House-Style Linting

Axis B enforces crisp, direct technical prose through deterministic pattern checks. It provides house-style linting rather than statistical AI detection. This reference details the underlying rules, the data model, and score calculation mechanics.

Deterministic House-Style Linting

In benchmark evaluations, heuristic pattern rules demonstrated limited classification power for detecting synthetic origin (AUC = 0.577 general, AUC = 0.403 technical). In technical writing, em-dash usage inverts: human technical authors employ em dashes as frequently as language models. Axis B makes no provenance determination; it enforces direct technical house style across all prose.

Rule Classification

The ten rules originated in the technical-post-editorial skill. Machine-detectable rules feed Axis B automated checks; human-only dimensions rely on editorial review.

#	Rule	Verification Method	Axis B Signal	Notes / Exceptions
1	No em dashes in prose	Automated	em_dash_count	Excludes code spans and tables; allows markdown list separators.
2	Active voice, named actors	Human-assisted	passive_hint_count (weak)	Passive detection is heuristic; false agency requires human judgment.
3	No filler adverbs	Automated	adverb_ly_rate	Excludes technical allowlist (atomically, synchronously, recursively, etc.).
4	No throat-clearing openers	Automated	throat_clearing_count	Phrase matching: “Here’s the thing”, “It’s worth noting”, “It turns out”.
5	No binary contrast framing	Automated	binary_contrast_count	Pattern matching: “not X, it’s Y”, “isn’t … it’s”, “not only … but”.
6	No staccato fragmentation	Automated (heuristic)	fragment_count	Flags verbless sentence fragments as advisory hints.
7	No Wh- sentence starters	Automated	wh_starter_rate	Sentence-initial What, When, Where, Which, Who, Why, How.
8	Vary sentence rhythm	Automated (advisory)	sentence_len_cv, list_of_three_count	Advisory hint. Prompting models with hard CV targets degrades rhythm variance.
9	No vague declaratives	Automated (heuristic)	vague_declarative_count	Flags announcement phrasing (“The implications are significant”).
10	Trust the reader	Human judgment	None	Evaluates permission-granting, hand-holding, and meta-commentary.

Pattern Scores Model

Axis B executes through the shared analysis engine alongside Axis A:

Field	Type	Description
em_dash_count	int	Em dashes in prose (excludes code blocks and tables).
adverb_ly_rate	float	Non-technical -ly adverbs per 100 words.
throat_clearing_count	int	Matched opening preamble phrases.
binary_contrast_count	int	Matched binary contrast frames.
wh_starter_rate	float	Wh- sentence openings per 100 sentences.
fragment_count	int	Heuristic sentence fragment count.
list_of_three_count	int	Three-item parallel lists in prose.
sentence_len_cv	float	Coefficient of variation of sentence length (advisory).
vague_declarative_count	int	Vague significance announcements.
passive_hint_count	int	Heuristic passive-voice occurrences.
total_tells	int	Total high-confidence tell count.
ai_tell_score	float	Aggregate style conformity score (0–10 scale, 10 = clean).
confidence	str	Evaluation confidence (“high” or “low” for word_count < 100).

Detectors execute on extracted prose only. Code blocks, inline code spans, and table cells are stripped beforehand.

Style Conformity Score Computation

ai_tell_score maps to a 0–10 scale, where 10 represents clean prose without style violations. It penalizes normalized rates according to confidence tiers:

• High-confidence patterns (rules 1, 3, 4, 5, 7): Full penalty weight.
• Heuristic patterns (rules 6, 9) and passive hints: Reduced penalty weight.
• Rhythm metrics (sentence_len_cv): Evaluated as advisory indicators without penalties.

Passing threshold: ai_tell_score >= 7.0

Axis B Verdict Thresholds

• Pass: ai_tell_score >= 7.0 with no single high-confidence category elevated.
• Warn: ai_tell_score between 5.0 and 6.9, or one elevated high-confidence category.
• Fail: ai_tell_score < 5.0, or multiple elevated high-confidence categories.

Editorial Override Protocol

Human reviewers may retain flagged patterns when context justifies the device (such as an intentional contrast or technical emphasis). The scorecard records overrides with an explicit rationale, preserving auditability while respecting editorial discretion.

Rhythm Variance Guidelines

sentence_len_cv measures sentence length variation. Low coefficient of variation values indicate potential monotone pacing. Research (experiment E4 in the research program) demonstrated that prompting language models with hard CV targets degrades natural sentence variation. Use CV exclusively as an advisory diagnostic.