# AI Writing Pattern Reference for Medical/Radiology Manuscripts

Detailed reference for the 27 AI writing patterns, with expanded examples and suggested
rewrites specifically tailored for medical imaging and radiology research. Patterns 1-18 are
the general set; 19-21 are senior-MA-reviewer red flags; 25-27 are style/structure tells that
apply to any prose — typographic (25), rhythmic (26), and syntactic (27); 22-24 are
response-to-reviewers (R2R) letter patterns.

Sources (grounding differs by pattern — treat accordingly):
- Patterns 1-18: matsuikentaro1/humanizer_academic (English 18 patterns) + Wikipedia,
  "Signs of AI writing". Inherited lists — the thresholds they carry (em-dash per 1000 words,
  overall density) are conventional, not measured on a medical-manuscript corpus.
- Patterns 19-21, 22-24: observed in this user's own circulation and rebuttal rounds
  (senior-MA-reviewer comments; machine-drafted response letters).
- Pattern 25: observed in a co-author's track-changed edits of AI-drafted prose.
- Patterns 26-27: a prose-rhythm (26) and a prose-structure (27) tell, found when a
  native-fluent reader flagged AI-drafted argumentative prose that had already cleared the
  lexical and typographic sweeps. Pattern 27's antithesis rewrite test (M2) is adapted from
  the SNL-UCSB paper-writing skill's `gate_mechanical.md` (MIT-licensed).
- Adapted for radiology/medical imaging context

---

## Content Patterns

### Pattern 1: Significance Inflation

AI inflates importance with vague claims about "broader impact" instead of stating specific
clinical or scientific significance.

**Watch words:** pivotal, evolving landscape, underscores, highlights its importance, setting
the stage for, deeply rooted, focal point, indelible mark, paradigm shift, unprecedented

| # | BAD (AI-generated) | GOOD (Human-written) |
|---|-------------------|---------------------|
| 1 | "CT plays a pivotal role in the evolving landscape of oncologic imaging" | "CT is the primary imaging modality for cancer staging" |
| 2 | "This underscores the critical importance of early detection in improving patient outcomes" | "Early detection reduced mortality by 20% in the NLST trial" |
| 3 | "AI represents a paradigm shift in radiology practice" | "AI-assisted detection increased sensitivity from 0.78 to 0.91" |
| 4 | "Deep learning has made an indelible mark on medical image analysis" | "Deep learning models have achieved diagnostic accuracy comparable to radiologists in several tasks" |
| 5 | "Setting the stage for transformative advances in precision radiology" | "These methods may improve lesion detection in low-dose protocols" |

**Radiology-specific note:** Radiology papers are particularly susceptible to this pattern
in introductions discussing AI/deep learning. State the specific performance metric or
clinical outcome instead.

---

### Pattern 2: Notability Claims

AI labels studies, institutions, or researchers with unearned superlatives.

**Watch words:** landmark, renowned, prestigious, groundbreaking, impressive, seminal,
pioneering, state-of-the-art

| # | BAD | GOOD |
|---|-----|------|
| 1 | "This groundbreaking study from a prestigious institution" | "In a multicenter study of 12,000 patients (Smith et al., 2024)" |
| 2 | "The landmark NLST trial" | "The National Lung Screening Trial (NLST)" |
| 3 | "Using a state-of-the-art deep learning architecture" | "Using a ResNet-50 model pretrained on ImageNet" |
| 4 | "These impressive results demonstrate" | "The AUC was 0.94 (95% CI: 0.91-0.97)" |

**Radiology-specific note:** "State-of-the-art" is especially overused in AI radiology papers.
Name the specific architecture and training approach instead.

---

### Pattern 3: Superficial -ing Analyses

AI appends "-ing" participial phrases to sentences, creating fake analytical depth without
adding information.

**Watch words (at end of sentence):** highlighting, underscoring, emphasizing, showcasing,
fostering, reflecting, contributing to, demonstrating, suggesting, indicating

| # | BAD | GOOD |
|---|-----|------|
| 1 | "The AUC was 0.93 (95% CI: 0.90-0.96), highlighting the strong diagnostic performance of the model" | "The AUC was 0.93 (95% CI: 0.90-0.96)." |
| 2 | "Sensitivity improved from 78% to 91%, underscoring the value of AI-assisted detection" | "Sensitivity improved from 78% to 91%. This improvement corresponded to detection of 13 additional malignant nodules per 1000 screens." |
| 3 | "Inter-reader agreement was excellent (ICC = 0.92), demonstrating the reproducibility of the measurement technique" | "Inter-reader agreement was excellent (ICC = 0.92; 95% CI: 0.88-0.95)." |
| 4 | "Processing time decreased by 40%, showcasing the efficiency gains" | "Processing time decreased by 40% (from 5.2 to 3.1 minutes per case)." |
| 5 | "The false-positive rate decreased to 3.2%, reflecting improved specificity" | "The false-positive rate decreased to 3.2%." |

**Fix strategy:** End the sentence at the data. If interpretation is needed, start a new
sentence with a specific claim supported by the numbers.

---

### Pattern 4: Promotional Language

AI uses adjectives and adverbs that promote rather than describe.

**Watch words:** remarkable, dramatic, stunning, profound, extraordinary, exceptional,
breathtaking, robust (unless describing a statistical method), comprehensive, innovative,
novel (overused)

| # | BAD | GOOD |
|---|-----|------|
| 1 | "Our model achieved remarkable diagnostic accuracy" | "Our model achieved an AUC of 0.95" |
| 2 | "A dramatic reduction in false-positive rates was observed" | "The false-positive rate decreased from 12.3% to 4.1%" |
| 3 | "This comprehensive framework provides a robust solution" | "This framework reduced processing time by 60% while maintaining accuracy" |
| 4 | "The novel deep learning approach demonstrated exceptional performance" | "The proposed approach outperformed the baseline by 0.08 AUC points (p = 0.003)" |
| 5 | "A robust and comprehensive evaluation was conducted" | "We evaluated the model on three external datasets" |

**Radiology-specific note:** "Robust" is acceptable when describing a statistical method
(e.g., "robust standard errors") but not as a general-purpose adjective for frameworks,
pipelines, or results.

---

### Pattern 5: Vague Attributions

AI attributes claims to unnamed sources instead of citing specific studies.

**Watch words:** Studies have shown, Experts argue, Some researchers, It is widely accepted,
Several publications, The literature suggests, It has been reported, Research indicates

| # | BAD | GOOD |
|---|-----|------|
| 1 | "Studies have shown that AI can improve diagnostic accuracy" | "In a meta-analysis of 82 studies, AI systems achieved pooled sensitivity of 87% (Liu et al., 2019)" |
| 2 | "It is widely accepted that MRI is superior for soft tissue contrast" | "MRI provides superior soft tissue contrast resolution compared with CT (reference)" |
| 3 | "Several studies have demonstrated the utility of radiomics" | "Aerts et al. (2014) demonstrated that radiomic features predicted outcomes in lung and head-and-neck cancer" |
| 4 | "Research indicates that prompt engineering affects LLM output quality" | "Prompt structure affected diagnostic accuracy by up to 15 percentage points in GPT-4V evaluations (Wu et al., 2024)" |
| 5 | "It has been reported that CAD systems reduce reading time" | "Park et al. (2023) reported a 25% reduction in reading time with AI-assisted detection" |

**Fix strategy:** Replace every vague attribution with a specific citation. If you do not
know the specific reference, flag it for the user to fill in with `[CITE NEEDED]`.

---

### Pattern 6: Formulaic Challenges Sections

AI produces template limitation/future-work sections that could apply to any paper.

**Watch words:** Despite these limitations, Future outlook, Continues to provide valuable
insights, Further research is warranted, More studies are needed

| # | BAD | GOOD |
|---|-----|------|
| 1 | "Despite these limitations, our study provides valuable insights into AI-assisted diagnosis" | "The single-center design limits generalizability to community practice settings where case mix and image quality differ" |
| 2 | "Further research is warranted to validate these findings" | "External validation on a multi-institutional dataset with variable scanner protocols is needed before clinical deployment" |
| 3 | "More studies are needed to fully understand the potential of this approach" | "Prospective evaluation comparing AI-assisted and conventional reading in a screening population would quantify the clinical impact" |
| 4 | "Despite challenges, the future of AI in radiology is promising" | "Integration into the clinical PACS workflow remains an engineering challenge requiring vendor collaboration" |

**Fix strategy:** State the specific limitation, its consequence, and the specific study
design that would address it.

---

## Language Patterns

### Pattern 7: AI Vocabulary Words

Words whose frequency increased markedly in post-2023 AI-generated text. Their presence
at high density signals AI authorship.

