announcements also qualify. A second method may be anchored by a repeated two-noun object and explicit reader action. An adjacent anaphoric method may name a nominal syntax or configuration instrument; that instrument cannot establish a standalone warning or replace a passive actor. A discourse-marked unnamed piece, bit or part of logic or code may carry a transitive gerund through allows, enables or supports. Named components, numbered architectural parts and unmarked part descriptions remain excluded. Reader prerequisites with will need to have the following and an installed-state operand request direct wording; separate deadlines, qualified states and optional requirements remain excluded. A complete temporal state prerequisite and impersonal safe/possible action may introduce an immediate To do this/that passive invocation of a named method or function. The pair requests direct instruction wording while retaining the safety qualifier and prerequisite; automatic execution and unnamed operations remain excluded. A reader-intent condition followed by a distinct action with an exact code-form command or option is already a direct optional instruction. The either/or coordinator before alternative actions is not an action verb. A supported-action prefix followed by a modal relative helpfulness clause and an unqualified purpose nominal requests direct wording for both layers while retaining their modality and the relative antecedent. A configuration operand or opaque identifier used as a standalone subject does not establish an actor for a simple modal used-to/used-for capability. Preserve modal meaning, named actors and conditions. Named actor permissions and simple passive predicates do not. Adjacent method announcements remain related evidence. Failure, permission, negation and quoted constructions do not match. Reports complete matched clauses, once per clause. An immediate anaphoric method may be related across adjacent prose paragraphs, without crossing structural boundaries.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples. A conditional reader goal and a be-sure-to reminder are not reported.
filler.modern-world-opening
Replace the generic opening with the document's subject.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
filler.nominal-support
Consider expressing the action or frequency without this support noun.
An inflected make followed by use of an operand, an instrumental by/through/with the use of construction, or a frequency or execution-mode modifier carried by basis or manner. Modified use nouns, legal use restrictions, quantities and unrelated nominal subjects do not match. Reports complete matched clauses, once per clause.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
filler.numbered-section-framing
The introduction and heading emphasize a count instead of naming the section's subject.
A paragraph opens with a count and a generic content noun; the next heading repeats only that count and noun. Reports the opening clause and the related heading. Code, lists, intervening paragraphs, and excluded content break the pair.
An experimental editorial warning, not an authorship detector or a count consistency check. Recognizes two through twenty in words and 2 through 99 in digits. Limited to lines, steps, rules, points, things, ways, and tips. Explicit requirement and negation cues are excluded. The rule cannot establish whether an unstated count is technically necessary. Opening sentences are limited to 96 tokens, matched clauses to 48, and headings to four. A named task in the heading, a standalone numerical hook, and nonadjacent headings are outside its scope.
filler.section-announcement
Replace repeated section announcements with their subject.
Counts nonoverlapping configured phrases in bounded prose windows; one occurrence is allowed by default.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Only configured phrases are recognized; literal meanings and local exceptions require editorial review.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. Its list holds openings such as "in this section, we will".
filler.stacked-hedging
Check whether each hedge expresses a distinct uncertainty; retain needed qualifications.
Counts distinct configured modal/adverb cues in one clause; commas, punctuation, conjunctions, and code end a clause.
Alpha defaults are editorial policy proposals; corpus precision has not been established. A token boundary is not a dependency parse. Multiple independent uncertainties can be justified.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. It counts hedges stacked inside one sentence.
filler.unnamed-numerical-choice
Name the choice instead of introducing a count of unnamed answers or options.
A prose block opens with a count and answers/options/choices, followed by ', and/but the right/best one depends on' and a stated selection condition. Reports that clause; the actual task is left unnamed by the opening.
Experimental editorial policy, not authorship attribution, enumeration validation, or established population precision. Recognizes two through twenty in words and 2 through 99 in digits, only in the first sentence of a prose block. Clauses are limited to 48 tokens. Named alternatives, task headings, lists, requirements, quotations and questions do not qualify. Other numerical hooks remain unsupported. Keep the selection condition when naming the actual choice.
filler.unscoped-assurance
State the behavior, scope or evidence behind this quality claim.
