v0.9.0

sloplint 0.9.0 and the first release of the companion gem, sloplint-judge 0.1.0. Both are on rubygems: gem install sloplint sloplint-judge.

Added

• sloplint check --judge runs the rules of sloplint-judge alongside the
  regex catalog and merges the notes in document order. The judge is a second
  gem built from this repository (sloplint-judge.gemspec, exe/sloplint-judge)
  whose rules are questions put to a System One model. sloplint loads it by
  name only and exits 2 with an install hint when it is missing. A backend
  failure under --judge is exit 3 and writes no notes. See docs/JUDGE.md.
• Sloplint::Split, a paragraph and sentence splitter that keeps offsets into
  the source, and Engine.context_window, the note's context drawn for any
  span rather than only a regex match. Both are used by the judge.
• script/calibrate, which measures a judge backend against RAID and a
  current-model side generated from RAID's own prompts, and reports the
  pass lines from docs/JUDGE.md.

Fixed

• --markdown opens and closes a fenced code block only at the start of a
  line, at any indent, so a fence under 10. or a nested bullet still
  pairs. A fence quoted inside a sentence (```) used to open a
  block there, and every fence after it paired wrong for the rest of the file.
• The splitter also drops indented code blocks (four spaces or a tab), lines
  that are one HTML tag, and YAML front matter under --markdown, so none of
  them reach the judge as prose.

sloplint-judge 0.1.0

The judge is a second gem built from this repository. Its rules are questions put to a System One model (Jev, from TypeSafe) rather than regexes: does this paragraph name anything a reader could check, does this sentence tell its reader anything new. It adds sloplint check --judge and a sloplint-judge executable, needs an API key, and is the only part of sloplint that sends text off the machine. Fourteen rules in two categories; see docs/JUDGE.md.

Full notes for both gems are in CHANGELOG.md.

v0.8.1

Added

• cataphoric-teaser (self-rating, warning, high confidence): the
  forward-pointing tease that sells a claim as rare knowledge before making
  it. "Here's what nobody tells you about hiring.", "The part most people get
  wrong is the rollback." Three closed frames (a forward demonstrative, a
  noun slot, a scarcity subject and a verb of telling); the bare-subject
  form must reach a copula or colon, and "that's what nobody understands"
  points backward and is left alone.
• punch-sentence (cadence, info, medium confidence): the verbless beat of
  one to three words wedged between two long sentences in the middle of a
  paragraph, "Not anymore." and "Simple.", which mic-drop-closer and
  bare-auxiliary-closer only see at a paragraph's end. Two closed shapes:
  a negator with up to two words after it, or one capitalised word of five
  letters or more. Titles, initials and case citations before the stop are
  guarded, and short clauses with a subject ("He agrees.") are not matched.

v0.8.0

[0.8.0] - 2026-09-15

Changed (breaking)

• The four old categories are gone: rhetorical-tic, structure, puffery
  and hedging. Nine new categories replace them, each named for the move
  the construct makes: self-rating (the writer grades their own prose or
  claim), closer (closes by restating or announcing the point), cadence
  (rhythm: repetition, parallelism, the long-then-short kicker), puffery
  (inflates the subject), false-correction (corrects a reading nobody
  offered), false-concession (performs balance or candor and gives nothing
  up), reader-address (instructs or flatters the reader),
  borrowed-metaphor (an engineering term applied to an argument), and
  punctuation (the mark itself). Every note, --select, and --ignore now
  use these nine names; the four old names are gone, not aliased, so a
  script still passing one gets a usage error instead of a silent no-op.
• The rule ellipsis-closer is renamed bare-auxiliary-closer. Nothing
  about its pattern, message, or fixtures changed, only the id.
• severity now rates only what the construct costs the prose (error,
  warning, info). A new field, confidence, rates how likely a match is
  a false positive (high, medium, low), and a low rule replaces what
  used to be default_on: false — it stays out of a normal run and needs
  --strict or its own id named directly to turn on.
• sloplint rules --json drops the default_on field and adds confidence
  and rationale for every rule. Each note in check output also gains a
  confidence field alongside severity.
• --select <category> no longer turns on that category's low-confidence
  rules. Naming a category now selects only its default-on rules; to run a
  low-confidence rule, pass --strict or name the rule's own id.
• Every rule's severity was re-rated for what the construct costs the prose,
  now that confidence carries the false-positive question on its own. The
  catalog reads 11 error, 56 warning, 13 info. Nothing shipped at
  error before: the eleven that do now are the ones that hurt a sentence in
  any register, the puffery family (puffery-words, rich-tapestry,
  vital-role, impact-noun-vague,
  trailing-significance-participle), the tautology and self-ranking moves
  (the-x-is-the-x, cleanest-x, most-honest-x), the fake balance
  (two-things-true, none-of-this-is-to-say) and vague-attribution.
  Twenty rules moved from info to warning, among them em-dash-overuse:
  one em dash is a style choice, three in a paragraph is the tell, so the
  dense form now costs more than em-dash. Any tool that keys off severity
  sees different values for the same text.