**High-signal words to eliminate or replace:**

| AI Word | Replacement |
|---------|-------------|
| Additionally | (delete, or use "In addition," sparingly) |
| Furthermore | (delete, or restructure the sentence) |
| Moreover | (delete) |
| Crucial | Important (or delete) |
| Delve | Examine, analyze, investigate |
| Enhance | Improve |
| Fostering | Promoting, supporting (or delete) |
| Garner | Receive, attract |
| Highlight (verb) | Show, demonstrate (or delete) |
| Interplay | Interaction, relationship |
| Intricate | Complex (or delete if unnecessary) |
| Key (adjective, overused) | Important, main, primary |
| Landscape (abstract) | Field, domain (or delete) |
| Leverage | Use, apply |
| Multifaceted | (delete; describe the specific facets) |
| Pivotal | Important, central |
| Showcase | Show, demonstrate |
| Tapestry | (delete entirely) |
| Testament | Evidence, indication |
| Underscore (verb) | (delete; state the point directly) |
| Utilize | Use |
| Valuable | Useful (or delete) |

**Radiology examples:**

| BAD | GOOD |
|-----|------|
| "Additionally, we utilized a novel architecture to enhance detection" | "We used a ResNet-50 to improve detection" |
| "This crucial finding underscores the pivotal role of AI" | "This finding supports the use of AI" |
| "We delved into the intricate interplay between image quality and model performance" | "We examined how image quality affected model performance" |

**Rule:** If more than 3 of these words appear in a single page, the section needs revision.

---

### Pattern 8: Copula Avoidance (Avoiding "is")

AI substitutes elaborate verb constructions for simple "is/are."

**Watch words:** serves as, stands as, marks, represents [a], boasts, features, offers [a],
constitutes, functions as

| # | BAD | GOOD |
|---|-----|------|
| 1 | "CT serves as the primary imaging modality for lung cancer screening" | "CT is the primary imaging modality for lung cancer screening" |
| 2 | "This metric represents a significant improvement" | "This metric is a significant improvement" |
| 3 | "The dataset features 12,000 annotated images" | "The dataset contains 12,000 annotated images" |
| 4 | "DWI stands as the most sensitive sequence for acute stroke detection" | "DWI is the most sensitive sequence for acute stroke detection" |
| 5 | "This architecture constitutes a major advance" | "This architecture is a major advance" |

---

### Pattern 9: Negative Parallelisms

AI overuses "Not only X but also Y" constructions.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "The model not only improved sensitivity but also reduced false positives" | "The model improved sensitivity and reduced false positives" |
| 2 | "This approach not only streamlines workflow but also enhances diagnostic confidence" | "This approach reduces reading time by 30% and improves diagnostic confidence" |
| 3 | "AI not only assists in detection but also aids in characterization" | "AI assists in both detection and characterization" |

**Threshold:** More than 1 per section is a red flag.

---

### Pattern 10: Rule of Three Overuse

AI forces ideas into groups of exactly three.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "accuracy, efficiency, and reproducibility" (repeated across 4 paragraphs) | Vary: sometimes mention two, sometimes four, as the content demands |
| 2 | "detection, segmentation, and classification" | "detection and segmentation" (if classification is not relevant to the point) |
| 3 | "clinical, technical, and educational implications" | Group naturally: state the specific implications rather than labeling categories |
| 4 | "sensitivity, specificity, and accuracy" repeated 5 times | Report each metric where relevant; do not always bundle all three |

**Detection method:** Search for the pattern `X, Y, and Z` -- if it appears more than 3 times
in a section, the author (or AI) is forcing triples.

---

### Pattern 11: Synonym Cycling (Elegant Variation)

AI avoids repeating the same word by cycling through synonyms, which creates inconsistency
in medical writing where precision matters.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "patients... participants... subjects... individuals" | "patients" (consistent throughout, matching IRB language) |
| 2 | "lesions... abnormalities... findings... pathology" | "lesions" (if referring to the same thing) |
| 3 | "radiologists... readers... interpreters... physicians" | "radiologists" or "readers" (pick one per context) |
| 4 | "images... scans... examinations... studies" | "examinations" for the procedure, "images" for the pictures |
| 5 | "model... algorithm... system... tool... framework" | "model" (if it is a model) |

**Radiology-specific note:** Inconsistent terminology is a real problem in radiology
manuscripts. "Lesion" and "finding" have different meanings; "examination" and "image"
are not interchangeable. Pick the most precise term and use it consistently.

---

### Pattern 12: False Ranges

AI uses "from X to Y" where X and Y are not on a meaningful continuum.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "from improved detection to enhanced workflow efficiency" | "improved detection and workflow efficiency" |
| 2 | "ranging from data augmentation to transfer learning" | "including data augmentation and transfer learning" |
| 3 | "from clinical practice to research applications" | "in clinical practice and research" |

---

## Style Patterns

### Pattern 13: Em Dash Overuse

AI uses em dashes far more frequently than human academic writers.

**Threshold:** More than 2 em dashes per 1000 words.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "The model -- trained on 50,000 images -- achieved an AUC of 0.94" | "The model, trained on 50,000 images, achieved an AUC of 0.94" |
| 2 | "Three features -- size, shape, and margin -- were selected" | "Three features (size, shape, and margin) were selected" |
| 3 | "CT -- the most widely used modality -- remains essential" | "CT, the most widely used modality, remains essential" |

---

### Pattern 14: Title Case in Headings

AI capitalizes all major words in section headings. Most medical journals use sentence case.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "Statistical Analysis And Primary Endpoints" | "Statistical analysis and primary endpoints" |
| 2 | "Deep Learning Model Architecture" | "Deep learning model architecture" |
| 3 | "Inter-Reader Agreement Assessment" | "Inter-reader agreement assessment" |

**Note:** Follow the target journal's style guide. Most radiology journals (Radiology, AJR,
European Radiology) use sentence case for section headings.

---

### Pattern 15: Curly Quotation Marks

ChatGPT and similar tools produce curly (smart) quotes. Many journal submission systems
and LaTeX workflows expect straight quotes.

| BAD | GOOD |
|-----|------|
| \u201csensitivity\u201d | "sensitivity" |
| \u2018specificity\u2019 | 'specificity' |

**Fix:** Find-and-replace all curly quotes with straight quotes before submission.

---

## Filler and Hedging Patterns

### Pattern 16: Filler Phrases

Empty phrases that add words without adding meaning.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "It is important to note that the AUC exceeded 0.90" | "The AUC exceeded 0.90" |
| 2 | "It is worth noting that sensitivity decreased in smaller lesions" | "Sensitivity decreased in smaller lesions" |
| 3 | "In order to evaluate diagnostic performance" | "To evaluate diagnostic performance" |
| 4 | "Due to the fact that the dataset was imbalanced" | "Because the dataset was imbalanced" |
| 5 | "At the present time, no consensus exists" | "No consensus exists" |
| 6 | "With respect to image quality" | "For image quality" |
| 7 | "The model has the ability to detect" | "The model can detect" |
| 8 | "In the context of emergency radiology" | "In emergency radiology" |

**Detection method:** Search for "It is" at sentence start, "In order to," "Due to the fact,"
"With respect to," "has the ability to," "In the context of."

