sloplint 0.7.0 → 0.8.0

data/lib/sloplint/rules.rb

@@ -6,12 +6,14 @@ module Sloplint
   # count_group: a Regexp scanned over the matched text to tally items; when set,
   #   the Note carries `count` and the message may interpolate %{count}.
   # skip: Regexps that, if they match the matched text, drop the note (exclusions).
-  # default_on: false keeps a noisy rule out of the default run (still selectable).
+  # severity: what the construct costs the prose -- "error", "warning", "info".
+  # confidence: how likely a match is a false positive -- "high", "medium", "low".
+  #   A "low" rule stays out of the default run; --strict or its own id turns it on.
   Rule = Data.define(
-    :id, :category, :severity, :pattern, :message, :suggestion,
-    :examples_bad, :examples_ok, :count_group, :skip, :rationale, :default_on
+    :id, :category, :severity, :confidence, :pattern, :message, :suggestion,
+    :examples_bad, :examples_ok, :count_group, :skip, :rationale
   ) do
-    def initialize(count_group: nil, skip: [], rationale: nil, default_on: true, **rest)
+    def initialize(count_group: nil, skip: [], rationale: nil, **rest)
       super
     end
   end
@@ -83,11 +85,11 @@ module Sloplint
     "|time|world|numbers?|roots?|money|estate|user"
 
   RULES = [
-    # ── rhetorical-tic ────────────────────────────────────────────────────
     Rule.new(
       id: "no-x-no-y",
-      category: "rhetorical-tic",
+      category: "cadence",
       severity: "warning",
+      confidence: "high",
       # Comma chains only: "no fluff, no filler, no jargon". The comma is the
       # evidence -- it makes the parallelism deliberate. Fragment chains split
       # by sentence punctuation are a separate, quieter rule (no-x-no-y-frag),
@@ -104,10 +106,11 @@ module Sloplint
     ),
     Rule.new(
       id: "no-x-no-y-frag",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "cadence",
+      severity: "warning",
+      confidence: "medium",
       # The same cadence built from sentence fragments: "No fluff. No filler."
-      # Ships at info, not warning, because the shape is genuinely ambiguous --
+      # Medium confidence, because the shape is genuinely ambiguous --
       # two short "no" sentences in a row is also just writing ("No one moved.
       # No one spoke."). The agent reading the flag decides; see rationale.
       #
@@ -147,8 +150,9 @@ module Sloplint
     ),
     Rule.new(
       id: "thats-the-whole",
-      category: "rhetorical-tic",
+      category: "closer",
       severity: "warning",
+      confidence: "high",
       # The nouns are WHOLE_CLOSERS. "value" and "fix" are the closers agents
       # write in technical prose ("That's the whole fix"), where the
       # contraction keeps them out of is-the-whole-x, which sees only "is".
@@ -198,8 +202,9 @@ module Sloplint
     ),
     Rule.new(
       id: "is-the-whole-x",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "closer",
+      severity: "warning",
+      confidence: "medium",
       # The generic form of thats-the-whole: a subject, then "is the whole /
       # real / actual / entire N" with N from a closed abstract list. "That
       # periodicity is the whole tell.", "Consistency is the real test." The
@@ -216,7 +221,7 @@ module Sloplint
       # "cost" because "the real deal", "the real thing" and "the whole cost"
       # are idioms or quantities. The noun may not run on into a compound
       # ("problem-solver"). Gaps may cross a hard-wrapped newline but never
-      # a paragraph break. Ships at info because "the real question" and
+      # a paragraph break. Medium confidence: "the real question" and
       # "the whole point" are also how people talk.
       pattern: /(?<![\w'’-])
                 (?!(?:that|this)#{WRAP_GAP}+(?:is)#{WRAP_GAP}+(?:the)#{WRAP_GAP}+(?:whole)#{WRAP_GAP}+(?:#{WHOLE_CLOSERS})(?![\w'’-]))
@@ -265,8 +270,9 @@ module Sloplint
     ),
     Rule.new(
       id: "bare-equative",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "closer",