Fixed

• The test suite failed with no locale set, because spec/plugin_spec.rb
  read files without forcing UTF-8. Fixed to read as UTF-8 regardless of the
  environment's locale.
• The README and SPEC quoted an outdated rationale for no-x-no-y and an
  outdated sample of sloplint check output. Both are corrected and now
  pinned by tests, so a future drift between the docs and the code fails the
  build instead of sitting unnoticed.
• SPEC's description of the Rule model and of the package's file layout
  had fallen out of step with the code. Both are corrected to match.
• --help was missing count from its list of note keys, even though a
  note already includes it for rules that tally items. Added.

Added

• A test asserting that no rule id ever matches a category name, since
  --select and --ignore resolve a name against both and a collision
  would make one shadow the other.

v0.7.0

Added

• --strict runs the full catalog, including the five rules that ship off
  by default. Before this, the only way to turn one on was to name it in
  --select, which also turned off every other rule unless you listed
  those too. --ignore still subtracts on top of --strict, and
  --select is unchanged.
• ellipsis-closer (structure, info) flags the long-sentence-then-short-closer
  kicker mic-drop-closer catches, but built on the opposite verb shape: the
  closer's verb phrase is elided down to a bare auxiliary with no object,
  "The agent did." rather than a quantifier plus a full verb. No subject list
  is needed — the tell is the missing verb, not the subject — so "That",
  "This" and "It" are not excluded the way mic-drop-closer excludes them.
  A negative lookahead drops a closer that still holds a wh-word, since that
  marks a subordinate clause with its own complement rather than an elided
  one. Reuses mic-drop-closer's long-sentence prefix.
• actually-not-x: the adverb and a trailing "…, not X" in one clause, as in
  "The disclosure actually covered two incidents, not one". Either marker
  alone is ordinary technical English and neither is matched on its own; the
  tell is the two of them doubled up, correcting an alternative the reader was
  never offered. The narrowing is structural, not a word list: the comma in
  front of "not" has to be the first comma of its clause and has to follow a
  word, so a fronted setup ("Despite the name, …") and a parenthetical both
  drop it. Nothing crosses a line break, so two rows of a table cannot weld
  into one correction.
• real-x-real-y (rhetorical-tic, info) flags the same "real" used twice
  attributively in one sentence, in front of two different nouns: "it can
  trigger real API calls or hold real credentials". Bare "real" is an
  ordinary word, so the narrowing is the repetition itself, not a noun
  list. A hyphen on either side takes it out ("real-time", "non-real"), a
  backreference requires the second noun to differ from the first, and a
  short closed list drops the fixed senses of this register (real time,
  real-world, real numbers, real user monitoring, real estate, real money).
• phrase-echo (rhetorical-tic, info, off by default) flags a three-word
  phrase that comes back within a few hundred words, the model reusing a
  phrase it has just minted. Each word must be four characters or more and
  lowercase-led, and one of them six letters with nothing but letters, so "in
  order to", "couldn't have been" and "New York City" never match. The window
  does not cross into a list item or a table row. Off by default: a term of
  art repeats because it must, and in reference prose the pattern fires on
  names by the thousand per million words.
• trailing-restatement (structure, info, off by default) flags the tail
  that says the sentence again in other words: ", which means …", ", which
  is to say …", ", meaning …", and four participles that hang a result off
  the sentence (", making it easier", ", allowing us to", ", giving them
  more", ", leaving you with"). The connective is all a pattern can see, and the
  same connective states a real consequence in careful prose, so the rule
  runs only when selected.