---

### Pattern 17: Excessive Hedging

AI stacks multiple hedging words, making claims weaker than the evidence supports.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "may potentially suggest the possibility of improved outcomes" | "improved outcomes" (if the data clearly shows it) or "suggests improved outcomes" (if uncertain) |
| 2 | "It could be argued that this might have some impact" | "This may affect diagnostic accuracy" |
| 3 | "These findings seem to indicate that AI may perhaps assist" | "These findings indicate that AI assists" (if the data supports it) |

**Calibration guide for radiology manuscripts:**
- Strong evidence (p < 0.001, large effect, prospective design): State as fact. "AI improved detection."
- Moderate evidence (p < 0.05, moderate effect, retrospective): Single hedge. "AI may improve detection."
- Weak evidence (trend, small sample, pilot, single-center): Measured hedge. "These preliminary results suggest AI may improve detection."
- Exploratory (secondary analysis, post hoc): Explicit qualifier. "In exploratory analysis, AI-assisted reading was associated with higher sensitivity."

---

### Pattern 18: Generic Positive Conclusions

AI ends papers with content-free optimistic statements.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "The future looks bright for AI-assisted radiology" | (delete) |
| 2 | "This paves the way for transformative advances in diagnostic imaging" | "Prospective validation in a screening population is the next step before clinical implementation" |
| 3 | "AI continues to reshape the landscape of medical imaging" | "AI-assisted detection reduced missed cancers by 12% in this retrospective analysis" |
| 4 | "These findings open new avenues for research" | "Future work should evaluate the model on external datasets with heterogeneous scanner protocols" |
| 5 | "This work lays the foundation for future innovations" | "The trained model and evaluation code are publicly available at [URL]" |

**Fix strategy:** Replace with the single most specific implication or the concrete next step.

---

## Senior MA Reviewer and Typographic Patterns

### Pattern 19: § (Section-Sign) Marker

AI models frequently use the `§` symbol to point to sections in the body text. A senior MA reviewer recognizes this immediately as an "AI pattern" — a LaTeX-style notation that is almost never used in Korean or US medical-journal manuscripts.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "as described in §2.3" | "as described in the Methods" |
| 2 | "(see §Discussion)" | "(see Discussion)" |
| 3 | "§Results" | "Results" |

**Detection:** `grep -c "§" manuscript.md` → must be 0.
**Fix strategy:** Delete every § or replace it with the section name (`Methods`, `Results`, `Discussion`).

---

### Pattern 20: Methods/Results Self-Reference Parenthetical

Self-reference parentheticals of the form `(Methods §X)`, `(Results §Y)`, `(Methods, Section 2.3)`. Frequently co-occurs with Pattern 19 (§). A reviewer judges that there is no need to cross-reference one's own section.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "We applied bootstrap resampling (Methods §2.3) to..." | "We applied bootstrap resampling (described in Methods) to..." or simply delete |
| 2 | "The pooled estimate (Results §3.1) was..." | "The pooled estimate was..." |
| 3 | "As shown in Table 1 (Methods)..." | "As shown in Table 1..." |

**Detection:** `grep -inE "\((Methods|Results|Discussion|Introduction)\s*§" manuscript.md` → 0 lines.
**Fix strategy:** Delete the parenthetical when the flow is self-evident. When an explicit pointer is needed, shorten to "(Methods)" or "(see Methods)".

---

### Pattern 21: AI Disclosure Boilerplate (Body)

Boilerplate paragraphs such as "Artificial Intelligence Disclosure" / "Generative AI was not used to create..." / "AI Acknowledgement" appearing in the manuscript body. These are needed on submission forms and cover letters, but placing them in the body reads to a reviewer as a declaration that "this was written by AI" — an AI-generated signal.

| # | BAD | GOOD |
|---|-----|------|
| 1 | An "## Artificial Intelligence Disclosure" paragraph at the end of the body | Delete the whole paragraph → state it only in the cover letter or submission form |
| 2 | "Generative AI was not used to create, modify, or alter any images, figures, or tables in this manuscript." (in body Methods) | (delete) — the `~/.claude/rules/journal-ai-image-policies.md` boilerplate belongs only in the cover letter |
| 3 | "We acknowledge the use of ChatGPT for language editing" in the Acknowledgments section | Move to the cover letter per journal policy, or reduce to "Language editing was performed" |

**Detection:** `grep -inE "artificial intelligence disclosure|generative ai was not used|ai acknowledg(e)?ment" manuscript.md` → 0 lines (body).
**Fix strategy:** Remove from the body → keep only on the submission form / cover letter. Exception only when the journal requires an in-body statement.

---

### Pattern 25: Inline-Emphasis Over-Use (Typographic Over-Signposting)

LLM-drafted prose italicises words the sentence already stresses. A human copy-editor strips
almost all of it — which is how this pattern was found, in a co-author's track-changed edits of
an AI-drafted section. Unlike Patterns 19-21 this is not senior-MA-specific; it applies to any
prose, including response letters.

**Watch for:** single-word italics (*into*, *passive*, *same*, *not*), whole-clause italics
(*a redesign of the relationship itself*), bold used mid-paragraph to signpost a phrase.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "The model was trained *on* the derivation cohort, not *within* it." | "The model was trained on the derivation cohort, not within it." |
| 2 | "This is *a fundamental shift in how the reference standard is constructed*." | "This changes how the reference standard is constructed." |
| 3 | "Sensitivity improved, but **only in the subgroup with prior imaging**." | "Sensitivity improved only in the subgroup with prior imaging." |

**Legitimate italics — never strip these:** statistical symbols (*P*, *t*, *n*, *F*, *r*),
Latin (*in vivo*, *et al.*, *post hoc*), gene and species names (*BRCA1*, *E. coli*), journal
titles. A **bold run-in subheading** at line start is correct Nature/npj style
(`manuscript-style-classical.md` §1.2) and is not counted.

**Detection:** `/self-review` `scripts/check_emphasis_density.py --manuscript manuscript.md`
→ `EMPHASIS_OVERUSE` (Minor). Counts non-allowlisted *italic* spans per 1,000 body words;
bold is deliberately excluded so the gate does not fight the run-in-subheading rule. Fires
only when density and raw count both clear a floor, so one stray italic in a short note is
never flagged. A whole-clause italic span carries an escalation note — it is the strongest tell.

**Fix strategy:** Delete the emphasis and let word order carry the stress. If the sentence
genuinely needs the contrast, rewrite it so the contrast is structural rather than typographic.

### Pattern 26: Aphorism Density (Every Sentence Lands, None Explains)

LLM-drafted argumentative prose writes in epigrams. Each sentence is built to be quotable,
and the sentences that merely carry information forward — the ones a human writer puts
between the good lines — are missing. Two co-occurring signals make this measurable: the
**negative definition** (`X is not Y` as a whole short sentence) and a high share of **very
short declaratives** used as punchlines. Found when an external reader said a heavily
revised draft was "clearer, but still reads like AI", and the cause turned out to be rhythm
rather than grammar.

**Watch for:** "Authority is not cognition." "Involvement is not independence." "It did
not." "It is not one." "It cost more than it paid." Any paragraph whose every sentence
could be pulled out and put on a slide.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "Agreement is not confirmation. The circle does not merely overlap. It is dragged inward." | "Agreement here is not confirmation, because the tool is working from the reader's own framing; rather than overlapping the reader's error, its error is pulled toward it." |
| 2 | "That answer is only half-built. It is not sufficient." | "That answer is only half-built, and the missing half is the part the safety argument actually rests on." |
| 3 | "Involvement is not independence. Partial decorrelation is not independence either." | "Involvement is not independence, and partial decorrelation does not close the gap either, since the bound assumes the stronger property." |

**Calibration.** Across eight published *npj Digital Medicine* Perspectives, cleaned to body
prose, the negative-definition rate ran 0.00–0.45% of sentences (seven of eight were 0.00%)
and the short-declarative share 0.94–10.47%. A manuscript at 2% and 24% is outside that
range on both axes.

**Detection:** `/self-review` `scripts/check_aphorism_density.py --manuscript manuscript.md`
→ `APHORISM_DENSITY` (Minor). Fires only when BOTH rates clear thresholds set above the
observed published maxima, so ordinary emphatic prose does not trip it. Use `--calibrate`
to regenerate thresholds on your own corpus; the shipped numbers come from argumentative
Perspectives and another genre may sit elsewhere.

**Fix strategy:** Do not delete the epigrams — two or three are what a reader remembers.
Absorb most of them into the neighbouring sentence and write back the explanatory sentence
that was compressed out. Deleting them instead of absorbing them shortens the prose further
and makes the problem worse.

