zero-slop 2.5.8 → 2.7.0

package/data/learned-log.md

@@ -337,3 +337,56 @@ example that motivated it.
   penalty. Six community tells added, `--explain` and `--json --gate` repaired,
   `--dna` added, a ReDoS and two quadratics removed, and user prose stopped
   reaching git-tracked files.
+- 2026-08-26 — v2.5.9 incumbent audit. Added three precision-scoped shared
+  patterns from `conorbronsdon/avoid-ai-writing@40328bd`: noun-anchored
+  lingering-attention framing, generic social endorsement closers, and
+  asterisked chat-roleplay actions. Literal return-to-an-idea clauses with a
+  reason, ordinary instructional verbs, and ordinary italics stay silent.
+  Expanded the existing AI-tool residue rule to cover tracker parameters from
+  Claude, Copilot, Gemini, Perplexity, and Grok. Added mixed-script and
+  zero-width normalization, exact protection for
+  code/front matter/blockquotes/tables/inline identifiers/paths/heading levels,
+  and a weak long-form low-word-variety check. The latter was the only new
+  stylometric feature to clear the incumbent's 1,654-paragraph corpus with
+  useful lift (20/779 machine paragraphs, 1/875 human; 22.46x). Function-word
+  entropy, punctuation uniformity, and cross-paragraph rhythm were tested but
+  did not fire on the 38-item editorial panel; they remain unshipped. Frozen
+  editorial metrics were unchanged, 12/12 known-human files stayed below the
+  gate, and the adversarial fixtures moved from missed to caught.
+- 2026-08-26 — v2.5.10 unmarked-antithesis gap. A 209-word draft with four
+  antithesis pairs, one per paragraph, scored 13.0/100 "clear" with zero flagged
+  phrases. Root cause: all 17 `contrast` patterns were anchored on a literal
+  negation token (`not`, `n't`, `never`, `won't`), so the same figure with no
+  marker was unreachable, and the pairs' short-then-long cadence *raised*
+  burstiness from 0.332 to 0.527 — the tell paid off the channel built to catch
+  machine cadence, the failure mode already documented for broetry. Added three
+  patterns and widened one: `isocolon-ditransitive` (a give-you frame whose verb
+  repeats across a sentence break), `this-is-what-looks-like`,
+  `no-x-had-to`, and eight nouns on `performed-candor` so "Here's the detail
+  that matters" is caught. The draft moved 13.0 → 59.0.
+  `isocolon-ditransitive` is the first pattern to reach a figure rather than a
+  phrase; its safety property is that the backreference sits on the **verb**,
+  which anaphora never repeats ("dedicate / consecrate / hallow"). Three
+  frame-level loosenings were measured firing on `gettysburg.txt`,
+  `federalist.txt`, and `esl-engineer-email.txt` and were rejected;
+  `Detector.test_isocolon_rule_turns_on_verb_identity` pins that. All 114 frozen
+  document scores were unchanged, 18/18 known-human controls stayed below the
+  gate, and blind-panel accuracy held at 0.8421. These are the file's first
+  patterns to carry `hints`, which left throughput higher than the baseline.
+  `judgment/verdict-arithmetic.txt` graduated to `mechanical/` — the first span
+  to move from "no regex gates this safely" to caught. Bare subject swap
+  ("Llama is open-weights. Dolma releases the data.") stays unreachable and was
+  added as `judgment/bare-subject-swap.txt`; SKILL.md now makes the
+  performed-register pass mandatory and gives it a reported count, because a
+  clean scorer had been licensing the assistant to skip it entirely.
+
+- 2026-08-26 — incumbent completeness follow-up. Added four narrowly bounded
+  phrase families found in `conorbronsdon/avoid-ai-writing@40328bd`: leaked
+  reasoning narration, unsupported novelty formulas, generic emotional
+  reactions, and answer-restatement loops. Each rule names the full formula;
+  ordinary step-by-step instructions, literal staffing statements, explained
+  surprise, and direct answers remain silent. The accompanying AI review now
+  checks six context-dependent problems that are unsafe to infer from a phrase
+  alone: paragraph-order dependence, unsupported novelty, self-labelled
+  significance, moral-adjective category errors, recap-flattery, and
+  wall-of-text replies.