Fixed

• check's options now work without the command word: sloplint --markdown -
  and sloplint --strict - scan, as the README has said sloplint - does.
  Before, the global parser rejected any option it did not own with
  invalid option, so check had to be spelled out whenever an option
  came first. An option no command knows still exits 2.
• --markdown now blanks <!-- --> comment blocks along with fenced code,
  inline code and URLs. Planning notes kept in comments never render, so
  every rule that fired inside one was a permanent false positive.
• rule-of-three's message, rationale, README and SPEC entries said it
  flagged three parallel comma items. The pattern requires single words in
  the closing two slots, so a triad of phrases never matched; the text now
  says that.

v0.6.0

The catalog grows from 50 rules to 75, and sloplint ships as a Claude Code
plugin as well as a gem.

Added

• sloplint now ships as a Claude Code plugin as well as a gem. Add the
  marketplace benjaminjackson/sloplint and install sloplint; the plugin
  carries the Ruby source with it, so nothing is installed from a package
  index and it works where gem install cannot reach the network — Claude
  Cowork above all. The plugin adds one skill, /sloplint:check, which runs
  the scanner and reports what it returns. The skill refuses to assess prose
  on its own when the scanner cannot run, because a review with no scan
  behind it reads exactly like one with a scan behind it.
• not-by-x-but-by-y: the corrective frame on a repeated preposition, "not
  by luck, but by design". The copula rules cannot see it because nothing
  anchors "not" but a verb or a dash, so the anchor is the preposition, which
  must repeat after "but". Ships at info: Thoreau and Melville write it too.

Fixed

• check now reads files and stdin as UTF-8 whatever the locale says. A
  sandbox with no LANG set — Claude Cowork's, among others — leaves Ruby's
  default external encoding at US-ASCII, and the first em dash in a draft then
  failed the whole scan with invalid input: invalid byte sequence in US-ASCII. Prose is the only input sloplint takes, so UTF-8 is the
  assumption rather than the locale's guess. Genuinely invalid UTF-8 is still
  exit 2.

Changed

• thats-the-whole now takes value and fix after "that's the whole",
  alongside point, game, thing, deal and story. The contracted
  closer is the one an agent writes in technical prose ("That's the whole
  value of the info tier", "That's the whole fix"), and is-the-whole-x
  sees only the uncontracted "is", so the sentence slipped both rules. The
  uncontracted "That is the whole fix." moves with it, from info under
  is-the-whole-x to warning. The two new nouns must end the sentence or
  the paragraph, or run into one of the words a closer trails off on, so
  thats-the-whole leaves "value chain", "value-add" and "fix list" alone
  however the line wraps (is-the-whole-x still reads the uncontracted
  "That is the whole value chain" at info, as it did), and a curly
  apostrophe now counts as a contraction. Both rules interpolate one WHOLE_CLOSERS
  fragment, so the lists cannot drift, and a spec walks every noun through
  every form the two wrappers treat differently. Neither noun appears after
  "that's the whole" in 1.55M words of public-domain prose.

• Empty or whitespace-only input
  is now exit 2 with empty input: nothing to check in ..., instead of exit 0. Reporting a scan of nothing as a clean
  scan is the same trap a mistyped rule id sets, and gets the same answer. The
  text is tested before --markdown blanks code and URLs, so a file holding
  only a fenced code block still exits 0; and only an all-empty set of sources
  is an error, so one empty file among several named ones is left alone.

• The executable moved from bin/sloplint to exe/sloplint. claude.ai
  rejects a plugin with a top-level bin/ directory, which would block
  organization distribution. gem install sloplint is unaffected.

• exe/sloplint now checks RUBY_VERSION before loading anything and aborts
  with an explanation, rather than raising NoMethodError from Data.define.
  The Ruby macOS ships at /usr/bin/ruby is 2.6 and hits this.