### Pattern 27: Antithesis-and-Cleft Density (Structural AI Tells)

LLM-drafted argumentative prose over-builds two marked constructions that no per-instance
rule can flag, because each single occurrence is grammatical and often functional:

- **antithesis parallelism** — "authority *rather than* cognition", "*not* a check *but* a
  second road", "involvement, *not* independence".
- **cleft / pseudo-cleft** — sentence-initial "*What* matters *is* …", "*It is* X *that* …".

One of each is legitimate; a run of them is an epigram machine. Found when a native-fluent
co-author's comprehension review flagged prose that had already cleared the em-dash, passive
and vocabulary sweeps (Patterns 7, 13) — the residue was structural, not lexical (one draft
carried 28 "rather than" and roughly ten sentence-initial clefts). Unlike Pattern 26
(whole-sentence negative definitions), this counts *within-sentence* antithesis markers and
fronting constructions, and is a density measure — never a single instance.

**Watch for:** "X rather than Y" (the dominant tic), "not X but Y", "X, not Y", "What X is
Y", "It is X that Y". Do NOT touch "instead of" — it is the functional, non-decorative
sibling of "rather than" and its presence is not a tell.

| # | BAD (antithesis / cleft density) | GOOD (plain construction) |
|---|---|---|
| 1 | "The model earns trust by what it measures rather than by what it claims. What matters is the validation set, not the training curve. It is the external cohort, rather than the internal split, that carries the evidence." | "The model earns trust from what it measures. The validation set matters more than the training curve, and the external cohort carries the evidence the internal split cannot." |
| 2 | "This is calibration, not discrimination. What the reader needs is a probability, rather than a rank." | "This is calibration rather than discrimination; the reader needs a probability, and a rank does not supply one." |
| 3 | "It is the disagreement that informs, not the agreement." | "Disagreement between the two readers is what carries the information here." |

**Calibration.** Across this toolkit's three published-quality demo manuscripts, "rather
than" runs 1.4–3.8 per 1,000 body words and sentence-initial clefts are absent. The gate's
thresholds (antithesis 6.0 / 1,000, cleft 2.5 / 1,000, each behind a raw-count floor) sit
above that range, so a lone functional "rather than" or one pseudo-cleft never trips it.

**Detection:** `/self-review` `scripts/check_rhetorical_density.py --manuscript manuscript.md`
→ `ANTITHESIS_DENSITY` / `CLEFT_DENSITY` (both Minor, independent). Each fires only when the
per-1,000 rate AND the raw count both clear a floor, so functional use in ordinary prose does
not trip it.

**Fix strategy (M2 heuristic):** For each negative/antithesis construction, delete the
negative half and rewrite the clause in the positive. If a fact disappears the contrast was
functional — keep it; if nothing disappears it was decoration — cut it. Judge by the
manuscript's overall rate, not instance by instance, and keep two or three for emphasis.
Rewrite clefts in plain subject-verb order ("What matters is X" → "X matters"). The
negative-form test is adapted from the SNL-UCSB paper-writing skill's `gate_mechanical.md`
(MIT-licensed).


---

## Response-Letter Patterns (R2R)

These three patterns are specific to response-to-reviewers (R2R) letters and editor cover
letters. They rarely appear in manuscript bodies but dominate machine-drafted rebuttals,
where the model narrates the editing process instead of the science. Apply them whenever the
text under review is a response letter or cover letter. See the revise skill's
`references/r2r_voice.md` for full before/after skeletons. Examples below are synthetic
(a fictional deep-learning lung-nodule CT study).

**Related (triage, not a fixed-string pattern):** *defensive over-elaboration* — pre-emptive
cross-reviewer lobbying ("Reviewers 2 and 3 also accepted this"), defensive meta-comments ("we
confirm this is unchanged and not softened"), manufactured paragraphs answering a satisfied
reviewer, or a separate cover letter on an R2+ round when its content belongs in the
response-letter head — is a succinctness / round-discipline issue, not a regex pattern, and is
most common on R2+ rounds. The normative guidance lives in the revise skill (Step 5 +
Response-Letter Voice), its `references/r2r_voice.md`, and the `rebuttal-letter-style` rule;
humanize only cross-references it (no hard fail, no detection regex added here).

### Pattern 22: Editing-Mechanism / Change-Log Narration

The response prose narrates *how the text was edited* — what was added, where, how many phrases
were swapped, which pass produced it — instead of stating what changed and why. The reviewer
reads this as auto-generated and as checklist-clearing rather than scientific engagement.

**Scope — this targets editing-mechanism narration only (avoid over-flagging):** narrating a
*new analysis you ran* ("we performed a sensitivity analysis restricted to one eye per patient,
and the result held") is the science the reviewer asked for — never flag it. Likewise "we added
a sentence to the Methods: '...'", structured `Response:` / `Changes made:` blocks, and
`Original → Revised` before/after pairs are normal human conventions. The tell is the *editing
mechanism* layered on top, not the act of describing, quoting, or locating a change.

**Watch phrases (the tell):** version-prefixed edits ("v2 Methods adds one sentence"), "we
softened six phrases", "demoted the term at all N locations", "a grep-and-soften pass", "the v2
revision changes", bare "No further manuscript change" stubs, "reframed via the vocabulary cascade".

| # | BAD (editing-mechanism) | GOOD (substantive / science) |
|---|---|---|
| 1 | "v2 Methods adds one sentence: '...'. This is a short visible clarification rather than only in the Limitations." | "We agree the design is observational; we have added to the Methods: '...'" |
| 2 | "We softened six over-interpretive phrases in Results and Discussion." | "We have rephrased the over-interpretive passages; for example, '...' now reads '...'." |
| 3 | "No further manuscript change was applied." | "The existing Discussion text already addresses this; the relevant statement is '...'." |

**Detection (triage, not auto-fail):** `grep -inE "v[0-9]+ .*(add|demote|soften|reframe)|softened [0-9a-z]+ phrases|no further (manuscript )?change|grep-and-soften|vocabulary cascade|demoted .* at all [0-9]+" response_to_reviewers.md` — review each hit and flag only genuine editing-mechanism narration, never analysis narration or quoted additions.
**Fix strategy:** Delete the mechanism narration. State the substantive change and quote the new sentence; omit how it was found or made.

### Pattern 23: Internal Line-Number Reference Tone

Pointing reviewers to internal draft/markdown line numbers ("(line 43)", "at lines 43, 49,
58, 60, 77–79"). These never match the reviewer's view of the revised manuscript and read as
a diff log. Section names are what authors actually use.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "We clarified this at line 43." | "We clarified this in the Methods (Design subsection)." |
| 2 | "Single-sentence clarifications were added at lines 43, 49, 58, 60, and 77–79." | "We added short clarifications to the Methods covering the design, the intervention, and the outcome definitions." |
| 3 | "v2 line 127 retains this sentence and adds the adjustment after it." | "In the Results (Between-group comparison) we retained the original sentence and added the adjustment immediately after." |

**Not a tell:** a **revised-manuscript** page/line ("page 7, lines 177-178") when the letter
states once that all page/line numbers refer to the revised manuscript the reviewer is reading.
The tell is the *internal draft* line number that will not match the reviewer's PDF.

**Detection:** `grep -inE "\(line [0-9]+|at lines? [0-9]+|line [0-9]+(–|-)[0-9]+" response_to_reviewers.md` → review each hit; keep only those that demonstrably point to the revised manuscript.
**Fix strategy:** Replace internal/draft line numbers with a section name; keep revised-manuscript page/line only if it matches what the reviewer sees.

