Skip to content

House-Style and Clarity Gate (`Syntaxis` & Editorial Standards)

This reference defines the mandatory prose and clarity rules for academic preprints and technical manuscripts. Every draft must pass these checks before publication.


1. Prohibited Rhetorical Devices

  1. Zero Em-Dashes (—) in Prose:
    • Do not use — (U+2014) anywhere in abstract, body paragraphs, list items, or figure/table captions.
    • Why: Models overuse em-dashes as a generic rhetorical splice instead of specifying whether the second clause is a causal explanation (use a colon :), a coordinate independent clause (use a semicolon ;), or a parenthetical aside (use parentheses (...)).
  2. Zero Throat-Clearing Openers:
    • Ban openers that delay the subject and verb:
      • "It is important to note that..."
      • "It is worth mentioning that..."
      • "In the rapidly evolving landscape of..."
      • "Needless to say..."
    • Fix: Delete the preamble and begin directly with the actor and action.
  3. Zero Lazy Binary-Contrast Frames ("Not X, but Y"):
    • Avoid formulaic contrast templates such as "This is not merely X, it is Y" or "The goal is not to X, but rather to Y" unless contrasting two formal mathematical definitions ($P = 0$ vs. $\mathcal{T} = 1$).
  4. Zero Unanchored Evaluative Adjectives:
    • Never write "stunning improvement", "massive speedup", or "dramatic reduction" without immediately binding the noun phrase to the exact numeric multiplier and baseline comparison in the same sentence (for example: "cuts wall-clock latency by $4.23\times$ (54.10s vs. 229.02s)").

2. Active Voice & Named Actors

Every sentence describing system behavior or experimental methodology must name the concrete actor:

  • Weak (Passive / Unattributed): "Candidate outputs are validated and errors are fixed in subsequent turns."
  • Strong (Named Actor): "In Conditions C and D, the executor $M$ invokes run_validator on workspace $W_k$, receives line-level AST diagnostics from oracle $V$, and patches the offending struct field in Turn $k+1$."

3. Clean-Prose Readability & Technical Density Calibration

  1. Always Measure Clean Prose (Avoid the Raw-Markdown Illusion):
    • Running Flesch-Kincaid (FK) on a raw Markdown file (paper.md) produces an artificially low grade (FK ~11–13) because Markdown table rows and headers are parsed as short sentences, masking Very Dense or Impenetrable prose (FK 20+).
    • Always strip Markdown tables, fenced code blocks, inline code tokens, and LaTeX display/inline math ($$...$$, $...$) before evaluating readability (e.g., via ~/projects/docstats or scripts/validate_preprint.py).
  2. Target Clean-Prose Metrics (Dense Systems-Paper Band):
    • Flesch-Kincaid Grade (flesch_kincaid_grade): 12.0 to 14.5 (warn if $> 16.0$).
    • Flesch Reading Ease (flesch_reading_ease): 30.0 to 50.0.
    • Average Sentence Length (ASL): 15.0 to 19.0 words/sentence (never exceed 22.0 w/s overall; avoid sentences $\ge 38$ words).
    • Average Syllables per Word (ASW): 1.75 to 1.88 syllables/word.
    • Axis B (ai_tell_score in docstats): $\ge 9.5 / 10.0$ (0 em-dashes, 0 throat-clearing openers, 0 binary contrasts, 0 offender adverbs, -ly adverb rate $\le 1.0 / 100\text{w}$, sentence-length CV $\ge 0.25$).
  3. Three Levers to Lower Technical Density Without Losing Rigor:
    • Unpack Multi-Clause Sentences: Do not pack a conceptual mechanism, 4 parenthetical metrics, and 3 CLI flags into a single 60-word sentence. State the mechanism in one crisp 12–16 word sentence, then report the numeric comparison in the next sentence.
    • Replace Polysyllabic Nominalizations with Plain Verbs and Short Nouns: After defining a formal term once, prefer "syntax checks" over "syntactic validity verification", "token generation" over "autoregressive generation", "rounding coordinates" over "coordinate canonicalization", and "matching path commands" over "command-signature parity".
    • Paragraph Unity & Rhythm: Open each paragraph with a direct 10–15 word topic sentence stating the claim, develop the formal or empirical mechanism in the middle sentences, and close with the quantitative consequence or table/figure reference.