package/data/learned.json

@@ -120,6 +120,76 @@
    "first_seen": "2026-08-04",
    "last_confirmed": "2026-08-04",
    "source": "community-taxonomy"
+  },
+  {
+   "name": "lingering-attention",
+   "cat": "performed",
+   "rx": "\\b(?:the|that|this)\\s+(?:one\\s+)?(?:line|quote|bit|part|idea|point|framing|comment|thing|phrase)\\s+(?:that\\s+)?i\\s+keep\\s+(?:coming\\s+back\\s+to|thinking\\s+about)\\b|\\bi\\s+can(?:'|’)?t\\s+stop\\s+thinking\\s+about\\b|\\b(?:has|have|had|been|be)\\s+(?:been\\s+)?rattling\\s+around\\s+(?:in\\s+)?my\\s+(?:head|brain)\\b|\\bi(?:'|’)?ve\\s+been\\s+chewing\\s+on\\s+(?:this|that)\\b",
+   "hints": ["i keep", "i can't", "i can’t", "rattling", "been chewing"],
+   "w": 3.5,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26",
+   "source": "conorbronsdon/avoid-ai-writing@40328bd"
+  },
+  {
+   "name": "social-endorsement-closer",
+   "cat": "linkedin",
+   "rx": "\\bthis\\s+one(?:'|’)?s?\\s+(?:is\\s+)?(?:well\\s+|really\\s+|definitely\\s+)?worth\\s+(?:your\\s+time|the\\s+read|a\\s+read|reading|watching|a\\s+listen|a\\s+watch|a\\s+look)\\b|\\bdo\\s+yourself\\s+a\\s+favou?r\\s+and\\s+(?:read|watch|check\\s+out)\\s+(?:this|it)\\b|\\byou\\s+(?:really\\s+)?(?:won(?:'|’)?t|do(?:n(?:'|’)?t|\\s+not)|will\\s+not)\\s+want\\s+to\\s+miss\\s+this(?:\\s+one)?\\s*(?:[:.!?]|$)|\\bdo(?:n(?:'|’)?t|\\s+not)\\s+sleep\\s+on\\s+this(?:\\s+one)?\\b",
+   "hints": ["worth", "favor", "favour", "miss this", "sleep on"],
+   "w": 4,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26",
+   "source": "conorbronsdon/avoid-ai-writing@40328bd"
+  },
+  {
+   "name": "chat-roleplay-action",
+   "cat": "artifact",
+   "rx": "(?:^|[^*])\\*(?:nods?|sighs?|laughs?|smiles?|frowns?|shrugs?|grins?|winks?|chuckles?|gasps?|pauses?|thinks?|wonders?|whispers?|shouts?|gestures?|raises?|leans?|turns?|looks?|glances?|smirks?|blinks?|nodding|sighing|laughing|smiling|thinking|gesturing)\\b[^*\\n]{0,70}\\*(?:$|[^*])",
+   "hints": ["*"],
+   "w": 8,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26",
+   "source": "conorbronsdon/avoid-ai-writing@40328bd"
+  },
+  {
+   "name": "reasoning-artifact",
+   "cat": "artifact",
+   "rx": "\\b(?:let me think (?:this through|step by step)|here(?:'|’)s my thought process|working through this logically|to approach this systematically)\\b",
+   "hints": ["let me think", "thought process", "working through", "approach this systematically"],
+   "w": 6,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26",
+   "source": "conorbronsdon/avoid-ai-writing@40328bd"
+  },
+  {
+   "name": "novelty-inflation",
+   "cat": "rhetorical",
+   "rx": "\\b(?:the (?:failure mode|problem|insight) nobody(?:'|’)?s? (?:is )?(?:naming|talking about)|what nobody tells you|the insight everyone(?:'|’)?s? missing)\\b",
+   "hints": ["nobody", "everyone"],
+   "w": 4,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26",
+   "source": "conorbronsdon/avoid-ai-writing@40328bd"
+  },
+  {
+   "name": "emotional-flatline",
+   "cat": "performed",
+   "rx": "\\b(?:what surprised me most|i was fascinated to (?:discover|learn)|what struck me was|i was excited to learn|the most interesting part)\\b",
+   "hints": ["surprised", "fascinated", "struck me", "excited", "interesting"],
+   "w": 3,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26",
+   "source": "conorbronsdon/avoid-ai-writing@40328bd"
+  },
+  {
+   "name": "acknowledgment-loop",
+   "cat": "artifact",
+   "rx": "\\b(?:to answer your question|you(?:'|’)re asking (?:about|whether)|the question of whether)\\b",
+   "hints": ["answer your question", "asking", "question of whether"],
+   "w": 3,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26",
+   "source": "conorbronsdon/avoid-ai-writing@40328bd"
   }
  ],
  "lexicon": {

