zero-slop 2.5.8

package/package.json

@@ -0,0 +1,54 @@
+{
+  "name": "zero-slop",
+  "version": "2.5.8",
+  "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",
+    "agent-skills",
+    "skills",
+    "claude-code",
+    "claude-skill",
+    "codex",
+    "cursor",
+    "ai-slop",
+    "de-slop",
+    "humanizer",
+    "humanize-ai-text",
+    "ai-detection",
+    "ai-writing",
+    "writing-tools",
+    "editing",
+    "prose"
+  ],
+  "homepage": "https://zero-slop.ai",
+  "bugs": "https://github.com/manavmishra/ZeroSlop/issues",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/manavmishra/ZeroSlop.git"
+  },
+  "license": "MIT",
+  "author": {
+    "name": "Manav Mishra",
+    "url": "https://zero-slop.ai/about/"
+  },
+  "files": [
+    "SKILL.md",
+    "references/",
+    "scripts/",
+    "data/",
+    "README.md",
+    "LICENSE",
+    "!scripts/build_bundle.py",
+    "!scripts/build_onepager_pdf.py",
+    "!scripts/build_plugin.py",
+    "!scripts/build_skill_zip.py",
+    "!scripts/check_svg.py",
+    "!scripts/contextual.py",
+    "!scripts/contextual-signals.md",
+    "!references/contextual-signals.md",
+    "!scripts/growth-snapshot.mjs",
+    "!data/voices/",
+    "!**/__pycache__/"
+  ],
+  "private": false
+}

package/references/copy-desk.md

@@ -0,0 +1,87 @@
+# The final copy desk
+
+Run this pass on every deliverable after the rewrite and initial verification
+work. It corrects mechanics and line-level style before the final read-aloud
+pass. Its job is to produce corrected copy, not an assessment of copy that still
+needs work.
+
+Use a dedicated copy-editor agent with fresh eyes when the harness supports
+subagents. Otherwise, set aside the writing mindset and perform the same review
+as a separate, role-isolated pass. Either way, apply every accepted correction
+to the actual deliverable before sending it to the final read-aloud pass.
+
+## What to correct
+
+Read the complete piece from title to final line, including headings, captions,
+list items, table labels, calls to action, and user-authored alt text. Correct:
+
+- spelling, typos, duplicated or missing words, and inconsistent regional usage;
+- grammar, agreement, tense, articles, prepositions, pronouns, and modifiers;
+- punctuation, capitalization, quotation marks, hyphenation, and compound terms;
+- broken parallelism, unclear antecedents, misplaced clauses, and ambiguity;
+- awkward syntax, needless repetition, imprecise diction, and choppy transitions;
+- phrasing that is stiff, clumsy, fussy, unintentionally casual, or unprofessional;
+- internal scoring, editing, or workflow labels that have leaked into reader-facing
+  prose; replace each with plain language unless the audience needs the technical
+  term;
+- inconsistencies in terminology, headings, labels, voice, and editorial style.
+
+Aim for tasteful, elegant, professional prose in the register the piece calls
+for. Elegance means precise, natural, and restrained. It does not mean ornate,
+formal, promotional, or generic. Preserve warmth, humor, contractions, fragments,
+dialect, and technical language when they are intentional and suit the genre.
+Keep the author's regional spelling unless the user or publication specifies a
+different house style.
+
+## Boundaries
+
+- Preserve every fact, claim, qualifier, number, name, date, link, and source.
+- Do not invent detail, strengthen certainty, soften a warranted conclusion, or
+  change the author's position for smoother prose.
+- Do not edit quotations, code, commands, identifiers, file paths, data, legal
+  boilerplate, or proper names unless the user explicitly asks.
+- Preserve the input format and all non-prose structure. Correct content inside
+  a structured form without changing the form itself.
+- Do not replace a distinctive human phrase merely because a more conventional
+  phrase exists. Correct errors and genuine awkwardness, not personality.
+- When a sentence is ambiguous and the intended meaning cannot be recovered
+  safely, leave it unchanged and flag the ambiguity instead of guessing.