+      severity: "warning",
+      confidence: "medium",
       # A sentence that opens on an abstract head noun and equates it with
       # something: "The tell here is the periodicity.", "The problem is not
       # the tool.", "The lesson is the handoff." The head noun list is closed
@@ -280,8 +286,8 @@ module Sloplint
       # answer is a mess"), a pointing complement ("the same", "the one",
       # "the first"), and a proper noun ("the Slack thread") are all out.
       # Gaps may cross a hard-wrapped newline but never a paragraph break.
-      # Ships at info: "The problem is the cost" is how people write too, at
-      # about four per million words on Hacker News; it is the density that
+      # Medium confidence: "The problem is the cost" is how people write too,
+      # at about four per million words on Hacker News; it is the density that
       # tells.
       pattern: /(?:^|(?<=[.!?])[ \t]{1,2})(?:[-*+•][ \t]+|\d+[.)][ \t]+)?\KThe(?:[ \t]|\r?\n(?!\s*\n))+
                 (?:tell|point|question|problem|issue|lesson|difference|trick|move|risk|goal|reason|pattern|insight|takeaway|shift|bet|catch|gap|bottleneck|failure|mistake|secret|magic|challenge|tension|trap)(?:[ \t]|\r?\n(?!\s*\n))+
@@ -322,9 +328,9 @@ module Sloplint
     ),
     Rule.new(
       id: "epistrophe",
-      category: "rhetorical-tic",
+      category: "cadence",
       severity: "info",
-      default_on: false,
+      confidence: "low",
       # Two clauses that end on the same two-word phrase, the second closing
       # the sentence: "built for one desk, and almost no job is done at one
       # desk." Two backreferences catch the repeat, one per word, so the
@@ -374,9 +380,9 @@ module Sloplint
     ),
     Rule.new(
       id: "phrase-echo",
-      category: "rhetorical-tic",
+      category: "cadence",
       severity: "info",
-      default_on: false,
+      confidence: "low",
       # The same three words again a few paragraphs on. Three consecutive
       # words, each four characters or more and lowercase-led, one of them
       # six letters with nothing but letters, and the same three again
@@ -458,8 +464,9 @@ module Sloplint
     ),
     Rule.new(
       id: "did-not-x-did-not-y",
-      category: "rhetorical-tic",
+      category: "cadence",
       severity: "warning",
+      confidence: "high",
       pattern: /\b(?:did\s+not|didn't)\s+[\w'-]+(?:,?\s+(?:and\s+)?(?:did\s+not|didn't)\s+[\w'-]+)+/i,
       message: '"did not X, did not Y" chain (%{count} items) reads as AI cadence.',
       suggestion: "Cut the chain or make it one plain sentence.",
@@ -470,8 +477,9 @@ module Sloplint
     ),
     Rule.new(
       id: "from-x-to-y-chain",
-      category: "rhetorical-tic",
+      category: "cadence",
       severity: "warning",
+      confidence: "high",
       # Two or more "from X to Y" spans in a row, comma-separated: "from
       # private notes to shared files, from personal memory to team context".
       # Each span is "from", one to three words, "to", one to three words; the
@@ -546,8 +554,9 @@ module Sloplint
     ),
     Rule.new(
       id: "one-x-one-y",
-      category: "rhetorical-tic",
+      category: "cadence",
       severity: "warning",
+      confidence: "high",
       # Three or more "one X" items in a comma chain that stands on its own:
       # "One owner, one repository, one weekly prune." or, after a colon,
       # "the setup is narrow: one reviewer, one queue, one deadline". The
@@ -617,8 +626,9 @@ module Sloplint
     ),
     Rule.new(
       id: "and-what-it-should",
-      category: "rhetorical-tic",
+      category: "closer",
       severity: "warning",
+      confidence: "high",
       # The elliptical tail: "List what the assistant knows about the client,
       # and what it should." The second "what" clause borrows its verb from the
       # first and ends on a bare modal or a negated auxiliary, so the sentence
@@ -664,8 +674,9 @@ module Sloplint
     ),
     Rule.new(
       id: "abstract-lives-in",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "puffery",
+      severity: "warning",
+      confidence: "medium",
       # 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,
@@ -681,7 +692,7 @@ module Sloplint
       # lies in the assumption") and "resides in" are ordinary English for
       # where a fault or an authority is, at any register.
       #
-      # Ships at info. The same shape states where information literally is
+      # Medium confidence. The same shape states where information literally is
       # ("the knowledge lives in our heads", "the instructions live in the
       # README"), and a sample of pre-2022 Hacker News biased toward the
       # construction turns up a few of those per million words, all human.
@@ -724,8 +735,9 @@ module Sloplint
     ),
     Rule.new(
       id: "the-x-is-the-x",
-      category: "rhetorical-tic",
-      severity: "warning",
+      category: "cadence",
+      severity: "error",
+      confidence: "high",
       # The repeated-head equative: "the reason it holds up is the reason the
       # other half happens", "the problem with A is the problem with B". The
       # same abstract head noun sits on both sides of the copula, caught by a
@@ -787,8 +799,9 @@ module Sloplint
     ),
     Rule.new(
       id: "same-determiner-chain",
-      category: "rhetorical-tic",
+      category: "cadence",
       severity: "info",
+      confidence: "medium",
       # The quiet cousin of one-x-one-y and no-x-no-y: three or more items in
       # a comma chain that all open on the same determiner or quantifier,
       # caught with a backreference: "every faculty, every thought, every
@@ -849,8 +862,9 @@ module Sloplint
     ),
     Rule.new(
       id: "dont-verb-it",
-      category: "rhetorical-tic",
+      category: "false-correction",
       severity: "warning",
+      confidence: "high",
       pattern: /\b(?:don't|do\s+not)\s+(\w+)\s+it\b[^.!?]*[.!?]\s*\1\s+it\b/i,
       message: '"Don\'t X it. X it Y." reframing is a stock LLM move.',
       suggestion: "Drop the fake reframe and state the claim once.",
@@ -860,8 +874,9 @@ module Sloplint
     ),
     Rule.new(
       id: "sit-with-that",
-      category: "rhetorical-tic",
+      category: "reader-address",
       severity: "warning",
+      confidence: "high",
       # Two branches. The deictic object is the tic anywhere in a sentence, so
       # "sit with that/this/it" needs no anchor. Anything else needs the
       # sentence-initial imperative, which is where the tic lives and where the
@@ -888,12 +903,13 @@ module Sloplint
                  "The object does not matter: the tic is telling the reader to dwell instead " \
                  "of giving them something to dwell on. The sentence-initial imperative does " \
                  "flag the plain companion sense (\"Sit with the baby while I run out\"), " \
-                 "which is accepted -- hence warning, not error."
+                 "which is accepted."
     ),
     Rule.new(
       id: "hold-onto-that",
-      category: "rhetorical-tic",
+      category: "reader-address",
       severity: "warning",
+      confidence: "high",
       # Sentence-initial imperative only. Past tense and subordinate clauses
       # ("she held on to that letter", "if you hold onto that phrase") are a
       # different construction, not the tell. Everything AFTER that/this is
@@ -919,8 +935,9 @@ module Sloplint
     ),
     Rule.new(
       id: "cleanly",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "self-rating",
+      severity: "warning",
+      confidence: "medium",
       # The bare adverb was the whole rule, and the engineering idioms were
       # flagged on purpose. Technical prose says that was the wrong call: a
       # patch that applies cleanly, a build that compiles cleanly and a gear
@@ -936,7 +953,7 @@ module Sloplint
       # parts. The clause-final form ("the objection breaks down cleanly, and
       # neither half survives") is the same tell and is left out.
       #
-      # The frame is not the sense, and this rule ships at info because of it.
+      # The frame is not the sense, and the confidence here is medium because of it.