New rules

• from-x-to-y-chain (rhetorical-tic, warning) flags two or more
  comma-separated "from X to Y" spans in a row: "from guessing to measuring,
  from hoping to knowing". Each span names two poles and nothing between
  them, and stacking them sweeps across a change without describing it. Two
  human shapes are skipped: the relay, where each span starts where the last
  ended ("from the egg to the worm, from the worm to the fly"), and the
  reduplication, where a span has the same word at both ends ("from hummock
  to hummock"). Operands must open with a letter, so a list of ranges ("from
  1990 to 1995, from 1997 to 2001") is not a chain. With those out it flags once in 2.4M words of pre-2022
  Hacker News (a geographic sweep) and once in 1.25M words of public-domain
  prose (Joyce).

• one-x-one-y (rhetorical-tic, warning) flags three or more "one X" items
  in a comma chain that stands on its own: "One owner, one repository, one
  weekly prune." The word does no counting; it sets a rhythm. The chain must
  open a sentence or follow a colon, because after a verb ("the flat has one
  bedroom, one bathroom, one balcony") the word is counting. Items are
  letter-led, so an enumeration over numbers is out; the distributive "one
  for you, one for me, one for the pot" is skipped at any length. Nothing in
  2.4M words of pre-2022 Hacker News or 1.25M words of public-domain prose.

• everyone-nobody (structure, warning) flags the comma-spliced antithesis
  on quantifier subjects: "Everyone wants the dashboard, nobody maintains
  it." The
  balance is what makes the diagnosis sound settled. The comma splice is
  required (with "and" it is a sentence, with a period it is two), the two
  subjects must differ in polarity, and the second clause must close the
  sentence. Nothing in 2.4M words of pre-2022 Hacker News or 1.25M words of
  public-domain prose.

• and-what-it-should (rhetorical-tic, warning) flags the elliptical tail:
  "List what the assistant knows about the client, and what it should." The
  second clause borrows its verb from the first and closes on a bare modal
  or a negated auxiliary, so the sentence ends on a contrast it never
  states. The comma, the conjunction, and the full stop right after the
  modal are all required; the affirmative copula and do-verb ("and what he
  does.") are complete clauses and stay out, and so does a question.
  Nothing in 2.4M words of pre-2022 Hacker News or 1.25M words of
  public-domain prose.

• abstract-lives-in (rhetorical-tic, info) flags an abstraction given an
  address: "the craft that lives between the two desks", "its context lives
  in a folder nobody else can open", "the value sits in the follow-up". The
  subject list is closed and abstract, so people and dogs living and sitting
  places never match, and a capitalised subject is a proper noun. "with" is
  left out of the prepositions because "the decision sits with the board" is
  ordinary English for who is responsible, "at" because "the value sits at
  ten million" is a quantity, and "lies in" because "the problem lies in
  the assumption" is where a fault is. It ships at info because the same
  shape says where information literally is ("the knowledge lives in our
  heads"), and a sample of pre-2022 Hacker News biased toward the
  construction turns up a few of those per million words, all human. A
  draft that keeps giving ideas addresses should be read as a warning.

• np-fragment-and (structure, info) flags a whole sentence made of two
  noun phrases and an "and": "A named owner and a quarterly review." It is
  the fix half of a model's problem-then-fix pair with the verb left out,
  and its usual habitat is a bulleted list, so a list marker may open it.
  Each phrase is one to three words, and no auxiliary or modal may appear,
  contractions included. It ships at info because a lexical verb is
  invisible to the pattern: "A car and a truck collided." has the same
  shape and flags. That sentence is rare in the corpora (nothing in 2.4M
  words of pre-2022 Hacker News, once in 1.25M words of public-domain
  prose), but it is a sentence, so one flag is a question; a draft full of
  them should be read as a warning.

• the-x-is-the-x (rhetorical-tic, warning) flags the repeated-head
  equative: "the reason it holds up is the reason the other half happens",
  "the problem with the tool is the problem with the team". A backreference
  catches the same abstract head noun on both sides of the copula, so the
  sentence equates two things while naming neither. The head list holds
  only nouns that cannot name an object, since "the key to the front door
  is the key on the red fob" is an identity statement; the two heads must
  share a clause, so an earlier "the cost was low, but shipping is the
  cost" never pairs; and the second head must be followed by a preposition,
  determiner, quantifier, pronoun, plural noun or punctuation, so "the
  answer key" is not a repeat. Once in 2.4M words of pre-2022 Hacker News;
  nothing in 1.25M words of public-domain prose.