+
+## Copy-editor agent brief
+
+Give the agent the genre, intended audience, original draft, selected rewrite,
+known voice signals, and immutable facts when available. Then use this brief:
+
+> You are the copy editor for a publication-ready piece. Edit the complete
+> text directly and return the corrected version in the same format. Fix every
+> genuine error in spelling, grammar, punctuation, capitalization, agreement,
+> tense, modifiers, syntax, diction, consistency, and usage. Improve awkward,
+> unclear, repetitive, clumsy, or unprofessional phrasing where the meaning is
+> certain. Remove internal evaluation or workflow language from outward-facing
+> prose when plain language says the same thing. Make the prose tasteful, elegant,
+> precise, natural, and appropriate for its genre and audience. Preserve the
+> author's voice, regional spelling, facts, claims, qualifiers, names, numbers,
+> links, structure, and intentional fragments. Do not add detail, hype, certainty,
+> or generic polish. Do not alter
+> quotations, code, commands, identifiers, or data. If a correction would
+> require guessing what the author means, leave that span unchanged and flag it.
+> Return: (1) the fully corrected text, not a list of suggestions; and (2) a
+> 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
+artifact returned by the read-aloud editor:
+
+1. Rerun the heuristic surface scorer and scripted fidelity check.
+2. Compare it directly with both the original and selected rewrite for preserved
+   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
+   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
+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/evidence.md

@@ -0,0 +1,382 @@
+# Evidence Base — why the ladder is ordered the way it is
+
+Every rule in this skill traces to measured findings. This file is the chain
+of custody. Full dossier with per-paper notes lives in the research archive;
+citations here are the load-bearing ones.
+
+## The central finding
+
+Commercial detectors (GPTZero, Pangram) rate *base-model* text as ~97–99%
+human while flagging *instruction-tuned* output from the same model
+(arXiv:2605.19516). Detectors are classifiers of the **post-training (RLHF)
+register** — the polished, uniform, preference-optimized voice — not of
+machine generation itself. That register lives entirely in surface
+realization, which is why rewriting can remove it without touching meaning.
+It is also why "humanizing" is legitimate editing: the target is a register,
+not a deception.
+
+## Detection features, ranked by evidence strength
+
+1. **Token-level predictability.** AI text sits at local maxima of model
+   log-probability (DetectGPT, arXiv:2301.11305; Binoculars, 2401.12070).
+   Human text doesn't. **Now measured** (v2.2, `scripts/predictability.py`) —
+   see the model-channel note below. Counter: Ladder L1: specific, slightly
+   surprising phrasing and concrete facts are the direct counter.
+2. **The LLM lexicon.** A few hundred style words carry huge evidential
+   weight: "meticulous" +34.7x, "commendable" +9.8x, "intricate" +11.2x in
+   post-ChatGPT scientific text (Liang, 2403.07183); ~900 excess words
+   catalogued across 15M PubMed abstracts (Kobak, 2406.07016 —
+   github.com/berenslab/llm-excess-vocab); 21 focal words traced to RLHF
+   (Juzek & Ward, 2412.11385). Era-dependent: delve peaked 2023–24;
+   enhance/highlight/showcase dominate 2025+. Counter: Ladder L5 + the scorer's
+   weighted lexicon + the learning loop's era updates.
+3. **Surprisal uniformity.** Humans spike information density unevenly; LLMs
+   smooth it (GPT-who/UID, 2310.06202 — beats commercial detectors by >20%). Counter: L3's "don't pad every claim to equal weight".
+4. **Low burstiness.** Uniform sentence length/structure: GPTZero's founding
+   feature, corroborated independently (Muñoz-Ortiz, 2308.09067; Reinhart,
+   PNAS 2410.16107). Human sentence-length CV is simply higher. Counter: the
+   scorer's burstiness metric and the ≥0.45 gate.
+5. **Register rigidity.** LLMs hold one polished expository voice regardless
+   of situation; humans shift register (Reinhart). Counter: L4.
+6. **Stance asymmetry.** Humans: first-person stance, modals, selective
+   epistemic hedging, discourse-marker cohesion. LLMs: nominalizations,
+   formal connectives, paragraph-architecture cohesion (Herbold, 2304.14276). Counter: L4's hedging rules and "connective texture over scaffolding".
+7. **Syntax signature.** Nominalization density, present-participial
+   clauses, longer constituents (Reinhart; Muñoz-Ortiz). Counter: de-nominalize;
+   kill participial openers.