       # "The argument splits cleanly into two parts" and "the gear retracted
       # cleanly into the well" are one pattern apart only in their subject,
       # and a regex cannot see the subject -- the same limit that keeps the
@@ -971,8 +988,9 @@ module Sloplint
     ),
     Rule.new(
       id: "clean-count",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "warning",
+      confidence: "high",
       # Needs a partition noun. The bare count reaches the laundry: Ulysses
       # has "four clean strokes", Jane Eyre "two clean tuckers".
       pattern: /\b(?:two|three|four|five|six|seven|2|3|4|5|6|7)\s+
@@ -998,8 +1016,9 @@ module Sloplint
     ),
     Rule.new(
       id: "cleanest-x",
-      category: "rhetorical-tic",
-      severity: "warning",
+      category: "self-rating",
+      severity: "error",
+      confidence: "high",
       # Noun list only. "The cleanest way to install the driver" and "the
       # cleanest cut of meat" are ordinary English, so "way" is admitted only
       # in front of a speech verb, and the concrete-capable nouns (cut, line,
@@ -1027,10 +1046,11 @@ module Sloplint
     ),
     Rule.new(
       id: "clean-x",
-      category: "rhetorical-tic",
-      severity: "info",
-      # The quiet half of cleanest-x, at info because the positive degree is
-      # where ordinary usage lives. "A clean separation of concerns" is
+      category: "self-rating",
+      severity: "warning",
+      confidence: "medium",
+      # The quiet half of cleanest-x, at medium confidence because the positive
+      # degree is where ordinary usage lives. "A clean separation of concerns" is
       # standard engineering English, so "separation" stays out of the list;
       # "break" is admitted only in "clean break between", never bare, because
       # "make a clean break with the past" is an idiom and not a tell.
@@ -1052,13 +1072,14 @@ module Sloplint
         "The report gives a clean bill of health."
       ],
       rationale: "'Clean' in front of an idea is evaluation, not description -- it says the " \
-                 "writer approves, and nothing about the idea. It ships at info because the " \
-                 "same words carry a plain sense a regex cannot separate from the tic."
+                 "writer approves, and nothing about the idea. The same words carry a plain " \
+                 "sense a regex cannot separate from the tic."
     ),
     Rule.new(
       id: "you-already-know",
-      category: "rhetorical-tic",
+      category: "reader-address",
       severity: "warning",
+      confidence: "high",
       pattern: /\byou\s+already\s+know\b/i,
       message: '"You already know…" is a stock LLM rhetorical setup.',
       suggestion: "Just make the point; don't tell the reader they know it.",
@@ -1068,8 +1089,9 @@ module Sloplint
     ),
     Rule.new(
       id: "is-the-entire",
-      category: "rhetorical-tic",
+      category: "closer",
       severity: "warning",
+      confidence: "high",
       pattern: /\bis\s+the\s+entire\s+(?:point|game|thing|business\s+model|deal|story)\b/i,
       message: '"X is the entire point/game/…" is an LLM emphasis tic.',
       suggestion: "State the point plainly without the superlative frame.",
@@ -1079,8 +1101,9 @@ module Sloplint
     ),
     Rule.new(
       id: "the-entire-is",
-      category: "rhetorical-tic",
+      category: "closer",
       severity: "warning",
+      confidence: "high",
       pattern: /\bthe\s+entire\s+(?:point|game|thing|business\s+model|deal|story)\s+is\b/i,
       message: '"The entire point/game/… is" is an LLM emphasis tic.',
       suggestion: "State the point plainly without the superlative frame.",
@@ -1090,8 +1113,9 @@ module Sloplint
     ),
     Rule.new(
       id: "is-real-and-not",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "false-concession",
+      severity: "warning",
+      confidence: "medium",
       pattern: /\bis\s+real,?\s+(?:and|but|not)\b/i,
       message: '"The X is real, and…" is a stock LLM concession move.',
       suggestion: "Drop the 'is real, and' scaffolding; assert the point directly.",
@@ -1112,8 +1136,9 @@ module Sloplint
     ),
     Rule.new(
       id: "real-x-real-y",
-      category: "rhetorical-tic",
+      category: "cadence",
       severity: "info",
+      confidence: "medium",
       # Ships at info, not warning: the only two hits the probe found in
       # 1.28M words were both false positives, and one corpus of one
       # register is thin evidence next to honestly, which sits at warning
@@ -1210,8 +1235,9 @@ module Sloplint
     ),
     Rule.new(
       id: "the-punchline-is",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "warning",
+      confidence: "high",
       # Same reveal, three nouns. "honest answer" and "honest version" join
       # "punchline" because they do the identical job: rate the sentence as
       # the candid one before the reader gets it. "short version" was tried
@@ -1236,8 +1262,9 @@ module Sloplint
     ),
     Rule.new(
       id: "worth-naming",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "self-rating",
+      severity: "warning",
+      confidence: "medium",
       # Widened to optionally include a trailing "names" so the "naming
       # names" idiom is part of the matched text -- skip: checks the matched
       # text itself, and the tighter /\bworth\s+naming\b/ never captured
@@ -1274,8 +1301,9 @@ module Sloplint
     ),
     Rule.new(
       id: "worth-saying-plainly",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "warning",