A positive copular clause announces intentional quality without a mechanism, calls a choice nearly always right, or promises the fastest/best/easiest way to understand everything. Also recognizes affirmative extreme quality predicates and unrestricted claims about what everybody or nobody knows, wants or can verify, unsupported majority claims about reader preferences or practice, and anaphoric result tails predicting generic readers' preferences or understanding. Ordinary copular quality or difficulty judgments require a main subject and a complete predicate without a stated local mechanism. Testing-to-guarantee and unrestricted-trust constructions receive explanations of their specific assurance. Grouped occurrences use a specific explanation only when all occurrences support the same explanation; mixed groups retain the general diagnostic. Technical adjective-noun phrases, conditional, quantified and protected constructions are excluded. Reports complete matched clauses, once per clause. Local emphatic capability modifiers, optimization degree, attributive ease-to-use software descriptions, and deictic superlative benefits request criteria for their extent or ranking. Bare quality fragments may occur in headings or lists. Local measurements, criteria, comparisons, mechanisms, negation, attribution and quoted operators remain excluded; a bare count in an adjacent sentence and a reader-intent condition do not establish a quality criterion. Formal technical properties are not inferred from degree words alone. Testing or review activities guaranteeing abstract quality, and bare assertions that nothing is taken on trust, also request explicit scope. Concrete invariants, proofs, qualifications, attribution and protected code are excluded. Source-local quality claims about future-proofed APIs, unbounded feature-set rankings, paired subjective software qualities, reasonable defaults and important effects on quality request a criterion or concrete behavior. Named comparison sets, local mechanisms and quotations are excluded.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
Off by default: On real pages two reviewers both accepted 14 of 48 findings and both rejected 21. A neighboring sentence explained most of the rejected claims, and a local rule does not read it.
filler.unspecified-update
Identify the assets and version behind this update announcement.
Matches a standalone affirmative update or upgrade of an indefinite asset set to an unnamed latest version. A second clause, another sentence in the same block, a stated version, attribution, quotation or condition excludes it. The rule does not determine whether an update happened or infer its destination version. Reports complete matched clauses, once per clause.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
filler.weak-intensifiers
Check whether these intensifiers add a measurable distinction.
Counts configured adverbs before adjectives/adverbs per 100 prose words, after the minimum word count.
Alpha defaults are editorial policy proposals; corpus precision has not been established. POS roles are heuristic. The rule does not require deleting distinctions with technical meaning.
Off by default: 0.019 per thousand words in human prose against 0.007 in the controlled arm.
filler.wordy-phrase
Consider a shorter phrase that preserves the condition or meaning.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
format
2 rulesformat.em-dash-density
Em dashes exceed the selected local density; punctuation can be intentional.
Counts U+2014 in selected prose per 100 prose words, with mapped source ranges including encoded Markdown entities.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This surface measurement does not establish a grammatical error, a need for revision, or authorship.
Off by default: Surface measurement. It is the one measure elevated in generated prose, 0.070 per thousand words against 0.001 in human prose, so the strict profile turns it on. Elsewhere it stays opt-in: it reports density, not an error.
format.list-fragmentation
Several short lists fragment this section; check whether the grouping helps readers.
Counts complete short flat unordered lists in a bounded section window using grammar-derived list ownership.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This surface measurement does not establish a grammatical error, a need for revision, or authorship. Ordered/task/nested lists, sentence items, procedure openings, identifiers, and selected reference headings are excluded.
Off by default: Surface measurement of list shape. No occurrence in 5.6 million measured prose words, human or generated.
hype
4 ruleshype.absolute-claim
Check whether this guarantee has conditions or exceptions.
Matches configured phrases in affirmative prose, excluding questions and explicit qualifications conservatively.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This lexical screen does not determine truth, semantic vagueness, or whether every condition is stated.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. It looks for a guarantee stated without conditions.
hype.metaphor-cluster
Replace clustered stock metaphors with specific properties.