+8. **Affect skew.** LLM text is joy-skewed and uniformly positive
+   (Muñoz-Ortiz). Counter: widen affect.
+
+## Why the scorer measures features, not detector verdicts
+
+Detectors are brittle: RAID (2405.07940) shows trivial perturbations fool
+them, and DIPPER (2303.13408) shows one paraphrase pass drops DetectGPT from
+70% to 4.6% detection. Verdicts are therefore neither necessary nor
+sufficient. The features they key on, however, are exactly what human readers
+report as "sounds like AI" — so the scorer tracks the features directly:
+weighted tell density, lexicon hits, burstiness, formatting densities,
+register signals. Passing the gate means "the measurable tells are gone",
+which is the honest, robust target.
+
+## Span-first diagnosis, not a binary vibe check
+
+Shaib et al., *Measuring AI “Slop” in Text* (arXiv:2509.19163), built a
+taxonomy from 19 experts and span annotations by professional copy editors.
+The useful split is broader than style alone: information utility (density and
+relevance), information quality (factuality and appropriate perspective), and
+style quality (repetition, templatedness, coherence, fluency, verbosity, word
+complexity, and tone). Which dimensions mattered changed by domain. Factual and
+structural problems mattered most in short answers; utility and tone mattered
+more in news.
+
+The negative results are just as important. Pairwise agreement on the binary
+slop label was poor to fair, automatic linear models reached only 0.52 and 0.55
+AUPRC on the two datasets, and zero-shot LLM judges under-predicted slop with
+recall of 0.08–0.12. Prompted span extraction also aligned poorly with the human
+annotations. That evidence rules out an ungrounded "does this feel like slop?"
+model verdict as a reliable gate. Zero Slop therefore uses a span-first,
+category-specific diagnosis, keeps deterministic surface measurements separate,
+and reserves relevance, coherence, tone, and factual judgments for explicit
+review rather than laundering them into the 0–100 meter.
+
+Source: <https://arxiv.org/abs/2509.19163> (CC BY 4.0).
+
+## Cross-draft templating
+
+The Slop Index evaluates 19,928 model generations against pre-ChatGPT human
+baselines. Its most portable idea for an editor is not its composite ranking;
+it is measuring repeated five-word openings across several responses to the
+same prompt. A single document cannot expose that failure. The same project
+also found that some plausible measures reverse direction by genre, so its
+rhythm axis is used only for email and its weights are renormalized when a
+baseline cannot support an axis.
+
+Zero Slop adopts the conservative part of that method in `--portfolio`: report
+exact repeated openings and shared multiword templates across three or more
+related drafts. The result does not change the surface score. There is not yet
+enough labeled, cross-genre evidence to assign it a safe universal weight, and
+necessary domain phrases can legitimately recur.
+
+Source: <https://github.com/hgaddipati1118/slop-index> (MIT).
+
+## What rewriting cannot do (the honesty boundary)
+
+- Retrieval/watermarking by providers survives any rewrite (DIPPER's
+  conclusion). Fine — this skill's goal is reader-experienced quality, not
+  evasion.
+- Character-level tricks and fake typos fool detectors (RAID) but degrade
+  writing. Banned.
+- Hollow content scores clean on every surface metric. Only the removal test
+  catches it, which is why the judgment pass can never be skipped and why
+  hollow spans are flagged, not padded.
+- The fidelity check compares tokens and entities, not meaning. v2 hardened it
+  against the two false alarms it raised most — a number spelled out ("18"
+  versus "eighteen") and a common word capitalised at a sentence start read as
+  an invented name — but a paraphrase that keeps every entity while bending a
+  claim can still pass it. Semantic fidelity, entailment between the draft and
+  the rewrite, is the roadmap; until then the judgment pass carries what the
+  token check cannot see.
+
+## Reader-reported salience: the Reddit study
+
+JCarterJohnson's 2026 analysis pulled 89,239 Reddit posts from 47 subreddits
+covering 2021–2026, filtered 7,984 posts about recognizing AI-flavoured
+writing, and manually reviewed a 600-post high-engagement sample. In that
+sample, readers cited flat rhythm, reflexive praise or agreement, formulaic
+shape, and fluent-but-empty prose more often than most individual words. The
+keyword pass produced the opposite error: ordinary words such as "however",
+"thus", "hence", "nuanced", "comprehensive", and "utilize" matched often
+even though audited readers almost never named them as tells.