package/data/patterns.json

@@ -157,7 +157,7 @@
   {
    "name": "performed-candor",
    "cat": "performed",
-   "rx": "\\b(?:let(?:(?:'|\u2019)s| us) be (?:honest|real|clear)|here(?:(?:'|\u2019)s| is) the (?:thing|truth|kicker|deal|catch)|i(?:'|\u2019)ll be honest|the (?:uncomfortable|hard) truth is|truth be told|no fluff|real talk|not gonna lie|if i(?:'|\u2019)m being honest|let me be (?:crystal |perfectly )?clear)\\b",
+   "rx": "\\b(?:let(?:(?:'|\u2019)s| us) be (?:honest|real|clear)|here(?:(?:'|\u2019)s| is) the (?:thing|truth|kicker|deal|catch|detail|part|wrinkle|nuance|upshot|punchline|rub)|i(?:'|\u2019)ll be honest|the (?:uncomfortable|hard) truth is|truth be told|no fluff|real talk|not gonna lie|if i(?:'|\u2019)m being honest|let me be (?:crystal |perfectly )?clear)\\b",
    "w": 4,
    "first_seen": "2026-08-03",
    "last_confirmed": "2026-08-15"
@@ -523,6 +523,33 @@
    "first_seen": "2026-08-14",
    "last_confirmed": "2026-08-14"
   },
+  {
+   "name": "isocolon-ditransitive",
+   "cat": "contrast",
+   "rx": "\\b(let|give|hand|show|tell|offer|teach|buy|cost|earn|save|bring|send|win)s? you\\b[^.!?]{0,70}[.!?]\\s+[^.!?]{0,70}\\b\\1s? you\\b",
+   "hints": ["you"],
+   "w": 4.5,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26"
+  },
+  {
+   "name": "this-is-what-looks-like",
+   "cat": "contrast",
+   "rx": "(?:^|[.!?]\\s+|\\n)(?:this|that) is what [\\w '\u2019-]{2,40} looks like\\b",
+   "hints": ["looks like"],
+   "w": 4,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26"
+  },
+  {
+   "name": "no-x-had-to",
+   "cat": "contrast",
+   "rx": "(?:^|[.!?]\\s+|\\n)No [\\w '\u2019-]{2,40} (?:had to|needed to|has to|ever) [\\w '\u2019-]{2,60}\\.\\s+[\\w '\u2019-]{2,40} (?:did|made|chose|decided|built|released|shipped|published)\\b",
+   "hints": ["had to", "needed to", "has to", "ever"],
+   "w": 4,
+   "first_seen": "2026-08-26",
+   "last_confirmed": "2026-08-26"
+  },
   {
    "name": "not-the-but-the",
    "cat": "contrast",
@@ -1709,7 +1736,7 @@
   {
    "name": "chatgpt-artifact",
    "cat": "artifact",
-   "rx": "\\boaicite\\b|\\boai_citation\\b|\\bciteturn\\d|\\butm_source=(?:chatgpt(?:\\.com)?|openai)\\b|:contentReference\\[|\\battributableIndex\\b|\\[cite: ?\\d|\\bturn\\d+(?:search|news|image|view)\\d+\\b|\\[span_\\d|\\(start_span\\)|\\bgrok_(?:card|render)|\\bppl-ai-file-upload\\b|\\battached_file\\b",
+   "rx": "\\boaicite\\b|\\boai_citation\\b|\\bciteturn\\d|\\butm_source=(?:chatgpt(?:\\.com)?|openai(?:\\.com)?|copilot(?:\\.com)?|claude\\.ai|perplexity\\.ai|gemini\\.google\\.com)\\b|\\breferrer=grok\\.com\\b|:?contentReference\\[(?:oaicite:)?|\\battributableIndex\\b|\\[cite: ?\\d|\\bturn\\d+(?:search|news|image|view)\\d+\\b|\\[span_\\d|\\(start_span\\)|\\bgrok_(?:card|render)|\\bppl-ai-file-upload\\b|\\battached_file\\b",
    "w": 10,
    "first_seen": "2026-08-14",
    "last_confirmed": "2026-08-22"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zero-slop",
-  "version": "2.5.8",
+  "version": "2.7.0",
   "description": "An Agent Skill that scores AI-sounding prose 0-100 locally and rewrites it without losing a fact. Runs inside Claude Code, Codex, Cursor, Warp, Zed and other SKILL.md agents. Offline, zero dependencies, MIT.",
   "keywords": [
     "agent-skill",

package/references/copy-desk.md

@@ -69,7 +69,7 @@ known voice signals, and immutable facts when available. Then use this brief:
 > terse note only for unresolved ambiguities or unusual forms deliberately kept.
 
 Apply the corrected version to the deliverable, then send that exact artifact
-through the final read-aloud pass in `references/readalong.md`. Verify the
+through the read-aloud pass in `references/readalong.md`. Verify the
 artifact returned by the read-aloud editor:
 
 1. Rerun the heuristic surface scorer and scripted fidelity check.
@@ -77,11 +77,13 @@ artifact returned by the read-aloud editor:
    claims, qualifiers, intended voice, regional spelling, format, and non-prose
    structure. The script cannot detect every semantic or stylistic change.
 3. If any check or comparison requires a repair, apply it, send the repaired
-   text through the copy desk and final read-aloud pass again, and repeat all
+   text through the copy desk and read-aloud pass again, and repeat all
    final checks.
 