Counts nonoverlapping configured phrases in bounded prose windows; one occurrence is allowed by default.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Only configured phrases are recognized; literal meanings and local exceptions require editorial review.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. Its list holds stock metaphors such as "rich tapestry of possibilities".
hype.modifier-cluster
Replace the cluster of evaluative modifiers with specific properties.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
hype.vague-praise
Replace general praise with a specific property or measured result.
Matches configured phrases in affirmative prose, excluding questions and explicit qualifications conservatively.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This lexical screen does not determine truth, semantic vagueness, or whether every condition is stated.
policy
1 rulepolicy.banned-phrases
Remove a phrase forbidden by this editorial policy.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
readability
2 rulesreadability.grade-metric
The ARI formula exceeds the selected level; consider the audience and keep necessary technical terms.
Applies 4.71*characters/words + 0.5*words/sentences - 21.43 to local extracted prose words, excluding identifier-shaped tokens and counting each part of a hyphenated compound as a word.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This surface measurement does not establish a grammatical error, a need for revision, or authorship. Characters are Unicode letters and digits within prose tokens. A token with an inner uppercase letter, a digit, an underscore, a period, or a slash is an identifier and is not counted. No age, comprehension, quality, or authorship is predicted.
Off by default: Surface measurement of sentence and word length. It runs at 2.964 per thousand words in the controlled arm against 0.711 in human prose, but it was frozen hypothesis H2 and returned +0.32 points on the confirmation partition at p 0.747. Length is not wording.
readability.long-paragraph
Consider paragraph boundaries while keeping each condition with the statement it qualifies.
Requires both paragraph length and a configured count of long sentences; reports local descriptive measurements.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This surface measurement does not establish a grammatical error, a need for revision, or authorship. Length, lexical diversity, and POS ratios are descriptive; reference material can justify them.
Off by default: 0.177 per thousand words in human prose against 0.040 in the controlled arm.
repetition
16 rulesrepetition.adjacent-word
This word occurs twice in succession; check for an accidental duplicate.
Locates a prose word written twice in succession, separated only by whitespace. The two copies have the same case; a sentence may open with the capitalized copy. A repeated capitalized word inside a sentence is read as a name or a title. In Markdown, emphasis, code, a closing link, or a table boundary between the copies separates them. Quoted examples, protected text, repeated adjectives/adverbs, and ambiguous grammatical repetitions are excluded. Quotation protection follows the repeated tokens, including quotes spanning sentences in one block. Numeric operands and formatting conversions are not prose words; uppercase Boolean operator names remain distinct from their lowercase grammatical uses.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Deliberate repetition and technical compounds can still match; this is not an authorship signal. Unclosed quotations and ambiguous plural possessives inside single quotes protect the remaining passage.
repetition.definition-echo
The definition repeats its subject; state the distinguishing information directly.
A positive copula repeats the same nominal phrase, ignoring articles, immediately before ', not' and a nominal alternative. Also recognizes a still-qualified repetition of the same nominal or gerund subject, including a bounded both-object ellipsis. Added modifiers, conditions and code identifiers are excluded; bare uncontrasted identities stay outside the rule. Reports complete matched clauses, once per clause.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
repetition.duplicate-list-item
This complete list item is repeated in the same list; check whether both entries describe the same change.
Compares complete simple unordered items within one grammar list and section, preserving case, punctuation and references.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Only a single identifier may compare across plain text and inline-code formatting. Other protected content, ordered/task lists and complex items are inapplicable. Identity is not a semantic paraphrase test.
repetition.exact-sentence
This sentence is repeated; consider whether every occurrence is needed.
Compares sentences within one structural section, with separate table-cell, leading-condition, and Markdown details scopes. Protected operands are compared by literal source identity without interpreting their contents.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Protected operands do not contribute to the prose word minimum. Identity comparison shares a bounded work budget.
repetition.explanatory-restart
The explanation repeats a judgment; state it once and name the concrete cause.
A nominal subject and a copular adjective are repeated as 'and it/they is/are ADJECTIVE because'. The repeated copula and adjective must agree; negation, quantities, conditions and ambiguous noun antecedents are excluded. A reason-subject setup also includes a final abstract optimization or design judgment repeating the same adjective. Reports complete matched clauses, once per clause.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
repetition.heading-echo
The first paragraph closely echoes its heading; check whether it adds information.