+
+This is useful evidence about reader salience, not a prevalence estimate or
+authorship detector. The source describes its audience as vocal, online
+people; the sample is biased toward recent, high-engagement posts, and its
+keyword analysis can confuse using a term with discussing it. Zero Slop uses
+the relative ordering to put meaning, stance, rhythm, and document shape
+before isolated vocabulary. It adds three context-only review traits:
+reflexive agreement, communicative drift, and rhetorical scale mismatch. It
+does not import the study's percentages as weights, ban em dashes, or promote
+ordinary connectors into the mechanical scorer.
+
+The accompanying `unslop-ai-text` skill and scanner were audited at commit
+`f7c4aefc2c797a66e55b49354a93917ab60d33ac`. Its severity levels, JSON/CI
+interface, sample-based register guidance, and read-aloud review are useful
+corroboration. Zero Slop already covered all 25 categories in the published
+tally through its six-family taxonomy, contextual review, scorer, voice
+profile, and final editorial passes. We adopted the useful machine-readable
+batch contract and reader-priority lesson, but no code, weights, or blanket
+single-mark rule. On Zero Slop's 13-file known-human safety set, that upstream
+scanner reported three high, two medium, and one low finding and returned a
+"strong" verdict; Zero Slop kept all 13 files below its gate. That comparison
+is a false-positive safety check, not an independent test of editorial quality.
+
+Sources:
+
+- Study and methodology: <https://www.reddit.com/r/ClaudeAI/comments/1ucpw87/i_pulled_90000_reddit_posts_about_what_makes/>
+- Skill comparison: <https://www.reddit.com/r/ClaudeAI/comments/1uel1dc/unsloptext_skill_vs_humanizer_skill_part_2/>
+- Launch discussion: <https://www.reddit.com/r/ClaudeAI/comments/1udl9hg/unsloptext_a_claude_skill_that_flags_and_removes/>
+- Data, scanner, and skill: <https://github.com/JCarterJohnson/vibecoded-design-tells/tree/main/unslop-ai-text> (MIT).
+
+## Practitioner corroboration
+
+Paul Graham's "Write Like You Talk" and "How to Write Usefully" predate LLMs
+and independently prescribe the same counters: spoken register (high
+perplexity relative to formal boilerplate), maximal-strength claims without
+overclaiming, qualification as precision. Wikipedia's WP:AICATCH — the
+largest human-curated corpus of caught-in-the-wild AI text — converges on the
+same tell families and adds the cluster rule this skill inherits: one tell is
+coincidence; repeated agreement across several tells is meaningful.
+
+## Negative results: trained classifiers
+
+Several trained-classifier approaches were evaluated for an additional
+detection channel. None shipped, and the reasoning generalises beyond the
+specific methods.
+
+**The ones that added nothing.** Two well-established classifier families were
+ruled out on published evidence: one shows no measurable gain over a simpler
+model on text classification while adding a dependency to a package whose value
+includes having none, and the other models a sequence signal that the burstiness
+and followability statistics already capture more cheaply and more legibly.
+
+**The one that worked, and was cut anyway.** A trained channel built on lexical
+frequency and stylometric features performed well in-domain: 0.985 AUC and 94%
+accuracy on held-out data, properly calibrated with an abstain band. It was
+built, integrated, and then removed. The transfer test is why. Trained on
+2022-era text, it rated 2026-era AI drafts as human at a mean probability of
+0.038, and in a live check returned 0.33 on a passage the pattern meter scored
+100 out of 100. Detector decay across model generations is well documented
+(RAID, arXiv:2405.07940); this was that decay measured directly. A channel that
+is confidently wrong on current text is worse than no channel, even reported
+separately and labelled a second opinion.
+
+The lesson generalises. Interpretable surface features degrade gracefully as
+models change, because updating them means editing a data file and the drift is
+visible in the diff. A trained classifier degrades silently, and silence is the
+failure mode you cannot audit.
+
+## Fairness: who a false positive lands on
+
+Liang et al. ran seven commercial GPT detectors over two human-written corpora,
+US eighth-grade essays and TOEFL essays by non-native English speakers. Native
+samples scored near-perfectly. **More than half of the non-native samples were
+misclassified as AI-generated** (arXiv:2304.02819; *Patterns* 4:100779, 2023).