-Stop only when the same artifact clears the copy desk, final read-aloud pass,
-semantic and format review, scorer, and fidelity check. Limit this repair loop to
+After verification, send that exact text through the finalizer in
+`references/fresh-eyes.md`. Stop only when the same artifact clears the copy desk,
+read-aloud pass, semantic and format review, scorer, fidelity check, and receives a
+no-change fresh-eyes approval. Limit this repair loop to
 three rounds. If an issue still cannot be resolved without guessing, return the best
 source-preserving version that completed both editorial passes, flag the unresolved
 span and failed check plainly, and do not describe the fallback as fully verified.

package/references/eval.md

@@ -0,0 +1,250 @@
+# Zero Slop eval
+
+Answer every check with pass or fail. Where a check asks for a count, write the
+number down; a count is evidence, and a missing count means the pass did not run.
+
+Any fail sends the text back through the copy desk and read-aloud pass, after which
+every check here runs again on the new text. Limit that loop to three rounds.
+
+**This file exists because the meter cannot see most of what is on it.** The scorer
+reads the lexically anchored subset: listed phrases, sentence-length variance,
+readability, formatting density. Everything in section A is a property of the whole
+document rather than of any span, so no pattern can reach it. A clear score is a
+reason to work through section A carefully, never permission to skip it.
+
+The ratchet: every miss an audit, a competitor, or a reader catches becomes a
+deterministic detector or a `data/corpus/must-flag/` fixture in the same change.
+A note is not a fix; `register.py --recall` proves each recorded miss still
+gets caught, and the release suite runs it.
+
+Families here are drawn from the Zero Slop tell catalogue, from Wikipedia's
+[Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing)
+maintained by WikiProject AI Cleanup, and from the checks in petergyang/no-ai-slop
+and blader/humanizer. Every check that needs one carries a worked before and after,
+because an abstract instruction gets a partial answer.
+
+Roles 7 and 8 own this file. Role 8 answers section F. A role that generated text
+never grades that text.
+
+Scoring this file is meaningless: it quotes tells as examples, so the meter flags its
+own catalogue. That is the documented exception for material quoted as an example, and
+it is a good candidate for `data/corpus/must-not-flag/`.
+
+## A. Register, which the meter cannot reach
+
+Run every item on every draft, including one that scored clear. Record the counts in
+the report even when they are zero.
+
+1. **Antithesis pairs. Count: ___** Two balanced statements where the second lands a
+   twist. Count all four shapes, and note that most carry no negation marker at all:
+   marked ("Not perfect. Honest."), bare subject swap ("Llama is open-weights. Dolma
+   releases the data."), isocolon with the arguments swapped, and unmarked reversal.
+   Budget is one per piece. Two is a finding. Three or more under 500 words means the
+   register failed whatever the score said.
+2. **Binary contrast, the subtractive form. Count per 1,000 words: ___** The
+   corrective appositive: "X, not Y." "A rather than B." Each instance is usually
+   good, careful writing, which is why the meter has no rule for it. The defect is
+   density. Above roughly 4 per 1,000 words, the document has one rhetorical move
+   instead of a voice.
+3. **Significance scaffolding. Count: ___** A sentence announcing that a point
+   matters instead of delivering it. Budget is zero.
+4. **Negative definition.** Does the draft repeatedly say what a thing is not before
+   saying what it is? One is a useful disambiguation, a habit is a tic. Stacked
+   negation counts: "No file monitoring, no browser hooks, no reaching into where you
+   publish" becomes "Zero Slop reads only what you hand it."
+5. **Comma-series density. Count per 1,000 words: ___** Enumerations of three or more
+   noun phrases. Technical writing legitimately enumerates, so this is a reading
+   judgment rather than a threshold: are the lists carrying information, or has the
+   sentence shape become the default? Convert the longest to real lists.
+6. **Robotic symmetry.** Repeated sentence shapes, identical paragraph geometry, stacked
+   punchy fragments, uniform paragraph length. Check table columns too: seven of eight
+   cells opening with a verb and a comma list of abstract nouns is the same defect
+   inside a grid.
+7. **Colon reveals.** A noun phrase, a colon, then a lowercase dramatic payoff.
+   Colons are for lists, labels, and quotations.
+8. **Throat-clearing openers.** "Here's the thing." "Let me be clear." "To be
+   honest." Cut and state the point.
+9. **Faux-insight setups.** "What most people get wrong." "The part everyone misses."
+   These flatter the writer as sole expert. Make the claim stand alone.
+10. **Rhetorical setups.** "What if I told you." "Think about it:" Self-answered
+    question and answer pairs.
+11. **Fake-profound kickers.** Delete the closing aphorism rather than rewriting it
+    into a better metaphor. End on the clearest concrete sentence already present.
+12. **Summary-recap endings.** The reader was just there.
+13. **Theatrical framing and extended conceit. Count: ___** An ordinary process
+    dressed as a courtroom, a heist, or a diagnosis. Budget is one metaphor.
+14. **Superficial analysis.** Trailing participial clauses that pretend to explain
+    significance: highlighting, underscoring, reflecting, showcasing.
+15. **Importance puffery.** "Marks a pivotal moment." "Stands as a testament." State