Compares a selected heading with its immediately following paragraph using exact set Jaccard.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Lexical and POS signals do not establish repeated meaning; technical exceptions require a declared policy. Short definitions may be useful; code and intervening blocks break adjacency.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. It compares a heading against the paragraph below it.
repetition.near-sentence
These sentences have high lexical overlap; check for repeated information.
Compares indexed sentence candidates with set Jaccard and a shared technical-contrast signature.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Modal, condition, and state cues are retained in token order. Numeric-unit protection retains the token after a number, not a complete unit expression. Equal signatures do not establish semantic equivalence; comparisons stay within structural and task scopes. Separate Markdown details choices have independent comparison scopes.
repetition.ngram-density
This phrase recurs in a short prose window; check whether each use is needed.
Groups repeated consecutive prose n-grams in sentence windows; longer covering phrases take precedence.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Lexical and POS signals do not establish repeated meaning; technical exceptions require a declared policy. Activation counts occurrences; reported coverage uses the prose words between the first and last matching sentence.
Off by default: Fires an order of magnitude more on human prose, 0.603 findings per thousand words against 0.056 in the controlled arm. Parallel technical wording is often deliberate.
repetition.paragraph-openers
Several paragraphs repeat the same opening formula.
Groups opening words from each paragraph's first sentence. The word minimum applies to the complete paragraph.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Repeated openings do not establish duplicate meaning. Structural sections and leading conditions bound comparisons. Separate Markdown details choices have independent groups.
Off by default: Its findings fall mainly on deliberate parallel wording. On the assistant-panel corpus two reviewers both rejected 32 of 39 findings and both accepted none. Repeated openings do not establish duplicate meaning.
repetition.paragraph-overlap
These prose blocks have high lexical overlap; check for repeated information.
Builds connected clusters of Jaccard-qualified block pairs through an expiring inverted index.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Lexical and POS signals do not establish repeated meaning; technical exceptions require a declared policy. The block window bounds each pair, not the complete cluster; not every pair in a connected cluster must qualify.
Off by default: 0.283 per thousand words in human prose and none in the controlled arm.
repetition.redundant-predicate
The predicate repeats the relation already named by its subject.
A category/type label describing its own unqualified category/type, or a path/location subject followed by is located/situated at/in, or a reason subject followed by is because. The grammatical subject must name the repeated relation; files, identifiers and control-flow conditions do not qualify. Reports complete matched clauses, once per clause.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
repetition.redundant-support
This clause marks the same relationship twice; check whether one marker is enough.
An additive also with a clause-final as well, an in-addition-to introduction with also in the same main clause, a like-introduced example immediately restated by for example, or adjacent then finally sequencing. Separate finite or to-infinitive predicates and literal as-well-as comparisons do not qualify. Reports complete matched clauses, once per clause.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
repetition.repeated-claim
These passages repeat an assertion; check whether the explanation can be shared.
Compares short complete assertions, assertions containing opaque inline-code atoms, and prefixes before a nonrestrictive ', which' clause, within one section and sentence window. Case, operands, punctuation, negation and conditions remain part of the exact identity. Also compares adjacent explicit reformulations of an only restriction as rejection of its negated property. Actors, predicates, tense, properties and objects must agree. A bounded possessive dependency-role conversion compares an owner's license restriction with its refusal of code not licensed under that property. A closed definition-benefit grammar compares the same opaque source and the same cross-page meaning invariant. A contribution section can connect an explicit add instruction and map-rendering statement to that source.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Requires a surface finite-verb cue; it does not resolve pronouns or prove general semantic equivalence. Questions, quoted text, lists, cells, arbitrary code, and claims over 64 tokens are excluded. Full sentences meeting the default twelve-prose-word minimum remain with repetition.exact-sentence. A relative-clause prefix must match a complete assertion, not another partial prefix. Restriction reformulations require That is or In other words, stay inside one block, and reject quantities, conditions, quotations, protected operands and unresolved scope or modality. Definition benefits accept only complete unqualified forms in prose paragraphs; they reject other headings, containers, unresolved map references and changed source operands. Other paraphrases remain unsupported.
repetition.sentence-openers
Several sentences repeat the same opening; consider removing repeated setup.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Groups stay within structural sections and explicit leading conditions. Separate Markdown details choices have independent groups. Repeated form need not mean redundant information.