• same-determiner-chain (rhetorical-tic, info) is the quiet cousin of
  one-x-one-y: three or more comma-separated items opening on the same
  determiner or quantifier, "every file, every branch, every deploy", caught
  with a backreference. The narrative possessives (my, his, her, their, its)
  are left out, since "his fame, his position, his life" is every
  novelist's. It ships at info because the device is one humans use on
  purpose: about 8 per million words of pre-2022 Hacker News and 14 per
  million
  in public-domain prose. One is a question; several in a draft should be
  read as a warning.

• quip-question (structure, info) flags the verbless question that opens
  a pitch: "No invite?", "New to the tool?", "Still stuck?". It must start a
  sentence, open on one of a short list of words, and close on the question
  mark within four more words with no auxiliary or contraction, so a real
  question stays out; "Need" and "Want" are left off the list because "Need
  help?" is a question with its verb elided. Ships at info: about six per
  million words of pre-2022 Hacker News, all of them replies asking the
  same shape of a person; nothing in public-domain prose outside Joyce's
  dialogue. Several in a draft should be read as a warning.

• is-the-whole-x (rhetorical-tic, info) is thats-the-whole on any
  subject: "That periodicity is the whole tell.", "Consistency is the real
  test." It yi...

v0.5.0

The catalog grows from 49 rules to 50, and check becomes the default
command.

Changed

• A first argument that is not a command name is now taken as a path, so
  sloplint draft.md and sloplint - scan instead of failing with
  unknown command. A mistyped command fails as a missing file, still exit 2.

New rules

• trailing-significance-participle (structure, warning): the participle
  clause hung off the end of a sentence to say what a fact means, "the sign
  carries both names, showcasing the range of travellers it draws". An event
  cannot showcase anything, so the claim belongs to a narrator who never
  appears in the text.

  The verb list is closed and narrow: highlighting, showcasing,
  reinforcing, shaping, enhancing, cementing, solidifying,
  embodying, fostering, facilitating, signalling. Verbs humans write
  in the same position -- driving, representing, reflecting, marking,
  contributing, illustrating, demonstrating, emphasising, echoing,
  affirming -- stay out, because they usually take a person as the subject
  and a regex cannot see the subject. underscoring is left to
  underscores-highlights.

  Two guards drop gerund lists: a preceding -ing word means the match is a
  middle list item, and a following comma, "and" or "or" means it is not the
  last item. "signalling to" is the physical gesture and does not flag.

v0.4.0

The catalog grows from 29 rules to 49. Every new rule was probed against
pre-2022 Hacker News comments or Project Gutenberg texts before shipping,
and each narrowing is pinned by an examples_ok fixture.