+      confidence: "high",
       # Two branches, both sentence-initial.
       #
       # First: evaluative adjective + speech verb + manner adverb, all three
@@ -1327,8 +1355,9 @@ module Sloplint
     ),
     Rule.new(
       id: "not-nothing",
-      category: "rhetorical-tic",
+      category: "false-concession",
       severity: "warning",
+      confidence: "high",
       # Two branches: spelled-out copula with an optional captured subject, and
       # the contracted "X's not nothing". The optional subject + skip: is the
       # load-bearing idiom -- skip: only sees matched text, so the personal
@@ -1382,8 +1411,9 @@ module Sloplint
 
     Rule.new(
       id: "exact-exactly",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "info",
+      confidence: "medium",
       # Rewritten after 2.2M words of technical prose left ~111 of 125 hits
       # false. The shape was the problem, not the entries: the pattern matched
       # "exact" everywhere and subtracted an allow-list, and an open pattern
@@ -1489,8 +1519,9 @@ module Sloplint
     ),
     Rule.new(
       id: "load-bearing",
-      category: "rhetorical-tic",
+      category: "borrowed-metaphor",
       severity: "warning",
+      confidence: "high",
       # Two guards, both structural, over the same noun list so they can't
       # drift apart. Forward: a physical building part right after it is the
       # literal sense, checked with a negative lookahead. Backward: the
@@ -1539,8 +1570,9 @@ module Sloplint
     ),
     Rule.new(
       id: "intersection-of",
-      category: "rhetorical-tic",
+      category: "borrowed-metaphor",
       severity: "warning",
+      confidence: "high",
       # "at" is not load-bearing -- "explores the intersection of art and
       # technology" is the same move -- so the anchor is "the intersection of"
       # and the two guards carry the whole burden of separating the literal
@@ -1626,8 +1658,9 @@ module Sloplint
     ),
     Rule.new(
       id: "impact-verb",
-      category: "rhetorical-tic",
+      category: "borrowed-metaphor",
       severity: "warning",
+      confidence: "high",
       # "impact" and "impacts" are also nouns, so they only count as verbs
       # behind an auxiliary or a subject pronoun. One optional pronoun may sit
       # between the auxiliary and the verb ("does this impact the date");
@@ -1731,10 +1764,11 @@ module Sloplint
     Rule.new(
       id: "impact-noun-vague",
       category: "puffery",
-      severity: "warning",
+      severity: "error",
+      confidence: "high",
       # Only the puffed shapes: an intensity adjective, or make/have plus an
-      # article. "the impact of X" is left to impact-noun-bare, which sits at
-      # info because it is the standard word in research prose. "positive" and
+      # article. "the impact of X" is left to impact-noun-bare, which takes the
+      # bare form because it is the standard word in research prose. "positive" and
       # "negative" stay out of the adjective list on purpose -- they name a
       # direction, which is more than the intensity words do. "statistically"
       # is pulled into the match as an optional leading word so skip: can see
@@ -1782,8 +1816,9 @@ module Sloplint
     ),
     Rule.new(
       id: "impact-noun-bare",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "borrowed-metaphor",
+      severity: "warning",
+      confidence: "medium",
       # Two shapes, each needing its own anchor: a measuring verb in front, or
       # "of" behind. That is what keeps the collision sense clear without a
       # list of collision verbs -- "the impact crushed the front bumper" has
@@ -1820,14 +1855,15 @@ module Sloplint
         "Torque the bolts with an impact wrench.",
         "Will the impact be permanent?"
       ],
-      rationale: "This one sits at info because 'the impact of X on Y' is the ordinary word in " \
-                 "research and policy writing, not a tell. Elsewhere it postpones the sentence: " \
-                 "the writer announces that an effect exists and stops before naming it."
+      rationale: "'The impact of X on Y' is the ordinary word in research and policy writing, " \
+                 "not a tell. Elsewhere it postpones the sentence: the writer announces that " \
+                 "an effect exists and stops before naming it."
     ),
     Rule.new(
       id: "thats-how-x",
-      category: "rhetorical-tic",
+      category: "closer",
       severity: "warning",
+      confidence: "high",
       pattern: /(?:\A|[.!?]\s+|\n\s*\n)\s*(?:that|this)(?:'s| is)\s+how\b/i,
       message: '"That\'s how…" opening a sentence is a stock LLM aphorism closer.',
       suggestion: "Cut the closer, or replace it with the concrete result you mean.",
@@ -1843,8 +1879,9 @@ module Sloplint
     ),
     Rule.new(
       id: "announced-takeaway",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "warning",
+      confidence: "high",
       pattern: /(?:\A|[.!?]\s+|\n\s*\n)\s*(?:here'?s\s+)?the\s+(?:loop|pattern|trick|lesson|takeaway|playbook|framing|insight|kicker)\b[^.!?\n]{0,60}:/i,
       message: "Colon-led takeaway label announces the lesson before making it.",
       suggestion: "Give the observation first; let the reader decide it's the takeaway.",
@@ -1860,8 +1897,9 @@ module Sloplint
     ),
     Rule.new(
       id: "is-is",
-      category: "rhetorical-tic",
+      category: "cadence",
       severity: "warning",
+      confidence: "high",
       # No anchor needed -- the doubled copula alone scored 0 across ~1.9M words.
       # The comma is allowed because "What it is, is a mystery" grates the same
       # way. Sentence and clause punctuation still block the weld ("what it is.
@@ -1896,8 +1934,9 @@ module Sloplint
 
     Rule.new(
       id: "earns-its-place",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "warning",
+      confidence: "high",
       # The possessive is the narrowing. "earned a place on the team" and
       # "earn a place in the final" are ordinary; "earns its place" is the
       # metaphor, a thing paying for the room it takes up.
@@ -1921,8 +1960,9 @@ module Sloplint
     ),
     Rule.new(
       id: "does-a-lot-of-work",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "warning",