+    the fact and let the reader weigh it.
+16. **Interpretive metadiscourse.** Lines that step outside the subject to tell the
+    reader what to notice or how much weight to give it.
+17. **Weasel attribution.** "Studies show." "Experts agree." Name the source, or flag
+    it for the writer. Never invent one.
+18. **Adjective inflation and hollow intensifiers.** "a real improvement", "actual
+   results", "a genuine breakthrough", "true value". Delete the adjective; the noun
+   carries the claim or it does not. Distinct from empty adverbs (next): adverbs pad
+   the verb, these inflate a noun, and a span that is neither belongs to importance
+   puffery. Measured by the meter; the reading pass owns only the shapes the noun
+   list misses.
+19. **Empty adverbs and filler frames.** Adverbs: "just", "simply", "actually",
+    "literally", "honestly", "fundamentally", "inherently", "inevitably". Frames:
+    "the reality is", "the truth is", "in terms of", "with regard to", "going
+    forward", "in this article". Cut each one that adds nothing; keep the ones
+    carrying real emphasis, uncertainty, contrast, or the writer's spoken rhythm.
+    Distinct from adjective inflation (previous): these modify verbs and clauses.
+    The meter does not flag these globally because they are ordinary in honest
+    writing, so this is a judgment per instance. "just", "simply", "actually", "literally",
+    "honestly", "fundamentally", "inherently", "inevitably". Cut each one that adds
+    nothing; keep the ones carrying real emphasis, uncertainty, contrast, or the
+    writer's spoken rhythm. This is a judgment per instance, which is why the meter
+    does not flag these words globally: they are ordinary in honest writing.
+20. **Nominalization and weak verb phrases.** "Made a decision" for "decided",
+    "needs removal from" for "you cut from". De-nominalize, and prefer a direct verb
+    with an actor.
+21. **Fake-strong verbs.** Prefer "is" and "has" where they are clearer.
+    "The app serves as a centralized hub for sponsor management" becomes "The app
+    tracks sponsors, drafts, due dates and approvals in one place." Monument verbs
+    count too: "stands on", "sits atop", "marks a".
+22. **Reasoning-chain artifacts.** Working-out left in the answer: "Let me think
+   through this", "First I'll consider", a restated plan before the plan. The reader
+   wants the conclusion, not the deliberation.
+23. **Emotional flatline.** Uniform affect across a piece that should vary. Nothing
+   irritates the writer, nothing surprises them, nothing costs anything. Real writing
+   has range.
+24. **False concession.** "While X has merit, Y" where X was never a live position
+   and nothing follows from conceding it. Keep a real counterargument; cut a straw one.
+25. **Confidence-calibration phrases.** "I'm fairly confident", "with high certainty",
+   "to be clear, this is my read". Calibrated hedges are fine when they carry real
+   uncertainty; these announce a posture instead.
+26. **Parenthetical hedging.** Qualifiers tucked in brackets so the sentence can claim
+   more than it supports: "The fix works (in most cases) for every driver."
+27. **Engagement hooks and endorsement closers.** "You won't believe", "Sound
+   familiar?", "Give it a try and let me know", "Trusted by teams everywhere".
+   Infomercial register, wherever it appears.
+28. **Lingering-attention claims.** "This will stay with you", "you'll be thinking
+   about this for weeks". The writer cannot know the reader's future.
+29. **Generic positive endings.** A closing paragraph that resolves into optimism
+   without a concrete next step: "The future looks bright for teams willing to adapt."
+30. **Synonym cycling.** If the clear word is right, repeat it. Do not rotate terms
+    for style. "The agent reviews the draft. The assistant scores the piece. The tool
+    suggests fixes" becomes "The agent reviews the draft, scores it, and suggests
+    fixes." Check every referent, not the first one you notice: a role table is where
+    a second and third name usually appear.
+
+## B. Substance
+
+31. **Removal test.** Does every paragraph lose something real when deleted?
+32. **Relevance test.** Does every paragraph serve the brief, audience, and argument?
+33. **Front-loading, applied selectively.** Conclusions arrive early where that helps
+    the reader, without forcing every section into the same point-then-detail shape.
+34. **Paragraph-order dependence.** Could several prose paragraphs be shuffled without
+    harming the argument? Reference material, FAQs, and independent findings are
+    exempt.
+35. **Unsupported novelty.** "Nobody is naming this" needs an actual comparison.
+36. **Self-labeling significance.** "This matters" needs a consequence.
+37. **Moral-adjective category error.** Calling a technical choice brave or honest
+    needs a moral agent.
+38. **Portability test.** Could any sentence move unchanged to another company,
+    product, or person? Cut it or make it specific.
+39. **Statistics cohesion.** Does each test or dataset get its own paragraph opening
+    with what it checks in plain words, before the numbers?
+40. **Cross-references resolve.** Every named artefact, panel, dataset, or prior
+    result is defined on first use or linked to where it lives. "The prior 84.2%
+    result on the 38-item editorial panel" tells a reader nothing if neither the
+    panel nor the packet is named anywhere.
+41. **Internal pointers name their target.** "In the references", "see the docs",
+    "as documented elsewhere" point at nothing. Name the file and the section, and
+    link it, the way an external citation would be named.
