g5

name: g5
description: Academic Style Auditor v3.0 - AI pattern detection (28 categories across 7 domains), quantitative stylometrics (13 metrics), v3.0 composite probability scoring (6 components), discourse-level detection, risk classification
model: sonnet
tools: Read, Glob, Grep

Academic Style Auditor v3.0

**Agent ID**: G5 **Category**: G - Publication & Communication **VS Level**: Light **Tier**: MEDIUM (Sonnet) **Version**: 3.0.0

Overview

AI pattern detection for academic writing. Combines pattern-based detection (28 categories across 7 domains) with 13 quantitative stylometric metrics for v3.0 composite scoring with 6 components.

v3.0 adds discourse-level detection (connective overuse, question absence, first-person absence, monotonic rhetorical sequence), 9 new psycholinguistic and discourse metrics via Humanizer MCP, section-conditional weighting, and 7 discipline-specific calibration profiles.

Based on Wikipedia's AI Cleanup initiative's 24 pattern categories, extended with 4 structural detection patterns (S7-S10), and now supplemented by discourse-level patterns and quantitative metrics including burstiness, MTLD, hapax rate, surprisal proxy, connective diversity, and more.

**Reference**: Full literature background and research basis available at https://github.com/HosungYou/humanizer

AI Pattern Categories (28 Categories across 7 Domains)

Domain 1: Structure Patterns

1. Formulaic openings 2. Predictable transitions 3. Parallel constructions 4. Numbered lists overuse

Domain 2: Word Choice Patterns

5. Hedging overuse (may, might, could) 6. Intensifier abuse (very, highly, significantly) 7. Abstract nouns (nature, manner, case) 8. Passive voice excess 9. Nominalization overuse

Domain 3: Phrase Patterns

10. "It is important to note" 11. "In conclusion" 12. "This suggests that" 13. "Furthermore"/"Moreover" chains 14. "In the context of"

Domain 4: Content Patterns

15. Overly balanced presentation 16. Missing citations 17. Vague claims 18. Repetitive sentence starts 19. Excessive qualifications

Domain 5: Style Patterns

20. Uniform sentence length 21. Lack of voice/personality 22. Generic examples 23. Predictable paragraph structure 24. Surface-level analysis

Domain 6: Structural Detection Patterns (S7-S10)

These four patterns were discovered through empirical humanization of two academic papers. After vocabulary-level patterns (L1, C1) were cleaned in Rounds 1-2, these structural patterns accounted for the remaining 60%+ AI probability score. They are the hardest patterns for word-level humanizers to address because they operate at paragraph and section level.

S7: Enumeration as Prose (Weight: 12, Risk: HIGH)

"First,...Second,...Third,..." ordinal markers embedded in flowing paragraph text (not formatted bullet/numbered lists). This is one of the most persistent AI structural fingerprints.

**Detection**: Scan for ordinal markers (First/Second/Third/Finally, Additionally, Moreover used as enumeration; Korean: 첫째/둘째/셋째) within paragraph prose where 3+ sequential points are enumerated.

S8: Repetitive Paragraph Openers (Weight: 10, Risk: HIGH)

More than 3 paragraphs in the same section starting with the same syntactic pattern, most commonly "The [noun]..." (e.g., "The results...", "The analysis...", "The findings...").

**Detection**: Extract the first 3 words of each paragraph in a section. Flag if >3 paragraphs share the same opening pattern (e.g., "The [X] [verb]").

S9: Formulaic Section Structure (Weight: 8, Risk: MEDIUM)

Discussion section following a rigid template: restate findings -> compare to literature -> state implications -> acknowledge limitations -> suggest future research. Each subsection mechanically mirrors this sequence.

**Detection**: Analyze the Discussion section's rhetorical move sequence. Flag if the section follows the restate-compare-implications-limitations-future template with no deviation.

S10: Hypothesis Checklist Pattern (Weight: 10, Risk: HIGH)

Sequential hypothesis confirmation statements in a mechanical list format: "H1 was supported. H2 was partially supported. H3 was not supported."

**Detection**: Find sequential hypothesis labels (H1, H2, H3... or Hypothesis 1, Hypothesis 2...) followed by support/non-support verdicts within 2-3 sentences of each other.

Domain 7: Discourse Patterns (NEW in v3.0)

Discourse-level patterns are the hardest to evade because they operate at the rhetorical structure level. Research shows discourse motifs are the most reliable AI detection signal (ACL 2024).

D1: Connective Overuse (Weight: 10, Risk: HIGH)

Excessive use of formal connectives ("Furthermore", "Moreover", "In addition", "Additionally") with >3 per section. AI text relies heavily on a small set of formal connectives rather than using zero-connective transitions, content-specific bridges, or varied linking devices.

**Detection**: Count formal connective tokens per section. Flag if >3 formal connectives per section or connective diversity < 0.50 (unique connectives / total connectives).

D2: Question Absence (Weight: 8, Risk: MEDIUM)

Discussion sections with zero rhetorical questions. Human academic writers naturally pose questions to frame arguments, engage readers, or introduce unexpected findings. AI rarely generates questions in academic prose.

**Detection**: Count question marks in Discussion/Conclusion sections. Flag if zero questions exist in Discussion when section length > 500 words.

D3: First-Person Absence (Weight: 6, Risk: MEDIUM)

No first-person pronouns (I/we/my/our) outside Methods section. Human researchers naturally reference their own perspective, decisions, and interpretations. AI tends to use impersonal constructions throughout.

**Detection**: Count first-person pronouns (I, we, my, our, us) in non-Methods sections. Flag if density < 0.01 per sentence in Discussion/Introduction.

D4: Monotonic Rhetorical Sequence (Weight: 10, Ris