+      confidence: "high",
       # Two arms, both narrowed away from the ordinary sense.
       #
       # "a lot of work" needs a locative ("here", "in that sentence") because
@@ -1954,8 +1994,9 @@ module Sloplint
     ),
     Rule.new(
       id: "failure-mode-here",
-      category: "rhetorical-tic",
+      category: "borrowed-metaphor",
       severity: "warning",
+      confidence: "high",
       # "here" is the whole narrowing, and it is doing a lot -- the bare "the
       # failure mode is" is ordinary engineering writing about real systems,
       # where the phrase means what it says. The deictic is what marks the
@@ -1981,8 +2022,9 @@ module Sloplint
     ),
     Rule.new(
       id: "thats-the-tension",
-      category: "rhetorical-tic",
+      category: "closer",
       severity: "warning",
+      confidence: "high",
       # Sentence-initial, only two nouns, and the noun must end the clause.
       #
       # "tradeoff" and "catch" were tried and cut: both are ordinary English
@@ -2015,8 +2057,9 @@ module Sloplint
     ),
     Rule.new(
       id: "right-up-until",
-      category: "rhetorical-tic",
+      category: "closer",
       severity: "warning",
+      confidence: "high",
       # The intensifier is the tell, not the reversal. "It works until it
       # doesn't" is an old human idiom and stays clean; stacking "right up"
       # in front of it is the model's version, and the two together are two
@@ -2041,8 +2084,9 @@ module Sloplint
     ),
     Rule.new(
       id: "two-things-true",
-      category: "rhetorical-tic",
-      severity: "warning",
+      category: "false-concession",
+      severity: "error",
+      confidence: "high",
       # Closed phrase, no anchor needed. The optional "both" and the optional
       # "at once" tail are the two ways the sentence is padded; the count word
       # is fixed at two, because "three things can be true" is someone
@@ -2067,8 +2111,9 @@ module Sloplint
     ),
     Rule.new(
       id: "notice-what-there",
-      category: "rhetorical-tic",
+      category: "reader-address",
       severity: "warning",
+      confidence: "high",
       # The self-referential half of the attention cue: the sentence points
       # at the writing rather than at anything in the world. Two frames, both
       # sentence-initial -- "notice what X did there" and the bare "read that
@@ -2096,8 +2141,9 @@ module Sloplint
     ),
     Rule.new(
       id: "notice-what",
-      category: "rhetorical-tic",
+      category: "reader-address",
       severity: "info",
+      confidence: "medium",
       # The ambiguous half of the pair, and it ships at info because the
       # sentence-initial imperative is also how people point at something
       # real: "Notice what happens around Q3", "Notice what is not on the
@@ -2128,8 +2174,9 @@ module Sloplint
     ),
     Rule.new(
       id: "none-of-this-is-to-say",
-      category: "rhetorical-tic",
-      severity: "warning",
+      category: "false-concession",
+      severity: "error",
+      confidence: "high",
       # Only the "none of" form. Every neighbouring phrasing is ordinary
       # English by an order of magnitude -- "which is not to say", "this is
       # not to say", "that's not to say" -- and admitting any of them would
@@ -2155,10 +2202,11 @@ module Sloplint
     ),
     Rule.new(
       id: "if-im-being-honest",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "false-concession",
+      severity: "warning",
+      confidence: "medium",
       # The "being honest" frame only. Plain "to be honest" and "I'll be
-      # honest" are how people talk and are excluded. info, not warning:
+      # honest" are how people talk and are excluded. Medium confidence:
       # writers really do say this out loud.
       pattern: /\bif\s+(?:I['’]?m|I\s+am|we['’]?re|we\s+are)\s+(?:being\s+)?honest\b
                 |\bhonestly,\s+the\s+(?:answer|truth)\b/ix,
@@ -2180,8 +2228,9 @@ module Sloplint
     ),
     Rule.new(
       id: "honestly",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "warning",
+      confidence: "high",
       # The manner adverb, the way "cleanly" is the manner adverb: hung on a
       # subject that cannot be honest. Two guards, no verb list.
       #
@@ -2231,8 +2280,9 @@ module Sloplint
     ),
     Rule.new(
       id: "honest-x",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "warning",
+      confidence: "high",
       # The noun list is short on purpose, and the words left out are the
       # point. "An honest answer", "an honest assessment" and "an honest
       # account" are ordinary English about people -- six hits between them in
@@ -2274,8 +2324,9 @@ module Sloplint
     ),
     Rule.new(
       id: "most-honest-x",
-      category: "rhetorical-tic",
-      severity: "warning",
+      category: "self-rating",
+      severity: "error",
+      confidence: "high",
       # The superlative frame carries the tell on its own, so this noun list is
       # wider than honest-x's -- "the most honest assessment" is self-ranking
       # in a way "an honest assessment" is not. The list still has to keep out
@@ -2310,14 +2361,14 @@ module Sloplint
     ),
     Rule.new(
       id: "genuinely",
-      category: "rhetorical-tic",
+      category: "self-rating",
       severity: "info",
+      confidence: "low",
       # Off by default. There is no narrowing here: the word is ordinary
       # English at every frequency we measured, and the difference between
       # the tell and the real use is whether a contrast exists in the
       # surrounding argument, which no regex can see. Selectable when a
       # writer wants every occurrence listed back to them.
-      default_on: false,
       pattern: /\bgenuinely\b/i,
       message: '"Genuinely" asserts sincerity instead of earning it (heuristic; high false-positive).',
       suggestion: "Cut the adverb. If the sentence needs it, the claim is doing the work.",
@@ -2339,8 +2390,9 @@ module Sloplint
     ),
     Rule.new(
       id: "actually-not-x",
-      category: "rhetorical-tic",
+      category: "false-correction",
       severity: "warning",
+      confidence: "high",
       # Two markers that each correct the reader, doubled up in one clause:
       # the adverb and the trailing "…, not X". Bare "actually" is not the
       # tell and is not matched, because "the build actually failed on the
@@ -2405,8 +2457,9 @@ module Sloplint
     ),
     Rule.new(
       id: "and-thats-fine",
-      category: "rhetorical-tic",
-      severity: "info",
+      category: "false-concession",
+      severity: "warning",
+      confidence: "medium",
       # Three narrowings, and the rule needs all of them. "and" is required:
       # bare "that's fine" is a reply people write constantly. The match must
       # open a sentence and close it, so the concessive clause -- "and that's
@@ -2430,8 +2483,9 @@ module Sloplint
     ),
     Rule.new(
       id: "and-nothing-else",
-      category: "rhetorical-tic",
+      category: "closer",
       severity: "warning",
+      confidence: "high",
       # This rule is deliberately wide, and the cost is known. There is no verb
       # list and no imperative requirement, so the only narrowing is structural
       # -- which means the pattern cannot tell a leaked instruction from the
@@ -2489,8 +2543,9 @@ module Sloplint
     ),
     Rule.new(
       id: "nothing-else-frag",
-      category: "rhetorical-tic",
+      category: "closer",
       severity: "warning",
+      confidence: "high",
       # The same exclusion as its own sentence: "Return the JSON. Nothing
       # else." Built on the no-x-no-y-frag template -- the fragment must start
       # at a sentence boundary and the separator is at most two spaces or one
@@ -2526,18 +2581,18 @@ module Sloplint
         # letter guard is there to exclude.
         "and so on; nothing more."