+42. **Numeric precision matches the measurement.** Five significant figures on a
+    single run of a single machine claims a stability the method cannot support.
+    Round to what the measurement earns, or state the spread.
+43. **Hollow spans flagged, not filled.** Prose that makes no claim was flagged for
+    the writer rather than reworded into something that sounds like one.
+44. **Sycophantic tone.** Generic validation of the reader that carries no content:
+    "Great question", "Excellent point", "You're absolutely right". It echoes nothing
+    back, it only rewards. Cut it and start with the answer.
+45. **Recap-flattery.** An opener that echoes the reader's *own work* back at them
+    dressed as gratitude, ahead of the point: "Thanks for all the legwork here, the
+    migration script and the rollback plan you worked through are what made this
+    possible." They already know what they did. A real thank-you is one clause and
+    moves on: "Thanks for the legwork, this looks right, one comment below."
+46. **Acknowledgment loops.** Restating the question or the prior section before
+    answering it. This echoes the *context*, where recap-flattery echoes their *work*
+    and sycophancy echoes *nothing*. Ask what is being echoed; that is what separates
+    these three.
+47. **Wall-of-text reply.** Paragraphing that hides a sequence the reader needs. A
+    long narrative paragraph is not a wall of text merely because it is long; the tell
+    is a buried list of steps or options.
+
+## C. Fidelity
+
+48. **Scripted check run, not eyeballed.** `slopscore.py --fidelity` exits zero.
+49. **No invented specifics.** No number, name, anecdote, date, or source appeared
+    that the author did not supply.
+50. **No invented interior claims.** No stated feeling, motive, or experience the
+    author never wrote.
+51. **Qualifiers survive.** Hedges, scope limits, and caveats carry the same strength
+    as the source.
+52. **Claims not reframed.** Same names and numbers can still carry a changed
+    emphasis or implication. Compare meaning, not tokens.
+53. **Non-prose untouched.** Code, front matter, tables, blockquotes, identifiers,
+    paths, and heading hierarchy intact.
+
+## D. Voice and readability
+
+54. **Expert voice.** A respected practitioner sounds at home: precise terms used
+    correctly, authority earned through specifics, no hedging into mush.
+55. **Followability.** A smart first-time reader follows each sentence on the first
+    pass. One idea per sentence; every abstraction anchored in the same breath; never
+    three or more abstract noun phrases stacked.
+56. **Voice preserved.** Would the writer recognize this as theirs? Distinctive
+    vocabulary, cadence, bluntness, humor, uncertainty, and digressions survived.
+57. **Strong sentences left alone.** Nothing was rewritten merely for consistency.
+58. **Active voice with human subjects.** People do the verbs. Passive stays only
+    where the actor is unknown, irrelevant, deliberately withheld, or native to the
+    genre.
+59. **Proportional cutting.** No compression that stripped character. Density is
+    information per word, not fewer words.
+60. **Read aloud.** Would this sound natural read to a sharp colleague?
+61. **No over-correction.** AI slop was not traded for edgy slop: forced hot takes,
+    fake first person, performed candor, staccato drama.
+
+## E. Form
+
+62. **Returned in the format it arrived in.** A .docx comes back a .docx.
+63. **Reader language.** No evaluator or harness vocabulary leaked into the copy:
+    candidate, artifact, overlay, gate, scorecard, burstiness.
+64. **Formatting slop.** No emoji headings, decorative bold, or bullets that should be
+    prose. Count the headings over one or two sentences: four consecutive H3s each
+    wrapping a code block and one line is a section that wants to be one section.
+65. **Em dashes.** None in short copy. One or two in a long draft only where they
+    clearly beat a comma, period, or parenthesis.
+66. **Hyphenated modifier stacking.** "AI-powered cloud-native data-driven platform".
+   Two stacked compound modifiers in one noun phrase is a tell; keep the hyphens
+   grammar requires and cut the rest.
+67. **List-label periods and bare-noun bullets.** Bullets that are fragments ending in
+   full stops, or a list of bare noun phrases with no predicate. Either make them
+   sentences or make them a real list.
+68. **Numbered-list inflation.** A numbered list where order carries no meaning, or a
+   list padded to a round number. Use a bulleted list, or prose.
+69. **Chatbot artifacts.** Citation markup leaks, "Regenerate response", tool URL
+   parameters such as utm_source, and any assistant scaffolding left in the copy.
+70. **Attribution accurate.** The report names who did what without guessing which
+    model is running, and never implies a Zero Slop service read the draft.
+
+## F. Process integrity
+
+Role 8 answers this section. It is the check no single-agent eval can make.
+
+71. **Roles stayed separate.** The copy desk, read-aloud pass, verification, and
+    fresh-eyes review each ran as a distinct pass.
+72. **No self-certification.** No role graded text it generated.
+73. **Counts reported.** Every count in section A appears in the summary, including
+    the zeros.
+74. **The exact final text cleared every check.** Not an earlier draft, not a version
+    that was repaired afterward.
+75. **Role 8 approved without changes.** If it changed anything, roles 5 through 8
+    ran again on the revision.
+76. **Fallbacks named honestly.** If the three-round limit was reached, the report
+    says which check failed and does not describe the result as fully verified.