### Pattern 24: Tooling / Scaffolding Leak

Exposing internal tooling, verification mechanics, or draft scaffolding to the reviewer:
grep counts, internal FIX/category codes, references to "the circulated bundle" or "the
internal supplementary index", or `§` self-references carried into the response letter.

| # | BAD | GOOD |
|---|-----|------|
| 1 | "Final grep verification returned zero occurrences across the circulated bundle." | (delete — describe the corrected wording instead) |
| 2 | "Addressed via the FIX-1 vocabulary cascade." | "We replaced [old term] with [new term] throughout the manuscript." |
| 3 | "These strings remain only in the internal supplementary index, which is not part of the circulated bundle." | (delete — never reference internal scaffolding to a reviewer) |
| 4 | "as reframed in §Discussion" | "as reframed in the Discussion" |

**Detection (triage, except `§`):** `grep -inE "grep verification|grep-and-soften|circulated bundle|internal (supplementary )?index|FIX-[0-9]|vocabulary cascade|§" response_to_reviewers.md cover_letter.md` — only `§` is a hard 0 (always a tell). The other terms can rarely be legitimate (e.g., "a cascade detector", "we grep-checked our own data pipeline"), so confirm each hit is an internal-tooling reference before flagging.
**Fix strategy:** Strip confirmed references to internal tooling, verification passes, and draft scaffolding. The reviewer should see only the science and the substantive changes.

---

## Section-Specific Application Guide

### Abstract (ALL patterns)

The abstract is the most visible section and the most likely to be checked for AI writing.
Apply all applicable patterns (1-21 and 25) with zero tolerance.

**Common abstract issues:**
- Pattern 1 in the Background sentence.
- Pattern 4 in the Results sentence ("demonstrated remarkable performance").
- Pattern 18 in the Conclusion sentence ("paves the way").

### Introduction (Patterns 1, 2, 5, 7, 12)

- Opening paragraph: Check for significance inflation (Pattern 1) and AI vocabulary (Pattern 7).
- Literature review paragraph: Check for vague attributions (Pattern 5).
- Gap statement: Check for notability claims (Pattern 2).

### Methods (Patterns 8, 16)

- Methods should be the most straightforward section. Watch for copula avoidance ("serves as
  the reference standard" instead of "is the reference standard") and filler phrases.

### Results (Patterns 3, 4, 6, 10, 11)

- After every statistical result, check for appended -ing clauses (Pattern 3).
- Check for promotional adjectives before numbers (Pattern 4).
- Check for synonym cycling of key terms (Pattern 11).

### Discussion (Patterns 1, 5, 6, 17, 18)

- First paragraph: Check for significance inflation (Pattern 1).
- Comparison paragraphs: Check for vague attributions (Pattern 5).
- Limitations: Check for formulaic language (Pattern 6).
- Final paragraph: Check for generic conclusions (Pattern 18) and excessive hedging (Pattern 17).

### Conclusion (Patterns 1, 18)

- The conclusion should be 1-3 sentences stating the main finding and its specific implication.
- No significance inflation, no generic optimism.

### Response to Reviewers / Cover letter (Patterns 22, 23, 24 — plus 13, 16, 19)

- Response letters and cover letters are reviewer-facing argument, not change-logs. The
  dominant AI-tell here is the editing-mechanism register (Pattern 22), internal draft
  line-number pointers (Pattern 23), and tooling/scaffolding leaks (Pattern 24).
- Also sweep for em dashes (13), filler phrases (16), and `§` markers (19).
- Patterns 1-18 still apply where relevant, but 22-24 are the highest-yield checks for this
  document type.

---

## Quick Checklist (Pre-Submission)

Run this checklist on the final manuscript before submission:

- [ ] No "Additionally" / "Furthermore" / "Moreover" at sentence start (allow max 1 total)
- [ ] No "pivotal" / "crucial" / "landscape" / "delve" / "utilize" / "leverage"
- [ ] No "-ing" participial phrases appended to statistical results
- [ ] No "serves as" / "stands as" / "represents a" (use "is")
- [ ] No vague "Studies have shown" without a specific citation
- [ ] No "not only X but also Y" (allow max 1 total)
- [ ] No generic positive conclusions ("paves the way," "the future looks bright")
- [ ] Consistent terminology throughout (no synonym cycling)
- [ ] Em dashes: fewer than 2 per 1000 words
- [ ] Hedging calibrated to evidence level (no stacked hedges)
- [ ] Filler phrases eliminated ("It is important to note that," "In order to")
- [ ] Straight quotation marks (not curly)
- [ ] Sentence case in headings (unless journal requires title case)
- [ ] AI pattern density < 2.0 per 1000 words
- [ ] § (section sign): 0 occurrences (Pattern 19) — `grep -c "§"` = 0
- [ ] (Methods §X) / (Results §Y) self-reference: 0 occurrences (Pattern 20)
- [ ] AI Disclosure boilerplate in body: 0 occurrences (Pattern 21) — cover letter / submission form only
- [ ] Inline-emphasis over-use (Pattern 25) — `check_emphasis_density.py`; strip single-word and whole-clause italics, keep stat symbols / Latin / gene-species
- [ ] Aphorism density (Pattern 26) — `check_aphorism_density.py`; a run of negative definitions plus very short declaratives; absorb most into neighbouring sentences, keep two or three
- [ ] Antithesis / cleft density (Pattern 27) — `check_rhetorical_density.py`; "rather than" / "not X but Y" / "X, not Y" and sentence-initial "What … is …" / "It is … that …"; delete the negative half and rewrite positive, keep two or three
- [ ] Sentence-length variety (SKILL.md Fix rule 7) — `check_sentence_variety.py`; short (8-12 words) and long (25-35 words) sentences both present
- [ ] Rewrite footprint within bounds — `check_edit_footprint.py` when humanize rewrote a file; >30% changed warrants a re-read, >50% is a halt

### Response letters / cover letters only (Patterns 22-24)

- [ ] No editing-mechanism narration (Pattern 22) — "v2 adds one sentence", "softened N phrases", "No further manuscript change" (analysis narration and quoted additions are fine)
- [ ] No internal draft line-number pointers (Pattern 23) — "(line NN)", "at lines N, M" (revised-manuscript page/line is fine)
- [ ] No tooling/scaffolding leak (Pattern 24) — "grep", "FIX-N", "circulated bundle", "internal index", `§`

#!/usr/bin/env python3
"""Bound how much of a text a humanize rewrite is allowed to touch (humanize Phase 3).

De-AI editing is subtractive: strip the tells, keep the author's sentences. A model asked to
"make this sound human" will also rewrite paragraphs that had nothing wrong with them, and the
result reads fluently enough that the loss is invisible on review — the author's voice is gone
and nobody can point to the sentence where it went. Pattern-by-pattern fixes touch a small
fraction of the words; a wholesale rewrite touches most of them. That difference is measurable,
so this gate measures it instead of trusting the rewrite to have been restrained.

It also enforces the two invariants the humanize skill declares but never checked: every number
and every citation present before the rewrite must still be present after it.

Verdicts:
  NUMBER_DRIFT (Major)         a numeric token's count changed across the rewrite.
  CITATION_DROP (Major)        a citation present before is absent after.
  EDIT_FOOTPRINT_HIGH (Minor)  more than --warn-pct of the words changed — re-read the diff.

Why the footprint is advisory and the invariants are not: the two invariants are the skill's
own declared contract ("every number, statistic, p-value, confidence interval and clinical fact
must remain identical"; "do not remove or relocate citations"), so a violation is unambiguous.
The footprint percentage has no such backing. Measured on this skill's own fixtures, a *correct*
de-AI pass over an AI-inflated Discussion changed 61% of word tokens — because Patterns 6 and 18
require replacing formulaic limitation and conclusion paragraphs with specific content, which
rewrites whole paragraphs by design. A hard threshold would therefore fail exactly the edits the
skill asks for. The percentage is reported so a human can notice an implausible one; it is not
evidence of over-editing on its own, and the default is deliberately loose.

Exit codes: 0 clean or Minor-only, 1 with --strict when any Major fires, 2 usage error.

Scoped to keep false positives low:
  * Comparison is on WORD tokens, not characters, so punctuation-only fixes (Pattern 13
    em-dash -> parenthesis, Pattern 15 curly -> straight quotes) barely move the number.
  * Citation markers are stripped before numeric extraction, so a reference number is
    checked once as a citation and never again as a number.
  * Fenced code blocks are excluded from both sides.

Stdlib-only.