Off by default: Its findings fall mainly on deliberate parallel wording. On the assistant-panel corpus two reviewers both rejected 56 of 77 findings and both accepted four. Repeated form need not mean redundant information.
repetition.summary-echo
This summary has high lexical overlap with preceding prose; check what it adds.
Compares paragraphs under explicitly configured selected summary headings with preceding nonsummary paragraphs.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Lexical and POS signals do not establish repeated meaning; technical exceptions require a declared policy. Summary scope follows grammar-derived heading ancestry. Other titles require configuration; lexical overlap can be intentional.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. It compares a closing section against the body it summarizes.
repetition.syntax-template
Several sentences repeat a POS shape; check whether the structure serves the text.
Groups complete normalized POS shapes with at least two lexical realizations; imperative openings and lists are excluded.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Lexical and POS signals do not establish repeated meaning; technical exceptions require a declared policy. POS grouping is a surface heuristic, not a dependency analysis or a claim that the facts are duplicated.
Off by default: 0.010 per thousand words in human prose and none in the controlled arm.
scaffold
4 rulesscaffold.ai-self-reference
Remove the assistant's self-reference.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
scaffold.chat-preamble
Start with the subject instead of a conversational acknowledgment.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
scaffold.dive-in
Start with the subject instead of announcing the explanation.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
scaffold.follow-up-offer
Remove the conversational follow-up offer from this document.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
syntax
12 rulessyntax.long-sentence
Consider splitting this long sentence without losing its conditions.
Reports a sentence with more prose words than the onset. Table cells are not measured.
Alpha defaults are editorial policy proposals; corpus precision has not been established.
syntax.nominalization-chain
Consider a direct verb while preserving the subject, conditions, and technical meaning.
Matches a configured weak verb, a configured common noun, and an of-complement with a common-noun head.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This surface measurement does not establish a grammatical error, a need for revision, or authorship. Inflected verb forms are explicit dictionary entries. Suffixes alone never activate this rule.
Off by default: Effectively silent: 0.001 findings per thousand words in human prose and none in the controlled arm.
syntax.not-only-density
Repeated paired contrasts make this passage formulaic.
Counts ordered not-only/not-just followed by but in bounded prose windows, allowing one matching sentence.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Version 3 retains prose around inline code; candidates cannot cross protected tokens or structural boundaries.
syntax.noun-stack
Review this common-noun sequence; keep established technical terms.
Counts NN modifiers followed by an NN/NNS head within a shallow NP chunk, stopping at terms, identifiers, and configured verb forms in predicate position.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This surface measurement does not establish a grammatical error, a need for revision, or authorship. A shallow NP is not a dependency tree, and a noun sequence can be an appropriate domain term. Interior NNS tags cause abstention: they can be plural modifiers or misclassified finite verbs. The negative modal "cannot" breaks a noun run even when tagged NN; other tagging ambiguity remains possible. Configured verb forms are literal words; a form that is also a noun modifier shortens a genuine stack. Forms also listed in nouns remain candidates after nominal modifiers or common nouns; leading forms can break a run. Coordinated predicates require a following verb or modal; noun coordination stays intact.
Off by default: Version 4 did not establish a cross-family association in the complete-document study. Version 5 repairs noun/verb boundaries but has no independent editorial qualification.
syntax.paired-contrast-density
Check whether repeated paired contrasts carry distinct information.
Counts paired contrasts in bounded prose windows: 'rather than' and 'instead of' and comma-not alternatives within a sentence, and adjacent 'It is not about'/'It is about' sentence pairs including contractions.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Only the declared structures are recognized; no semantic equivalence is inferred. A single contrast is ordinary English and the window allows one; the rule measures repetition of the shape, not its use.