package/references/evidence.md

@@ -47,6 +47,28 @@ not a deception.
 8. **Affect skew.** LLM text is joy-skewed and uniformly positive
    (Muñoz-Ortiz). Counter: widen affect.
 
+## Incumbent audit: what transferred and what did not
+
+Zero Slop v2.5.9 audited
+[`conorbronsdon/avoid-ai-writing`](https://github.com/conorbronsdon/avoid-ai-writing)
+at commit `40328bd292bc682d46010a6f9ac2cdbf4fb4ceca`. The ideas that survived
+transfer tests were zero-width and mixed-script normalization, exact protection
+for structured document spans, broader AI-tool tracker residue, three narrowly
+defined phrase families, and one conservative long-form word-variety signal.
+On the incumbent's pinned 875-human/779-machine
+paragraph corpus, its low-TTR signal fired on 1 human and 20 machine paragraphs
+(22.46x lift). Zero Slop keeps it weak and corroboration-dependent.
+
+Function-word entropy, punctuation-distribution uniformity, and cross-paragraph
+rhythm were also tested. None fired on Zero Slop's 38 consensus-labelled
+editorial passages; on the Beemo paired set they were rare and only directional
+proxies because Beemo labels provenance and editing history, not slop quality.
+They were not promoted. The incumbent's hand-shaped class probabilities were
+also rejected: its documentation correctly says they are not calibrated against
+a labelled corpus, and its own composite reports paragraph ROC-AUC 0.501 and
+document ROC-AUC 0.623. Zero Slop does not turn an uncalibrated writing score
+into an authorship probability.
+
 ## Why the scorer measures features, not detector verdicts
 
 Detectors are brittle: RAID (2405.07940) shows trivial perturbations fool

package/references/fresh-eyes.md

@@ -0,0 +1,62 @@
+# The fresh-eyes finalizer
+
+This is the last role in the editing workflow. It receives the exact text that
+already cleared the copy desk, read-aloud editor, and verifier. Its job is to read
+that text as a first-time reader and catch the small but consequential problems an
+editor can stop seeing after repeated passes.
+
+Use a new, role-isolated editor when the harness supports agents or subagents.
+Otherwise start a separate pass with no drafting or scoring mindset. Do not ask the
+rewriter to approve its own work.
+
+## What to check
+
+- **The whole arc.** The opening sets up the real point, sections arrive in a useful
+  order, and the ending resolves rather than repeats the piece.
+- **First-use clarity.** Names, pronouns, abbreviations, references, and transitions
+  make sense when encountered, without knowledge from the editing process.
+- **Reader effort.** A smart first-time reader can follow the argument once through.
+  Unpack only a genuine stumble; do not dilute expert language the audience needs.
+- **Voice continuity.** The same person seems to have written the beginning, middle,
+  and end. Restore source-backed warmth or texture lost to over-editing.
+- **Residual process language.** Remove internal labels, scoring talk, evaluation
+  jargon, or comments meant for the editor rather than the reader.
+- **Genre and format.** The text looks and sounds native to its destination. Preserve
+  all legitimate lists, tables, code, headings, citations, links, and layout.
+- **Last-mile polish.** Correct a safe leftover repetition, vague handoff, awkward
+  phrase, or mechanical inconsistency that survived the earlier passes.
+
+## Boundaries
+
+Preserve every fact, claim, qualifier, name, number, date, link, quotation, source,
+voice signal, regional spelling choice, and protected non-prose element. Do not add
+examples, experiences, certainty, novelty, emotion, or specificity. Do not make a
+substantive rewrite. If the intended correction is uncertain, leave the text alone
+and name the ambiguity.
+
+## Finalizer brief
+
+> You are the fresh-eyes finalizer. Read the complete verified text as a first-time
+> reader, from title to final line. Apply only safe last-mile polish for overall arc,
+> first-use clarity, reader effort, voice continuity, genre fit, residual process
+> language, repetition, and mechanical consistency. Preserve every fact, claim,
+> qualifier, name, number, date, link, quotation, source, voice signal, regional
+> spelling choice, and non-prose element. Do not add detail, certainty, emotion,
+> novelty, or generic polish. Return the complete text in its original format,
+> followed by exactly one status: `approve without changes`, `changed — rerun final
+> checks`, or `unresolved — needs the writer`.
+
+## Closed finalization loop
+
+Approval is valid only when the finalizer returns `approve without changes`. If it
+changes even one character, apply the complete revision and rerun, in order:
+
+1. the copy desk;
+2. the read-aloud editor;
+3. the verifier, including local and contextual checks; and
+4. this fresh-eyes finalizer.
+
+Repeat for at most three rounds. The same exact text must clear all four roles, with
+the fresh-eyes finalizer approving it unchanged. If the remaining issue cannot be
+fixed without guessing, return the best source-preserving text that completed every
+pass, identify the unresolved span, and do not describe it as fully verified.

package/references/overcorrection.md

@@ -56,6 +56,11 @@ experience — these alone are NOT evidence of AI:
 Require corroboration. A paragraph needs multiple independent tells, or a failed
 removal test, before it's slop.
 