Usage:
    python3 check_rewrite_fidelity.py --before original.md --after humanized.md \
        [--out qc/rewrite_fidelity.json] [--warn-pct 70] [--strict] [--quiet]
"""
from __future__ import annotations

import argparse
import difflib
import json
import re
import sys
from collections import Counter
from pathlib import Path

DETECTOR_ID = "check_rewrite_fidelity"

FENCE_RE = re.compile(r"```.*?```", re.S)
# Pandoc citation keys and bare numeric markers, e.g. [@smith2020], [12], [3-5], [3–5].
CITEKEY_RE = re.compile(r"\[@[^\]\s]+\]")
NUMMARK_RE = re.compile(r"\[\d+(?:\s*[-–,]\s*\d+)*\]")
WORD_RE = re.compile(r"[A-Za-z0-9''-]+")
# A numeric token: integer, decimal, or percentage. Sign and thousands separators kept out
# so that "1,200" and "1200" compare equal after separator removal.
NUMBER_RE = re.compile(r"\d+(?:\.\d+)?")


def _strip_fences(text: str) -> str:
    return FENCE_RE.sub(" ", text)


def _citations(text: str) -> Counter:
    keys = CITEKEY_RE.findall(text)
    marks = NUMMARK_RE.findall(text)
    return Counter(k.strip() for k in keys + marks)


def _numbers(text: str) -> Counter:
    """Numeric tokens with citation markers removed first, so a reference number is not
    double-counted as a statistic."""
    without_cites = NUMMARK_RE.sub(" ", CITEKEY_RE.sub(" ", text))
    without_seps = without_cites.replace(",", "")
    return Counter(NUMBER_RE.findall(without_seps))


def _words(text: str) -> list[str]:
    return WORD_RE.findall(text.lower())


def _changed_fraction(before: list[str], after: list[str]) -> float:
    """Fraction of word tokens that differ, measured against the longer side so that a
    rewrite cannot lower the score by deleting text."""
    if not before and not after:
        return 0.0
    matcher = difflib.SequenceMatcher(a=before, b=after, autojunk=False)
    matched = sum(block.size for block in matcher.get_matching_blocks())
    denom = max(len(before), len(after))
    return 1.0 - (matched / denom) if denom else 0.0


def _diff_counter(before: Counter, after: Counter) -> list[dict]:
    out = []
    for token in sorted(set(before) | set(after)):
        b, a = before.get(token, 0), after.get(token, 0)
        if b != a:
            out.append({"token": token, "before": b, "after": a})
    return out


def main(argv: list[str] | None = None) -> int:
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--before", required=True, type=Path, help="text as it was before the rewrite")
    ap.add_argument("--after", required=True, type=Path, help="text after the humanize rewrite")
    ap.add_argument("--out", type=Path, help="write the JSON envelope here")
    ap.add_argument("--warn-pct", type=float, default=70.0,
                    help="Minor above this %% of words changed (default 70; advisory, see module docstring)")
    ap.add_argument("--strict", action="store_true", help="exit 1 when any Major fires")
    ap.add_argument("--quiet", action="store_true")
    args = ap.parse_args(argv)

    for path in (args.before, args.after):
        if not path.is_file():
            print(f"usage error: no such file: {path}", file=sys.stderr)
            return 2

    before_raw = _strip_fences(args.before.read_text(encoding="utf-8"))
    after_raw = _strip_fences(args.after.read_text(encoding="utf-8"))

    changed = _changed_fraction(_words(before_raw), _words(after_raw))
    changed_pct = round(changed * 100, 1)
    num_delta = _diff_counter(_numbers(before_raw), _numbers(after_raw))
    cite_delta = _diff_counter(_citations(before_raw), _citations(after_raw))

    claims: list[dict] = []
    if changed_pct > args.warn_pct:
        claims.append({
            "verdict": "EDIT_FOOTPRINT_HIGH",
            "severity": "Minor",
            "changed_pct": changed_pct,
            "threshold_pct": args.warn_pct,
            "message": (
                f"{changed_pct}% of word tokens changed (advisory threshold {args.warn_pct}%). "
                "A thorough de-AI pass can legitimately reach this level when Patterns 6 and 18 "
                "replace whole paragraphs; re-read the diff and confirm the author's argument, "
                "not just their phrasing, survived."
            ),
        })
    if num_delta:
        claims.append({
            "verdict": "NUMBER_DRIFT",
            "severity": "Major",
            "tokens": num_delta[:40],
            "message": (
                f"{len(num_delta)} numeric token(s) changed count across the rewrite. "
                "Humanize must never alter a number."
            ),
        })
    if cite_delta:
        claims.append({
            "verdict": "CITATION_DROP",
            "severity": "Major",
            "tokens": cite_delta[:40],
            "message": (
                f"{len(cite_delta)} citation(s) changed count across the rewrite. "
                "Humanize must never remove or relocate a citation."
            ),
        })

    envelope = {
        "detector": "check_rewrite_fidelity",
        "before": str(args.before),
        "after": str(args.after),
        "changed_pct": changed_pct,
        "words_before": len(_words(before_raw)),
        "words_after": len(_words(after_raw)),
        "claims": claims,
    }
    if args.out:
        args.out.parent.mkdir(parents=True, exist_ok=True)
        args.out.write_text(json.dumps(envelope, indent=2, ensure_ascii=False), encoding="utf-8")

    if not args.quiet:
        print(f"{DETECTOR_ID}: {changed_pct}% of words changed "
              f"({len(_words(before_raw))} -> {len(_words(after_raw))})")
        for claim in claims:
            print(f"  [{claim['severity']}] {claim['verdict']}: {claim['message']}")
        if not claims:
            print("  clean: footprint within bounds, numbers and citations preserved")

    if args.strict and any(c["severity"] == "Major" for c in claims):
        return 1
    return 0


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env python3
"""Verify the sentence-length variety the humanize skill already prescribes (Fix rule 7).

SKILL.md Phase 3 tells the rewriter: "Mix short declarative sentences (8-12 words) with longer
ones (25-35 words). Avoid uniform length." Until now nothing checked whether the rewrite obeyed
it, and a de-AI pass tends to flatten rhythm rather than restore it — the model shortens the
long sentences and pads the short ones toward a comfortable middle, which is itself a tell.

The threshold is not borrowed from any corpus. It is the skill's own specification: a text that
prescribes a mix of short and long sentences and then contains none of one kind has failed its
own rule. So this gate fires only on that unambiguous case — an absent band — and reports the
distribution for everything else rather than inventing a cutoff.

Rule 7 names a RANGE, and until 2026-07-29 this gate only checked its lower edge. A single
97-word sentence populates the ">= 25 words" band, so a text could satisfy the check while
containing a sentence nearly three times the top of the range rule 7 prescribes -- and the
gate printed `max_words: 97` in its own stats while reporting "clean". Two external reviewers
independently read such prose as machine-written on two different manuscripts while this
detector returned nothing.

Choosing the ceiling was the delicate part, and the measurement is recorded here so the number
is auditable rather than asserted. Rule 7's own top is 35, but firing there is unusable: across
the six manuscripts in this repository with enough sentences to measure, 97 sentences exceed 35
words (~24%, 16-22 per manuscript). Those are the project's own exemplars, so a gate that
condemns them is measuring the wrong thing -- 25-35 is a target band, not a ceiling. Doubling
rule 7's top gives 70, at which the same corpus yields 4 sentences in 6 manuscripts (~1%, at
most one per file). So the number is still derived from the skill's own specification (35 x 2),
and the corpus was used only to confirm it does not condemn good prose.

Verdicts:
  SENTENCE_UNIFORM (Minor)  the prose contains no short sentences (<= --short-max words) or no
                            long ones (>= --long-min), i.e. every sentence sits in the middle
                            band. Break up or combine sentences until both bands are populated.
  SENTENCE_OVERLONG (Minor) at least one sentence exceeds --long-max (default 70 = twice the top
                            of rule 7's range). Not "a longer one" in the sense rule 7 means; a
                            reader loses the subject before the verb arrives. Split it.

Scoped to keep false positives low:
  * Headings, list items, table rows, code fences, and YAML front matter are excluded — their
    lengths say nothing about prose rhythm.
  * Common academic abbreviations (et al., e.g., i.e., vs., Fig., approx.) and decimals do not
    split a sentence.
  * Citation markers are removed before counting, so [@key] and [12] never inflate a length.
  * Silent below --min-sentences (default 15): an abstract or a short note has too few
    sentences for rhythm to mean anything.