       ],
-      rationale: "The fragment form of the same leaked instruction, and it reads the same way: " \
-                 "a second sentence that only says the first one was complete. It ships at " \
-                 "warning rather than info because the capital letter and the whole-sentence " \
-                 "requirement keep it off the continuation sense, which is where ordinary prose " \
-                 "puts the phrase: one hit in 1.92M words of public-domain prose and one in " \
-                 "461k words of pre-2022 Hacker News."
+      rationale: "The fragment form of the same leaked instruction, and it reads the same " \
+                 "way: a second sentence that only says the first one was complete. The " \
+                 "capital letter and the whole-sentence requirement keep it off the " \
+                 "continuation sense, which is where ordinary prose puts the phrase: one hit " \
+                 "in 1.92M words of public-domain prose and one in 461k words of pre-2022 " \
+                 "Hacker News."
     ),
-    # ── puffery ───────────────────────────────────────────────────────────
     Rule.new(
       id: "puffery-words",
       category: "puffery",
-      severity: "warning",
+      severity: "error",
+      confidence: "high",
       # "nestled" alone is the literal verb as often as the puffery sense --
       # a head nestling against a shoulder, a kitten nestling into a blanket
       # -- so it requires a following in/among/between, the same shape the
@@ -2573,7 +2628,8 @@ module Sloplint
     Rule.new(
       id: "stands-serves-as",
       category: "puffery",
-      severity: "info",
+      severity: "warning",
+      confidence: "medium",
       pattern: /\b(?:stands|serves)\s+as\b|\bis\s+a\s+(?:testament|reminder)\s+to\b/i,
       message: '"stands/serves as", "is a testament/reminder to" is puffed AI framing.',
       suggestion: "Say what it does, not what it 'stands as'.",
@@ -2584,7 +2640,8 @@ module Sloplint
     Rule.new(
       id: "vital-role",
       category: "puffery",
-      severity: "warning",
+      severity: "error",
+      confidence: "high",
       pattern: /\bplays?\s+a\s+(?:vital|crucial|pivotal|significant|key|central)\s+role\b/i,
       message: '"plays a vital/crucial/… role" is a stock AI puffery phrase.',
       suggestion: "State the specific role or effect instead.",
@@ -2595,7 +2652,8 @@ module Sloplint
     Rule.new(
       id: "underscores-highlights",
       category: "puffery",
-      severity: "info",
+      severity: "warning",
+      confidence: "medium",
       # "underscored/underscoring" are unambiguously the verb and flag bare.
       # "underscore/underscores" is also the character noun (a leading
       # underscore, snake_case docs), so those forms require a following
@@ -2636,15 +2694,16 @@ module Sloplint
         "Numbers accept underscores\n\nThe next section covers floats."
       ],
       rationale: "Models reach for 'underscore' as an all-purpose emphasis verb -- findings " \
-                 "underscore, outages underscore -- asserting significance without earning it. " \
-                 "Sincere journalistic and academic use exists, hence info: a flag means the " \
-                 "move is present, not that it's slop. The character noun never takes the " \
-                 "verb's frame and stays out."
+                 "underscore, outages underscore -- asserting significance without earning " \
+                 "it. Sincere journalistic and academic use exists: a flag means the move is " \
+                 "present, not that it's slop. The character noun never takes the verb's " \
+                 "frame and stays out."
     ),
     Rule.new(
       id: "rich-tapestry",
       category: "puffery",
-      severity: "warning",
+      severity: "error",
+      confidence: "high",
       pattern: /\brich\s+tapestry\b|\btapestry\s+of\b/i,
       message: '"rich tapestry"/"tapestry of" is a signature AI cliché.',
       suggestion: "Cut the metaphor; name the actual things.",
@@ -2653,11 +2712,11 @@ module Sloplint
       rationale: "'tapestry of' is one of the most reliable single-phrase model tells."
     ),
 