+This governs lexical flags only. It does not apply to the performed-register
+family: register is a property of the piece, not of a paragraph. Four unmarked
+antithesis pairs across four paragraphs *is* the corroboration — each one is
+locally defensible, and the repetition is the whole finding.
+
 ## Signs of human writing — preserve on sight
 
 When a draft shows these, protect them through the rewrite; deleting them is

package/references/readalong.md

@@ -6,8 +6,8 @@ listener hears it: from the first line to the last, with each sentence setting u
 next. A stumble, cold transition, unclear antecedent, or change of voice can still
 survive a clean score and correct grammar.
 
-Run this pass on every deliverable after the copy desk. It is the last editorial pass
-before final verification and the Report step. Use a dedicated editor with fresh eyes
+Run this pass on every deliverable after the copy desk. It is the spoken-flow pass
+before verification and the separate fresh-eyes finalizer. Use a dedicated editor with fresh ears
 when the harness supports subagents; otherwise perform a separate, role-isolated pass.
 The editor must fix the actual deliverable and return the corrected artifact in the
 same format. An audit or list of suggestions is not a completed pass.
@@ -21,17 +21,33 @@ safe correction for:
   endings that make a reader stop or back up.
 - **Cold transitions.** Add or repair the hinge when a sentence, paragraph, or
   section arrives without a clear connection to what came before.
+- **Paragraph-order dependence.** If adjacent prose paragraphs can trade places
+  without changing the argument, rebuild the progression. Do not impose a story
+  arc on FAQs, reference entries, independent findings, or genuine lists.
 - **Performed candor.** Remove announcements of honesty such as "honestly" or "to be
   fair" when the sentence can simply make its point.
 - **Reflexive agreement.** Remove automatic praise or agreement that appears before
   the substance has been checked. Let the facts determine whether the draft agrees,
   qualifies, or disagrees.
+- **Recap-flattery.** In a reply, cut praise and a needless restatement of the
+  question before the answer. Keep setup that materially changes the answer.
+- **Wall-of-text reply.** Break an answer where the reader must change task,
+  decision, time, or subject. Do not split a coherent narrative paragraph merely
+  to make it look lighter.
 - **Communicative drift.** Stop when fluent sentences no longer advance a clear
   point or reader need. Cut the passage, rebuild it around its purpose, or flag the
   missing intent rather than inventing one.
 - **Rhetorical scale mismatch.** Reduce grand reveals, lessons, or contrasts when
   the underlying point is too small or unsupported for the framing. Preserve a
   contrast that corrects a real, supported misconception.
+- **Unsupported novelty.** Remove claims that nobody has noticed or named an idea
+  unless the source supplies a defensible comparison.
+- **Self-labeling significance.** Replace an unsupported declaration that a point
+  is important with the consequence that makes it important.
+- **Moral-adjective category error.** Do not call a system design, metric, or
+  technical choice brave, honest, ethical, or courageous when the sentence means
+  accurate, transparent, conservative, or reliable. Preserve a sourced moral
+  judgment about an actual agent or decision.
 - **Performed-writer register.** Flatten theatrical framing of ordinary work ("we
   hired an adversary"), epigram closers, staccato antithesis pairs ("Not perfect.
   Honest."), extended conceits (billing, courtroom, forensics, recipe), one-word
@@ -58,6 +74,8 @@ safe correction for:
 Do not merely flag these problems. Correct them directly wherever the intended
 meaning is certain.
 
+For consistent reporting, use these names: paragraph-order dependence, unsupported novelty, self-labeling significance, moral-adjective category error, recap-flattery, and wall-of-text reply.
+
 ## Boundaries
 
 - Preserve every fact, claim, qualifier, number, name, date, link, and source.
@@ -103,7 +121,8 @@ Verify the exact artifact returned by the read-aloud editor:
    this read-aloud pass again, and repeat every final check.
 
 Stop only when the same artifact has cleared the copy desk, final read-aloud pass,
-semantic and format review, scorer, and fidelity check. Limit this repair loop to
+semantic and format review, scorer, fidelity check, and the separate fresh-eyes
+finalizer in `fresh-eyes.md`. Limit this repair loop to
 three rounds. If an issue still cannot be resolved without guessing, return the best
 source-preserving version that completed both editorial passes, state the unresolved
 span and failed check plainly, and do not describe the fallback as fully verified.