+
+The mechanism matters more than the headline. The same study found that
+enriching word choice in the non-native samples reduced misclassification,
+while simplifying the native samples increased it. The detectors were keying on
+linguistic complexity, so anything measuring "does this read as polished" will
+penalize the writers with the least room to perform polish.
+
+That is a direct hazard for this skill, which measures surface features by
+design. Two consequences are load-bearing rather than decorative. Burstiness
+and followability are scored as *bands*, not as "more is better", so plain
+sentences are never evidence on their own. And
+`data/corpus/must-not-flag/esl-engineer-email.txt` is in the corpus every new
+pattern must clear, so the reflect loop cannot learn a rule that flags
+competent non-native writing however many people cut the phrase.
+
+## The reflect loop: why recurrence is the gate
+
+The learning path (`scripts/learn.py`) takes its labels from the difference
+between what the skill returned and what the author published. Two design
+choices come from the failure modes above rather than from preference.
+
+*Learning must be two-directional.* A loop that only adds patterns can only
+grow, and a detector that only grows converges on flagging everything —
+which, per Liang, lands hardest on the writers already over-flagged. So a
+pattern the author overrules repeatedly loses weight.
+
+*A single edit is not evidence.* Authors cut for length, fix facts, and change
+their minds, and none of that is a style signal. Requiring the same span across
+three independent documents is a crude significance test: it separates a
+construction that recurs across unrelated writing from one person's sentence.
+The same property is what makes an upstream contribution shareable without
+carrying private text.
+
+The honest limit: this measures what *writers who use the skill* strike, not
+what readers detect. It tracks the register those authors are editing away
+from, which is the target, but it is a convenience sample and no substitute for
+the frequency work in `calibrate.py` against a real corpus.
+
+## Practitioner corroboration: Kagi SlopStop
+
+Kagi ships a production system for the same problem at web scale, and it
+converged on three of the design decisions here independently, which is the
+closest thing to external validation this architecture has.
+
+**Corroboration before classification.** A domain is downranked when it is *mostly*
+AI-generated, typically above 80% of its pages, rather than on a single hit.
+That is the same principle as requiring several signals to agree, arrived at from
+ranking rather than from linting.
+
+**Multiple reports accelerate review.** Kagi's community reporting treats
+repeated independent flags on one domain as stronger evidence than one flag.
+The reflect loop's three-document threshold is the same idea at document scale,
+and for the same reason: one reporter's judgment is taste, several unrelated
+reporters agreeing is signal.
+
+**An appeals path is a first-class feature, not an afterthought.** Users can
+report content as *not* slop, which triggers re-review and restores ranking.
+A system that can only add suspicion converges on suspecting everything. This
+is what `learn.py --demote` exists for, and Kagi treating it as core rather
+than optional is the argument for building it before you think you need it.
+
+**Downrank, never remove.** Flagged results stay visible and simply rank lower.
+The analogue here is hard rule 2 — flag hollow spans, do not fill them — and
+the reason both hold is that a false positive under a removal policy is
+unrecoverable, while a false positive under a flagging policy is an annoyance.
+
+Source: <https://help.kagi.com/kagi/features/slopstop.html>
+
+## Platform enforcement (August 2026)
+
+The target moved from reader disapproval to ranking. LinkedIn added a "Seems
+like AI slop" report control; reporting hides the post for that reader and
+trains LinkedIn's classifiers, flagged posts lose algorithmic reach beyond the
+author's own network, and repeat authors are notified privately in analytics.
+LinkedIn retired its "enhance your post" generator for a proofreader. Snapchat
+made wholly AI-generated video ineligible for Spotlight recommendation, ranking
+human-made above synthetic even with a disclosure label
+(Forbes, 1 August 2026).
+
+Two consequences for this skill's design.
+
+**The adversary is a reader-report classifier, not a word list.** LinkedIn's
+system is trained on what readers flag. That is a moving target defined by
+perception, which is an argument for the reflect loop — learning from what
+writers actually strike tracks the same perception — and against betting the
+detector on any fixed lexicon.