Off by default: Surface measurement of contrast frequency, not redundant information. On 36 complete exposed development pages, assistant review accepted one of 30 findings; 27 were rejected and two uncertain. Necessary technical distinctions dominate its review load. The unchanged matcher remains available through explicit configuration.
syntax.parenthetical-load
Review the load of balanced insertions; retain conditions, qualifications, and necessary definitions.
Measures nonexempt words, pair count, and nesting of balanced parentheses and square brackets within a prose block.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This surface measurement does not establish a grammatical error, a need for revision, or authorship. Unbalanced or protected-crossing frames are not matched. Depth is capped at 256 with an explicit error. An insertion below min_insertion_words, such as an issue reference or a license label, contributes to none of the three measurements, including nesting.
Off by default: Fires more on human prose, 0.499 per thousand words against 0.300 in the controlled arm.
syntax.passive-candidate-density
Several sentences contain be-plus-participle candidates; review voice only where the meaning allows it.
Counts sentences containing a be verb, up to four adverbs, and VBN in bounded contiguous prose windows.
Alpha defaults are editorial policy proposals; corpus precision has not been established. This surface measurement does not establish a grammatical error, a need for revision, or authorship. The same POS shape can describe a state. This rule is not syntax.passive-density and does not establish grammatical voice.
Off by default: Names candidates without a dependency parse, and says so in its own limitations. It runs at 1.953 per thousand words in the controlled arm against 1.112 in human prose.
syntax.repeated-reframing
Repeated denial and redefinition: review whether both clauses add information.
Groups adjacent copular denial/redefinition clauses, or a do-supported negative action followed by a finite positive action, with a repeated subject or anaphoric pronoun. Action pairs exclude attributed and quoted claims. A singular pronoun and determiner-led object can resolve a plural-noun tag on an s-ending finite predicate. Each pair stays in one block; repetition is counted within eight sentences of uninterrupted prose.
The denial-redefinition frame is a bounded token/POS heuristic, not a dependency or semantic parse. Clauses are limited to 48 tokens and subjects to eight. Protected text is opaque, never a lexical cue. Recognition does not establish redundancy, authorship, or a need for revision. Technical definitions and useful navigation can match; defaults carry no score or gate. Precision is unqualified.
syntax.rhetorical-question-density
Check repeated question-and-short-answer transitions.
Counts configured complete questions followed in the same block by a short nonquestion answer.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Arbitrary questions are not classified. Numeric/code answers are excluded; disable for FAQ paths.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. It counts questions a paragraph asks and answers itself.
syntax.slogan-contrast
The contrast names abstract actions without their objects; name the check and its failure condition.
A short nominal subject and a bare check/verify/prove verb opposed to a bare trust/assume/assert verb. Both verbs must lack objects and qualifiers; concrete alternatives and refusal/guessing contrasts are excluded. Reports complete matched clauses, once per clause.
Bounded token constructions, not a semantic or authorship classifier. Candidates are limited to 48 tokens; local tails may occur in sentences of up to 96 tokens. Protected text is never construction vocabulary. Warnings request an editorial review and do not prove that a sentence can be deleted. Defaults are experimental policy proposals; review results describe only their declared samples.
syntax.triad-density
Check repeated triads of evaluative modifiers against specific properties.
Counts three-item lists from the evaluative dictionary with at least two adjective tags; one triad is allowed.
Alpha defaults are editorial policy proposals; corpus precision has not been established. One noun tag is tolerated for an ambiguous modifier. This is a POS/dictionary heuristic, not a dependency parse.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. It counts three-item lists used as a rhythm device.
syntax.whether-preface-density
Reduce repeated whether-you prefaces while preserving alternatives.
Counts whether-you-are prefaces containing 'or' before a comma within 32 tokens; one occurrence is allowed.
Alpha defaults are editorial policy proposals; corpus precision has not been established. Recognizes one opening template, not arbitrary conditional clauses.
Off by default: No occurrence in 5.6 million measured prose words, human or generated. It counts prefaces of the form "whether or not X".