Exit codes: 0 clean or Minor-only, 1 with --strict when any Major fires (none — Minor only),
2 usage error. Stdlib-only.

Usage:
    python3 check_sentence_variety.py --manuscript manuscript.md \
        [--out qc/sentence_variety.json] [--short-max 12] [--long-min 25] \
        [--min-sentences 15] [--strict] [--quiet]
"""
from __future__ import annotations

import argparse
import json
import re
import statistics
import sys
from pathlib import Path

DETECTOR_ID = "check_sentence_variety"

FENCE_RE = re.compile(r"```.*?```", re.S)
FRONTMATTER_RE = re.compile(r"\A---\n.*?\n---\n", re.S)
SKIP_LINE_RE = re.compile(r"^\s*(#{1,6}\s|[-*+]\s|\d+\.\s|\||>\s|!\[|\[.*\]:)")
CITE_RE = re.compile(r"\[@[^\]\s]+\]|\[\d+(?:\s*[-–,]\s*\d+)*\]")
WORD_RE = re.compile(r"[A-Za-z0-9''-]+")
SENT_SPLIT_RE = re.compile(r"(?<=[.!?])\s+(?=[\"'(]?[A-Z0-9])")

# Abbreviations whose trailing period must not end a sentence. Order matters: longer first.
ABBREVIATIONS = (
    "et al.", "e.g.", "i.e.", "cf.", "vs.", "approx.", "Fig.", "Figs.", "Tab.", "No.",
    "Dr.", "Prof.", "St.", "Sr.", "Jr.", "Inc.", "Ltd.", "min.", "max.", "sec.",
)
_SENTINEL = "\x00ABBR%d\x00"
_DOT = "\x00DOT\x00"


def _prose_lines(text: str) -> str:
    """Body prose only: drop front matter, fences, headings, lists, tables, block quotes."""
    text = FRONTMATTER_RE.sub("", text)
    text = FENCE_RE.sub(" ", text)
    kept = [ln for ln in text.splitlines() if ln.strip() and not SKIP_LINE_RE.match(ln)]
    return " ".join(kept)


def _split_sentences(prose: str) -> list[str]:
    masked = prose
    for i, abbr in enumerate(ABBREVIATIONS):
        masked = masked.replace(abbr, _SENTINEL % i)
    # Protect decimals ("0.05", "P = .03") from the splitter. A lambda replacement is used
    # because a NUL byte cannot appear in an re.sub template string.
    masked = re.sub(r"(\d)\.(\d)", lambda m: m.group(1) + _DOT + m.group(2), masked)
    masked = re.sub(r"(?<=[=<>]\s)\.(\d)", lambda m: _DOT + m.group(1), masked)

    parts = SENT_SPLIT_RE.split(masked)

    out = []
    for part in parts:
        restored = part.replace(_DOT, ".")
        for i, abbr in enumerate(ABBREVIATIONS):
            restored = restored.replace(_SENTINEL % i, abbr)
        restored = restored.strip()
        if restored:
            out.append(restored)
    return out


def _word_count(sentence: str) -> int:
    return len(WORD_RE.findall(CITE_RE.sub(" ", sentence)))


def main(argv: list[str] | None = None) -> int:
    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    ap.add_argument("--manuscript", required=True, type=Path)
    ap.add_argument("--out", type=Path, help="write the JSON envelope here")
    ap.add_argument("--short-max", type=int, default=12, help="a short sentence is <= this (SKILL.md rule 7)")
    ap.add_argument("--long-min", type=int, default=25, help="a long sentence is >= this (SKILL.md rule 7)")
    ap.add_argument("--long-max", type=int, default=70,
                    help="a sentence over this is overlong (default 70 = twice the top of rule 7's 25-35 range)")
    ap.add_argument("--min-sentences", type=int, default=15, help="stay silent below this many sentences")
    ap.add_argument("--strict", action="store_true")
    ap.add_argument("--quiet", action="store_true")
    args = ap.parse_args(argv)

    if not args.manuscript.is_file():
        print(f"usage error: no such file: {args.manuscript}", file=sys.stderr)
        return 2

    sentences = _split_sentences(_prose_lines(args.manuscript.read_text(encoding="utf-8")))
    lengths = [n for n in (_word_count(s) for s in sentences) if n > 0]

    claims: list[dict] = []
    stats: dict = {"sentences": len(lengths)}

    if len(lengths) >= args.min_sentences:
        short = [n for n in lengths if n <= args.short_max]
        long_ = [n for n in lengths if n >= args.long_min]
        over_band = [n for n in lengths if n > 35]
        overlong = sorted((n for n in lengths if n > args.long_max), reverse=True)
        stats.update({
            "short_count": len(short),
            "long_count": len(long_),
            "median_words": round(statistics.median(lengths), 1),
            "min_words": min(lengths),
            "max_words": max(lengths),
            "stdev_words": round(statistics.pstdev(lengths), 1),
            # Reported, not judged. Rule 7's band tops out at 35, but ~24% of sentences in this
            # repository's own exemplar manuscripts exceed it, so the count is information for a
            # rewriter rather than grounds for a verdict.
            "over_rule7_band_count": len(over_band),
            "overlong_word_counts": overlong,
        })
        if not short or not long_:
            missing = "short" if not short else "long"
            band = (f"<= {args.short_max} words" if not short else f">= {args.long_min} words")
            claims.append({
                "verdict": "SENTENCE_UNIFORM",
                "severity": "Minor",
                "missing_band": missing,
                "message": (
                    f"No {missing} sentences ({band}) across {len(lengths)} sentences "
                    f"(median {stats['median_words']}, range {stats['min_words']}-{stats['max_words']}). "
                    "SKILL.md Fix rule 7 requires both bands; uniform length is itself an AI tell."
                ),
            })
        if overlong:
            claims.append({
                "verdict": "SENTENCE_OVERLONG",
                "severity": "Minor",
                "count": len(overlong),
                "word_counts": overlong,
                "message": (
                    f"{len(overlong)} sentence(s) over {args.long_max} words "
                    f"({', '.join(str(n) for n in overlong)}). Rule 7's long band is 25-35 words; "
                    f"{args.long_max} is twice its top, so these are not 'a longer one' in the sense "
                    "the rule means — the reader loses the subject before the verb arrives. Split them."
                ),
            })
    else:
        stats["skipped"] = f"only {len(lengths)} sentences (< --min-sentences {args.min_sentences})"

    envelope = {
        "detector": "check_sentence_variety",
        "manuscript": str(args.manuscript),
        "short_max": args.short_max,
        "long_min": args.long_min,
        "long_max": args.long_max,
        "stats": stats,
        "claims": claims,
    }
    if args.out:
        args.out.parent.mkdir(parents=True, exist_ok=True)
        args.out.write_text(json.dumps(envelope, indent=2, ensure_ascii=False), encoding="utf-8")

    if not args.quiet:
        if "skipped" in stats:
            print(f"{DETECTOR_ID}: skipped — {stats['skipped']}")
        else:
            print(f"{DETECTOR_ID}: {stats['sentences']} sentences, median {stats['median_words']} words, "
                  f"{stats['short_count']} short / {stats['long_count']} long")
        for claim in claims:
            print(f"  [{claim['severity']}] {claim['verdict']}: {claim['message']}")
        if not claims and "skipped" not in stats:
            print("  clean: both short and long sentence bands are populated")

    if args.strict and any(c["severity"] == "Major" for c in claims):
        return 1
    return 0


if __name__ == "__main__":
    sys.exit(main())