+
+**Reach is the cost of a false negative.** Before this, shipping a sloppy draft
+cost credibility with the people who noticed. Now it costs distribution to
+everyone. That raises the value of catching the structural tells, which is where
+a lexicon-only tool fails: the discrimination corpus contains a post scoring
+38.6 with zero weighted tells, caught on rhythm and shape alone.
+
+It also raises the cost of *over*-correction, since a rewrite that strips a
+writer's voice to pass a meter is a worse outcome than the tell it removed.
+That trade is why `references/overcorrection.md` exists and why the gate reports
+what it did not measure.
+
+## The host-model probe: token predictability without a shipped model
+
+The strongest signal in the table (feature 1) is the one a lexical linter cannot
+reach: whether a *model* finds the text predictable. DetectGPT and Binoculars read
+it from token log-probabilities. Zero Slop cannot — it ships no model, and it runs
+across harnesses where the model is Claude, whose API exposes no logprobs at all. So
+the predictability channel computes the same thing by generation instead of
+probability: it masks a spread of content words and asks the model *already running
+the skill* to guess each from context alone, then scores how often the guess lands on
+the word the author used (`scripts/predictability.py`). Machine text is easy to guess;
+human word choice is not.
+
+This is the DetectGPT insight — perturb, then ask the model how expected the original
+was — reduced to a cloze the host model can answer with generation alone. It works in
+any harness because it needs no logprobs and no bundled model; the scaffold (probe
+selection, scoring) is deterministic and offline, and only the guessing step needs the
+model, which the agent already is. A live check separates the extremes as expected: an
+AI-slop paragraph scored 66.7 predictability against 41.7 for a human bug-report of the
+same length. It is reported beside the surface score, never fused into it, so the
+0–100 score stays traceable to spans and this stays a second, independent opinion.
+
+The honest limit is calibration: the band cut-offs are set from small samples, not the
+tens of thousands of labelled documents the detection papers use, so read it as a
+corroborating signal, not a verdict — the same discipline the stylometric channels
+below are still waiting on.
+
+## The roadmap channels, and why they are not shipped yet
+
+Current work on interpretable detection converges on a small set of
+content-independent stylometric features, which is the direction this scorer
+should grow rather than adding more regex patterns.
+
+- **Function-word bigram frequency and average sentence length** are the two
+  highest-importance features in a lightweight interpretable detector, and the
+  paper reports 95–97% accuracy from them alone (NEULIF, arXiv:2511.21744).
+- A systematic cross-domain analysis finds the same family — function words,
+  punctuation, linguistic diversity — carries most of the interpretable signal
+  (arXiv:2606.04177).
+- These are attractive here because they are blind to specific wording, so
+  unlike the pattern meter they cannot be defeated by swapping synonyms, and
+  they degrade gracefully as models change.
+
+They are **not shipped**, and the reason is a measured negative result rather
+than a plan. Computing normalised sentence-opener entropy and function-word
+bigram entropy over the labelled discrimination corpus separates human from
+slop by essentially nothing (0.976 vs 0.986 opener entropy on twelve samples),
+with slop scoring marginally *higher*. That is not evidence the features fail;
+it is evidence the corpus is far too small to calibrate them — the cited work
+derives its thresholds from tens of thousands of documents. Adding an
+uncalibrated channel that shows no signal on the only data available would be
+exactly the hand-tuned rule this project tries to avoid.
+
+So the dependency is explicit: the labelled corpus (RAID arXiv:2405.07940, HC3,
+M4, AuTextification) comes first, the stylometric channels are calibrated
+against it second, and only then do they join the meter. Until then the honest
+statement is that the scorer is a lexical-and-structural linter with a
+research-backed roadmap, not a stylometric classifier.
+
+## Why register repair beats detection (2025-26)
+
+Two current results reframe what a slop tool should even try to do.
+
+**Detection is losing an arms race.** Adversarial paraphrasing — an LLM
+rewriting AI text under the guidance of a detector — cuts detector true-positive
+rates by 87-99% and transfers across neural, watermark and zero-shot detectors
+(Adversarial Paraphrasing, arXiv:2506.07001; StealthRL, arXiv:2602.08934).
+Anything that competes on catching a machine is chasing a target the literature
+is actively defeating. Zero Slop does not compete there: it repairs the register
+rather than classifying authorship, which is the half that stays useful when
+detection fails.