New rules

• worth-saying-plainly (rhetorical-tic, warning): the self-rating
  opener, "Worth saying plainly, ...". Needs all three slots -- evaluative
  adjective, speech verb, manner adverb -- at a sentence or paragraph start.
• hold-onto-that (rhetorical-tic, warning): the sentence-initial
  reader directive "Hold onto that" / "Hold on to this".
• cleanly (rhetorical-tic, warning): the bare manner adverb, "splits
  cleanly into", "maps cleanly onto".
• clean-count (rhetorical-tic, warning): a number plus "clean" plus a
  partition noun, "two clean buckets".
• cleanest-x (rhetorical-tic, warning): the superlative used to rank
  one's own claim, "the cleanest framing is".
• clean-x (rhetorical-tic, info): "clean" in front of an idea as
  praise, "a clean abstraction".
• earns-its-place (rhetorical-tic, warning): "earns its place" and
  "earns its keep", the metaphor of a sentence or feature paying rent.
• does-a-lot-of-work (rhetorical-tic, warning): the remark that points
  at a word and rates its load, "that qualifier does a lot of work here".
• failure-mode-here (rhetorical-tic, warning): "the failure mode here
  is", the engineering term borrowed for an argument or a person. The
  deictic "here" is the whole narrowing.
• thats-the-tension (rhetorical-tic, warning): the sentence-initial
  closer "That's the tension." and "That's the bet:".
• right-up-until (rhetorical-tic, warning): "right up until it
  doesn't". The intensifier is the narrowing -- the plain "until it doesn't"
  is an old human idiom and stays clean.
• two-things-true (rhetorical-tic, warning): "two things can be true",
  the concession that names neither half.
• notice-what-there (rhetorical-tic, warning): the self-referential
  attention cue, "Notice what that argument did there", "Read that again".
• notice-what (rhetorical-tic, info): the bare sentence-initial
  "Notice what ...", the quiet half of the pair above.
• none-of-this-is-to-say (rhetorical-tic, warning): the sweeping
  concession that retracts an argument nobody made.
• if-im-being-honest (rhetorical-tic, info): the candor preamble.
  Plain "to be honest" and "I'll be honest" stay clean.
• genuinely (rhetorical-tic, info, off by default): every
  "genuinely". No allowlist separates the intensifier from the contrastive
  use, so the rule runs only when named in --select, beside
  rule-of-three.
• question-isnt (structure, info): the corrective frame in
  interrogative dress, "The question isn't whether X, it's whether Y".
  Requires the resolving clause.
• and-thats-fine (rhetorical-tic, info): the permission-granting
  closer. The match must open and close a sentence.
• less-about-more-about (structure, info): the comparative
  reframe, "It's less about X and more about Y". Both halves required.

Changed rules

• sit-with-that broadened to any object. Two branches: the deictic object
  needs no anchor, any other object needs the sentence-initial imperative.
• the-punchline-is widened to take "the honest answer is" and "the honest
  version is".
• worth-naming and worth-saying-plainly widened together and stopped
  double-flagging. Both take more verbs; worth-naming now yields the
  adverb-bearing form to worth-saying-plainly.

Docs

• docs/SPEC.md credits Wikipedia's "Signs of AI writing" and slopwash.com's
  anti-slop ruleset, next to proselint, vale and write-good.
• The SPEC rule catalog had drifted ten rules behind the code. It is resynced,
  and spec/rules_spec.rb now asserts every rule id appears in docs/SPEC.md,
  the same way it already asserts the README counts.

v0.3.0

[0.3.0] - 2026-08-02

• New rule load-bearing (rhetorical-tic, warning): "load-bearing" used
  as a borrowed metaphor for anything important, outside its literal
  construction sense (a wall, a column, a beam). Guards on both sides —
  a building noun right after it, or the predicate form ("the wall is
  load-bearing") — leave the construction sense clean.
• New rule is-is (rhetorical-tic, warning): the doubled copula, "what it
  is is a mistake" / "the thing is, is that...". No anchor needed -- the bare
  pattern scored 0 hits across ~1.9M words of public-domain and modern prose.
• exactly-the replaced by the broader exact-exactly (rhetorical-tic,
  now info): flags "exact/exactly" generally, not just the fixed "exactly
  the point/kind/problem/…" phrase shape, with an allowlist for the places
  it's doing real work (numbers, times, same/opposite/way/etc.). Severity
  drops to info because a common word will still slip through on cases
  the allowlist hasn't seen yet.
• thats-not-nothing replaced by the broader not-nothing
  (rhetorical-tic, warning): catches the "is not nothing" litotes
  regardless of subject ("Fifty basis points is not nothing"), not only
  the demonstrative-subject form ("that's not nothing"). Personal-subject
  litotes ("he was not nothing to her") and the "there is not nothing"
  philosophy frame are allowlisted as legitimate human use.
• underscores-highlights broadened (puffery, still info): "underscores
  the need for", "underscored how fragile", and "underscoring the urgency"
  now match, not only the fixed "underscores/highlights/emphasizes its/the/
  their importance/significance" phrase. The character-noun sense
  ("a leading underscore", snake_case) stays excluded.