-    # ── structure ─────────────────────────────────────────────────────────
     Rule.new(
       id: "not-just-x-but-y",
-      category: "structure",
+      category: "false-correction",
       severity: "warning",
+      confidence: "high",
       # Two branches, both anchored on an explicit escalation word. (1) The
       # copula escalation: "is not just/only/merely/simply/solely A … but B".
       # (2) "not because A, but because B". The escalation word is what makes
@@ -2701,10 +2760,11 @@ module Sloplint
     ),
     Rule.new(
       id: "not-x-but-y",
-      category: "structure",
-      severity: "info",
+      category: "false-correction",
+      severity: "warning",
+      confidence: "medium",
       # The bare corrective: "is not A but B", no escalation word, comma or no
-      # comma. Ships at info because the line between a corrective ("not an
+      # comma. Medium confidence, because the line between a corrective ("not an
       # accident but a strategy") and an ordinary concession ("not warm but the
       # fire helped") is syntactic, and a regex cannot see syntax. What is here
       # is a set of cheap narrowings that cut the worst of the noise: A is
@@ -2718,7 +2778,8 @@ module Sloplint
       # ("was not warm but the fire helped") or a bare lexical verb ("was not
       # perfect but got us there"), because neither is distinguishable from the
       # corrective by surface form. That is the cost of the rule and the reason
-      # it is info: the agent reading the flag has the context to judge, and
+      # its confidence is medium: the agent reading the flag has the context
+      # to judge, and
       # should. Do not chase these by growing the B-list -- every word added
       # silently narrows recall with nothing pinning it.
       pattern: /(?:\bis|\bare|\bwas|\bwere|\bisn['’]t|\baren['’]t|\bwasn['’]t|\bweren['’]t|
@@ -2765,15 +2826,16 @@ module Sloplint
                  "escalation word dropped, and models reach for it constantly. It is the " \
                  "noisiest rule in the catalog by design: a loose version scored 174 hits " \
                  "before narrowing, cut to 15 here -- but not all 15 are correctives. Some " \
-                 "are concessions with an elided subject (Walden's " \
-                 "'It was not lonely, but made all the earth lonely beneath it'), which no " \
-                 "surface pattern can tell apart from the real thing. Hence info: a flag here " \
-                 "means 'this has the shape', not 'this is slop'."
+                 "are concessions with an elided subject (Walden's 'It was not lonely, but " \
+                 "made all the earth lonely beneath it'), which no surface pattern can tell " \
+                 "apart from the real thing. A flag here means 'this has the shape', not " \
+                 "'this is slop'."
     ),
     Rule.new(
       id: "isnt-x-its-y",
-      category: "structure",
-      severity: "info",
+      category: "false-correction",
+      severity: "warning",
+      confidence: "medium",
       # The corrective with the conjunction dropped: one clause rejects a
       # description, the next supplies the replacement through a second copula
       # ("It isn't the tool. It's the habit."). Every copula tense is covered,
@@ -2789,7 +2851,8 @@ module Sloplint
       # guards cut a loose version from 19 hits to 4 over 889k words of
       # 19th-century fiction.
       #
-      # What survives is the reason for info. An adjective outside the list
+      # What survives is the reason the confidence is medium. An adjective
+      # outside the list
       # still slips through, and some hits are the same shape written by a
       # person -- "was not the wife; it was the children" is Conan Doyle. A flag
       # here says the sentence has the frame, not that a model wrote it.
@@ -2861,13 +2924,14 @@ module Sloplint
                  "one, so the sentence sounds like a correction while correcting nobody. It " \
                  "is 'not A but B' with the conjunction dropped and the second half promoted " \
                  "to its own clause, which is the form models reach for most. Ordinary prose " \
-                 "contrasts two things this way too, so the rule ships at info: it reports " \
-                 "the shape, not a verdict."
+                 "contrasts two things this way too, so the rule reports the shape, not a " \
+                 "verdict."
     ),
     Rule.new(
       id: "not-by-x-but-by-y",
-      category: "structure",
-      severity: "info",
+      category: "false-correction",
+      severity: "warning",
+      confidence: "medium",
       # The corrective built on a repeated preposition: "not by A, but by B",
       # "not from A but from B". The copula rules above cannot see it because
       # nothing precedes "not" but the verb or a dash, so the anchor here is
@@ -2908,16 +2972,16 @@ module Sloplint
         "We won not by luck\n\nBut by then it hardly mattered."
       ],
       rationale: "The 'not A but B' corrective with the copula swapped for a repeated " \
-                 "preposition, which is how a model corrects a claim about means or " \
-                 "cause ('not by luck, but by design'). People write it too -- Thoreau " \
-                 "and Melville both lean on it -- so the rule ships at info. It reports " \
-                 "the shape; a human reader decides whether it earned its place."
+                 "preposition, which is how a model corrects a claim about means or cause " \
+                 "('not by luck, but by design'). People write it too -- Thoreau and Melville " \
+                 "both lean on it -- and the rule reports the shape; a human reader decides " \
+                 "whether it earned its place."
     ),
     Rule.new(
       id: "rule-of-three",
-      category: "structure",
+      category: "cadence",
       severity: "info",
-      default_on: false,
+      confidence: "low",
       pattern: /\b[\w'-]+,\s+[\w'-]+,\s+(?:and\s+)?[\w'-]+[.!?]/,
       message: "Three single words in a comma series closing a sentence (heuristic; high false-positive).",