#!/usr/bin/env bash
# Regression test for the humanize rewrite-fidelity gate.
# (bounded)   a correct de-AI pass -> no claim, even though it changed ~62% of the words;
# (wholesale) a full rewrite        -> EDIT_FOOTPRINT_HIGH only (Minor, never blocks);
# (numdrift)  a changed number and a dropped citation -> NUMBER_DRIFT + CITATION_DROP, exit 1.
set -u
HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
SCRIPT="$HERE/../scripts/check_rewrite_fidelity.py"
BEFORE="$HERE/fixtures/rewrite_before.md"
BOUNDED="$HERE/fixtures/rewrite_after_bounded.md"
WHOLESALE="$HERE/fixtures/rewrite_after_wholesale.md"
NUMDRIFT="$HERE/fixtures/rewrite_after_numdrift.md"
OUT="$(mktemp -t rwfid_XXXX).json"
trap 'rm -f "$OUT"' EXIT
fail=0
check() { local label="$1"; shift
  if "$@" >/dev/null 2>&1; then printf '  PASS  %s\n' "$label"
  else printf '  FAIL  %s\n' "$label"; fail=$((fail+1)); fi; }
[[ -f "$SCRIPT" ]] || { echo "ENV-ERR: script missing" >&2; exit 2; }

python3 "$SCRIPT" --before "$BEFORE" --after "$BOUNDED" --out "$OUT" --quiet >/dev/null 2>&1
check "no claim on a correct de-AI pass (numbers + citations preserved)" python3 -c "
import json
d=json.load(open('$OUT'))
assert not d['claims'], d['claims']
assert d['detector']=='check_rewrite_fidelity', d.get('detector')
"
check "a correct de-AI pass does not block under --strict" \
  python3 "$SCRIPT" --before "$BEFORE" --after "$BOUNDED" --strict --quiet

python3 "$SCRIPT" --before "$BEFORE" --after "$WHOLESALE" --out "$OUT" --quiet >/dev/null 2>&1
check "EDIT_FOOTPRINT_HIGH on a wholesale rewrite" python3 -c "
import json
d=json.load(open('$OUT'))
c=[x for x in d['claims'] if x['verdict']=='EDIT_FOOTPRINT_HIGH']
assert c, d['claims']
assert c[0]['severity']=='Minor', c[0]
"
check "footprint alone never blocks under --strict (advisory only)" \
  python3 "$SCRIPT" --before "$BEFORE" --after "$WHOLESALE" --strict --quiet

python3 "$SCRIPT" --before "$BOUNDED" --after "$NUMDRIFT" --out "$OUT" --quiet >/dev/null 2>&1
check "NUMBER_DRIFT when a statistic changed" python3 -c "
import json
d=json.load(open('$OUT'))
assert any(c['verdict']=='NUMBER_DRIFT' and c['severity']=='Major' for c in d['claims']), d['claims']
"
check "CITATION_DROP when a citation disappeared" python3 -c "
import json
d=json.load(open('$OUT'))
assert any(c['verdict']=='CITATION_DROP' and c['severity']=='Major' for c in d['claims']), d['claims']
"
if python3 "$SCRIPT" --before "$BOUNDED" --after "$NUMDRIFT" --strict --quiet >/dev/null 2>&1; then
  printf '  FAIL  %s\n' "--strict exits 1 on an invariant violation"; fail=$((fail+1))
else
  printf '  PASS  %s\n' "--strict exits 1 on an invariant violation"
fi

echo "fail=$fail"; [[ "$fail" -eq 0 ]] && echo "ALL PASS" || echo "FAILURES: $fail"
exit "$fail"

#!/usr/bin/env bash
# Regression test for the sentence-length variety gate (humanize SKILL.md Fix rule 7).
# (uniform) every sentence in the middle band, no long ones -> SENTENCE_UNIFORM;
# (mixed)   both short (<=12) and long (>=25) bands populated -> no claim;
# (short)   a text below --min-sentences is skipped rather than judged.
set -u
HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
SCRIPT="$HERE/../scripts/check_sentence_variety.py"
UNIFORM="$HERE/fixtures/variety_uniform.md"
MIXED="$HERE/fixtures/variety_mixed.md"
OUT="$(mktemp -t svar_XXXX).json"
SHORT="$(mktemp -t svarshort_XXXX).md"
trap 'rm -f "$OUT" "$SHORT"' EXIT
fail=0
check() { local label="$1"; shift
  if "$@" >/dev/null 2>&1; then printf '  PASS  %s\n' "$label"
  else printf '  FAIL  %s\n' "$label"; fail=$((fail+1)); fi; }
[[ -f "$SCRIPT" ]] || { echo "ENV-ERR: script missing" >&2; exit 2; }

python3 "$SCRIPT" --manuscript "$UNIFORM" --out "$OUT" --quiet >/dev/null 2>&1
check "SENTENCE_UNIFORM when no long sentences exist" python3 -c "
import json
d=json.load(open('$OUT'))
c=[x for x in d['claims'] if x['verdict']=='SENTENCE_UNIFORM']
assert c, d['claims']
assert c[0]['missing_band']=='long', c[0]
assert d['detector']=='check_sentence_variety', d.get('detector')
"
python3 "$SCRIPT" --manuscript "$MIXED" --out "$OUT" --quiet >/dev/null 2>&1
check "no claim when both bands are populated" python3 -c "
import json
d=json.load(open('$OUT'))
assert not d['claims'], d['claims']
assert d['stats']['short_count'] > 0 and d['stats']['long_count'] > 0, d['stats']
"
check "decimals and abbreviations do not split a sentence" python3 -c "
import json
d=json.load(open('$OUT'))
# 'R 4.3.1' and 'et al.' appear in the mixed fixture; a naive splitter inflates the count.
assert d['stats']['sentences'] < 25, d['stats']
"
printf 'One short line. Another one here. A third.\n' > "$SHORT"
python3 "$SCRIPT" --manuscript "$SHORT" --out "$OUT" --quiet >/dev/null 2>&1
check "a text below --min-sentences is skipped, not judged" python3 -c "
import json
d=json.load(open('$OUT'))
assert not d['claims'], d['claims']
assert 'skipped' in d['stats'], d['stats']
"
# Rule 7 names a RANGE and this gate only ever checked its lower edge: a single 97-word
# sentence populates the ">= 25 words" band, so the check passed while printing max_words: 97
# in its own stats. Two external reviewers read such prose as machine-written on two different
# manuscripts while the detector returned nothing.
OL="$HERE/fixtures/variety_overlong.md"
python3 "$SCRIPT" --manuscript "$OL" --out "$OUT" --quiet >/dev/null 2>&1
check "a 97-word sentence fires SENTENCE_OVERLONG" python3 -c "
import json
d=json.load(open('$OUT'))
c=[x for x in d['claims'] if x['verdict']=='SENTENCE_OVERLONG']
assert c, d['claims']
assert c[0]['word_counts']==[97], c[0]
"
check "...and the band check still says the prose is varied" python3 -c "
import json
d=json.load(open('$OUT'))
assert not [x for x in d['claims'] if x['verdict']=='SENTENCE_UNIFORM'], d['claims']
"
check "...and it stays Minor: --strict does not fail the build" test "$(python3 "$SCRIPT" --manuscript "$OL" --strict >/dev/null 2>&1; echo $?)" -eq 0

# The bound has to be a bound, not a synonym for "long". Same text, longest sentence at 64
# words -- over rule 7's band of 25-35, under the 70 ceiling -- must stay silent, or the
# verdict would just be re-reporting the long band it already checks.
WB="$HERE/fixtures/variety_long_within_bound.md"
python3 "$SCRIPT" --manuscript "$WB" --out "$OUT" --quiet >/dev/null 2>&1
check "a 64-word sentence does NOT fire (over rule 7's band, under the ceiling)" python3 -c "
import json
d=json.load(open('$OUT'))
assert not [x for x in d['claims'] if x['verdict']=='SENTENCE_OVERLONG'], d['claims']
"
check "...and the count over rule 7's band is reported, not judged" python3 -c "
import json
d=json.load(open('$OUT'))
assert d['stats']['over_rule7_band_count'] >= 1, d['stats']
assert not d['claims'], d['claims']
"
# The threshold is configurable, and lowering it must actually change the answer -- otherwise
# the default could be doing nothing and every assertion above would still pass.
python3 "$SCRIPT" --manuscript "$WB" --out "$OUT" --long-max 50 --quiet >/dev/null 2>&1
check "--long-max 50 makes the same 64-word sentence fire" python3 -c "
import json
d=json.load(open('$OUT'))
assert [x for x in d['claims'] if x['verdict']=='SENTENCE_OVERLONG'], d['claims']
"

echo "fail=$fail"; [[ "$fail" -eq 0 ]] && echo "ALL PASS" || echo "FAILURES: $fail"
exit "$fail"