+
+**Slop is a training-supply pollutant, not just an aesthetic one.** Model
+collapse work shows machine text re-entering training corpora degrades the next
+model, with as little as 1% synthetic data measurably harming quality
+(arXiv:2603.11784, arXiv:2510.16657). That makes de-slopping a contribution to
+the commons, not only to one post's reach — the cleaner the human writing that
+survives, the less the collapse.
+
+Together these are the argument for the product's shape: measure and rewrite the
+author's own draft, personalise to the author's own voice, and never optimise
+against a detector, because the detector is both defeatable and beside the point.

package/references/overcorrection.md

@@ -0,0 +1,77 @@
+# Over-correction — the second failure mode
+
+The classic humanizer failure is swapping AI-slop for a louder slop. Readers
+clock both. Everything here is a rewrite *output* ban: never introduce these
+into text that didn't have them.
+
+## The edgy-slop catalogue
+
+- **Forced contrarianism** — "Everyone says X. They're wrong." (unless the
+  source argued it)
+- **Fake first person** — "I've seen this a hundred times", "In my
+  experience…" injected into authorless prose. Manufactured war stories are
+  fabrication, the cardinal sin.
+- **Performed candor** — "Let's be real", "Here's the thing", "I'll be
+  honest": candor is shown, not announced.
+- **Staccato drama** — "This matters. A lot. More than you think." Broetry
+  fragmentation is the LinkedIn variant.
+- **Em-dash theatrics** — dashes manufacturing emphasis the content didn't
+  earn. (Yes, humanizers add these; yes, it reads as AI.)
+- **Binary-contrast reveals** — "The answer isn't more tools. It's
+  discipline." One per piece max; injecting them is over-correction.
+- **Manufactured stakes** — "In a world where…", "Now more than ever".
+- **Intensifier padding as personality** — "genuinely", "honestly",
+  "literally" sprinkled for flavor.
+- **Slang costume** — forced colloquialisms a professional author wouldn't
+  use ("chef's kiss", "hits different") unless the voice sample has them.
+- **Manufactured informality** — forced lowercase, stray "lol", conspicuous
+  swearing, or broken grammar added to look human. Preserve these when they are
+  already part of the writer's voice; never inject them as camouflage.
+- **Fake errors** — never inject typos or grammar mistakes to fool
+  detectors. That's adversarial evasion, not writing, and it degrades the
+  text.
+- **Performed-writer prose** — theatrical framing of ordinary work ("we
+  hired an adversary"), epigram closers, staccato antithesis ("Not perfect.
+  Honest."), extended conceits (billing, courtroom, forensics, recipe),
+  hyperbole ("nothing on earth"), slang-cute idioms ("has receipts"), and
+  cute meta-taglines. The detection-side rows live in `tells.md` §3;
+  injecting them is the same costume-swap.
+
+The bar is a *thinking* author, not a *loud* one.
+
+## What NOT to flag (false-positive guard)
+
+From Wikipedia's "ineffective indicators" plus detector-calibration
+experience — these alone are NOT evidence of AI:
+
+- Perfect grammar and spelling
+- Formal or technical register where the genre demands it
+- A transition word, an em-dash, a "however" in isolation
+- Long sentences that earn their length
+- Rule-of-three used once, deliberately, for rhythm
+- Domain jargon used correctly for a domain audience
+- Calibrated hedging in research/medical/legal writing
+- Text merely being unsourced (check it, don't flag it)
+
+Require corroboration. A paragraph needs multiple independent tells, or a failed
+removal test, before it's slop.
+
+## Signs of human writing — preserve on sight
+
+When a draft shows these, protect them through the rewrite; deleting them is
+damage:
+
+- A claim someone could disagree with, stated without cover
+- The specific odd fact ($1.1M, 4,000 users, "episode 142")
+- Selective hedging at the edge of the author's knowledge
+- Humor, irritation, dry asides, self-interruption
+- Digressions that carry personality; asymmetric structure
+- Insider references assumed, not explained
+- The author's pet phrases and punctuation habits (voice sample rules)
+- Mistakes of passion — a run-on in an excited passage. Leave it.
+
+## Idempotence check
+
+Run the finished rewrite through the scorer and this file once more. If your
+rewrite added any catalogue item above, you traded costumes. Prefer the
+smaller edit: the best de-slop is usually deletion of the hedge plus nothing.