       suggestion: "Fine in moderation; watch for the AI habit of ending on triplets.",
@@ -2943,8 +3007,9 @@ module Sloplint
     # CLAUDE.md: some tells can't be regexes; this was one.
     Rule.new(
       id: "everyone-nobody",
-      category: "structure",
+      category: "cadence",
       severity: "warning",
+      confidence: "high",
       # The comma-spliced antithesis on quantifier subjects: "Everyone wants
       # the dashboard, nobody maintains it." One clause opens on
       # everyone/everybody, the other on nobody/no one/none or "one N"
@@ -3016,8 +3081,9 @@ module Sloplint
     ),
     Rule.new(
       id: "np-fragment-and",
-      category: "structure",
-      severity: "info",
+      category: "cadence",
+      severity: "warning",
+      confidence: "medium",
       # A whole sentence that is 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 for the reader to supply.
@@ -3028,7 +3094,7 @@ module Sloplint
       # anywhere in it, contractions included, so "A man and a woman were
       # there." and "A man and a woman aren't here." never match.
       #
-      # Ships at info, and this is why: a lexical verb is invisible to the
+      # Medium confidence, and this is why: a lexical verb is invisible to the
       # pattern, so "A car and a truck collided." has the same shape and
       # flags. The corpora say that sentence is rare (one hit in 1.25M words
       # of public-domain prose, most of it narrative), but it is a complete
@@ -3071,8 +3137,9 @@ module Sloplint
     ),
     Rule.new(
       id: "quip-question",
-      category: "structure",
+      category: "reader-address",
       severity: "info",
+      confidence: "medium",
       # The verbless question that opens a pitch: "No invite?", "New to the
       # tool?", "Still stuck?", "Ready to start?". It must start a sentence,
       # open on one of a short list of words, run one to four more words, and
@@ -3118,8 +3185,9 @@ module Sloplint
     ),
     Rule.new(
       id: "mic-drop-closer",
-      category: "structure",
+      category: "cadence",
       severity: "info",
+      confidence: "medium",
       # The kicker: a sentence of at least sixty characters, then a closer of
       # two to eight words that ends the paragraph and opens on a quantifier
       # ("Nothing here needs a new login.", "Most teams end up
@@ -3142,7 +3210,7 @@ module Sloplint
       # position." The quantifiers carry the tell; the demonstratives are
       # ordinary, and no examples_bad used one.
       #
-      # Ships at info, and the rationale says why: people end paragraphs this
+      # Medium confidence, and the rationale says why: people end paragraphs this
       # way too, at about 150 per million words on Hacker News. One is
       # nothing. A draft where most paragraphs end this way is the tell, and
       # an agent that sees the flag repeat should read the family as a
@@ -3199,9 +3267,10 @@ module Sloplint
                  "warning, and the fix is usually to delete the closer outright."
     ),
     Rule.new(
-      id: "ellipsis-closer",
-      category: "structure",
+      id: "bare-auxiliary-closer",
+      category: "cadence",
       severity: "info",
+      confidence: "medium",
       # The same long-sentence-then-short-closer shape as mic-drop-closer,
       # but the tell lives in the verb, not the subject. mic-drop-closer's
       # closer opens on a quantifier and keeps a full verb with its object
@@ -3278,14 +3347,15 @@ module Sloplint
       ],
       rationale: "The closer withholds exactly the thing the long setup built toward -- the " \
                  "verb and its object are gone, and only the bare confirmation that something " \
-                 "happened is left standing. Ships at info: plenty of ordinary writing drops " \
-                 "the verb the same way when it confirms an expectation, and one flag proves " \
-                 "nothing; a draft where it repeats is the tell."
+                 "happened is left standing. Plenty of ordinary writing drops the verb the " \
+                 "same way when it confirms an expectation, and one flag proves nothing; a " \
+                 "draft where it repeats is the tell."
     ),
     Rule.new(
       id: "short-run",
-      category: "structure",
+      category: "cadence",
       severity: "info",
+      confidence: "medium",
       # Three consecutive sentences of thirty characters or fewer, each
       # opening on a letter and closing on a full stop, with no quotation
       # mark or digit in any of them: "Nobody used it. A named owner. Then a
@@ -3422,8 +3492,9 @@ module Sloplint
     ),
     Rule.new(
       id: "em-dash",
-      category: "structure",
+      category: "punctuation",
       severity: "info",
+      confidence: "medium",
       pattern: /—/,
       message: "Em dash — an AI punctuation tell.",
       suggestion: "Recast with a comma, parentheses, or a separate sentence.",
@@ -3442,8 +3513,9 @@ module Sloplint
     ),
     Rule.new(
       id: "em-dash-overuse",
-      category: "structure",
+      category: "punctuation",
       severity: "warning",
+      confidence: "high",
       pattern: /—(?:[^\n]|\n(?!\s*\n))*—(?:[^\n]|\n(?!\s*\n))*—/,
       message: "Three or more em dashes in one paragraph — an AI punctuation tell.",
       suggestion: "Recast with commas, parentheses, or separate sentences.",
@@ -3465,8 +3537,9 @@ module Sloplint
 
     Rule.new(
       id: "question-isnt",
-      category: "structure",
-      severity: "info",
+      category: "false-correction",
+      severity: "warning",
+      confidence: "medium",
       # The resolving clause is required, so a plain rhetorical question never
       # matches. "The real question is" is ordinary English and is excluded by
       # the adjacency of "question" to the negated copula.
@@ -3492,8 +3565,9 @@ module Sloplint
     ),
     Rule.new(
       id: "less-about-more-about",
-      category: "structure",
-      severity: "info",
+      category: "false-correction",
+      severity: "warning",
+      confidence: "medium",
       # The full frame is required at both ends. Bare "less about" and bare
       # "more about" are ordinary English on their own, and the subject slot
       # is limited to the pronouns so that a sentence with a real subject --
@@ -3519,8 +3593,9 @@ module Sloplint
     ),
     Rule.new(
       id: "trailing-significance-participle",
-      category: "structure",
-      severity: "warning",
+      category: "puffery",
+      severity: "error",
+      confidence: "high",
       # The verb list is closed and short on purpose. Wikipedia's "signs of AI
       # writing" names eight watch words for this construction; half of them
       # did not survive probing. In 1.85M words of pre-2022 Hacker News and
@@ -3597,9 +3672,9 @@ module Sloplint
     ),
     Rule.new(
       id: "trailing-restatement",
-      category: "structure",
-      severity: "info",
-      default_on: false,
+      category: "closer",
+      severity: "warning",
+      confidence: "low",
       # The restating tail: "…, which means working through the process
       # rather than around it". A regex sees the connective and not whether
       # the tail says the head again, and the sentence that prompted the rule
@@ -3687,11 +3762,11 @@ module Sloplint
                  "use the same connective to state a consequence, and the pattern cannot " \
                  "tell the two apart, so the rule is off by default."
     ),
-    # ── hedging ───────────────────────────────────────────────────────────
     Rule.new(
       id: "vague-attribution",
-      category: "hedging",
-      severity: "warning",
+      category: "false-concession",
+      severity: "error",
+      confidence: "high",
       pattern: /\bsome\s+(?:critics|experts|observers|scholars|analysts)\s+(?:argue|say|believe|contend|maintain)\b|\bit\s+is\s+widely\s+(?:regarded|considered|seen|believed|acknowledged)\b|\bmany\s+would\s+argue\b/i,
       message: "Vague attribution ('some critics argue', 'it is widely…') — an AI hedging tell.",
       suggestion: "Name the source, or drop the appeal to unnamed authority.",