v0.2.0

• is-real-and-not and worth-naming move from warning to info. Both
  patterns match on surface form alone and can't tell the AI cadence from an
  unrelated sentence that happens to share it: "is real, and/but/not" fires on
  any predicate-adjective sentence regardless of what follows the conjunction
  ("my debt to my senses is real and constant" -- Emerson), and "worth naming"
  collapses the AI meta-signpost with the plain sense of a thing worth
  mentioning ("the only thing worth naming to do that" -- Emerson). Probed
  against a 1.06M-word corpus (Austen, Melville, Madison, Thoreau, Emerson);
  the other 19 rules still at warning were probed too and held up clean.
• puffery-words' bare nestled matched the literal verb as often as the
  puffery sense -- a head nestled against a shoulder, a kitten nestled into a
  blanket. It now requires a following in/among/between, the shape the
  travel-brochure cliché actually takes ("nestled in the hills").
• New rule not-x-but-y (structure, info): the bare corrective "is not A
  but B" with no escalation word, comma before "but" or none. Ships at info
  because the line between a corrective and an ordinary concession is
  syntactic, and a pattern can only approximate it — a concession with an
  elided subject ("was not perfect but got us there") still gets through.
• New rule no-x-no-y-frag (rhetorical-tic, info): the "no X, no Y"
  cadence built from sentence fragments ("No fluff. No filler.") rather than
  commas. info, not warning, because two short "no" sentences in a row is
  also just writing.
• not-just-x-but-y now also catches "not because A, but because B" and the
  escalation words merely, simply, and solely alongside just/only.
  It keeps warning: the escalation word is a deliberate authorial move.
• Both not-…-but-… rules now stop at a paragraph break, so an unpunctuated
  heading or list item no longer joins up with the next paragraph's "But …".
• no-x-no-y-frag links must begin a sentence, so an ordinary sentence can no
  longer donate its tail to a chain ("There was no bread. No milk either." is
  one sentence and one fragment, not a chain), and the link separator is
  capped at two spaces so a code span blanked by --markdown cannot weld two
  distant fragments together.
• sloplint explain no longer breaks its aligned fixture block when a fixture
  contains a newline; those are escaped as \n.
• Notes carry a new context field: the match bracketed inside ~40 characters
  of surrounding prose. Human output shows it in place of the bare match, which
  told you nothing when the match was a single word or a lone em dash. excerpt
  is unchanged and still the bare match.
• New rule em-dash (structure, info): flags every em dash, not just
  paragraphs dense with them.
• em-dash-overuse (3+ em dashes in one paragraph) is now warning, up from
  info.

v0.1.0

First public release.

• 24-rule catalog across four categories: rhetorical-tic, puffery, structure,
  hedging. See sloplint rules or README.md for the full list.
• check, rules, explain, version commands; JSON and human-readable
  output; --select/--ignore by rule id or category; --markdown to skip
  fenced code, inline code, and URLs before scanning.
• Exit codes: 0 clean, 1 notes found, 2 bad arguments or invalid input
  (including invalid UTF-8 and unknown --select/--ignore ids).
• Zero runtime dependencies; requires Ruby >= 3.3.
• -v/--version print the version and exit 0. optparse auto-registers its
  own --version switch on any parser that doesn't define one, and that
  default printed "version unknown" to the real stdout and hard-exited,
  bypassing the out:/return-a-code contract every other path honors --
  caught by hand-testing the freshly built gem before this release shipped.
• Rationale text (sloplint explain) across six rules no longer narrates the
  rule's own implementation -- pattern scoping, anchoring, tuning history,
  corpus-driven thresholds. It states why the construct reads as AI writing
  and nothing else.

Rules narrowed against a corpus of real human prose (the Federalist Papers,
Moby-Dick, Walden) before release:

• not-just-x-but-y anchored on a preceding copula so it no longer flags
  ordinary correlative conjunctions ("not only... but...").
• puffery-words dropped the bare adjective "profound" and narrowed "in the
  heart of" to require a place object, both unguarded false-positive sources.
• em-dash-overuse, thats-how-x, and announced-takeaway scoped to real
  paragraphs (blank-line boundaries) instead of source lines, so hit counts
  no longer swing on Markdown line-wrapping.
• clause-triad-then removed: went 0-for-40 on its own description against
  the human corpus, and no reliable narrowing was found.
• no-x-no-y's rationale corrected to match what the pattern actually does
  (fires at two items; the frequency gap between human and model prose is
  the justification, not an item-count claim the pattern didn't enforce).