CLN Compare Hub · Research Series

LLM Output
vs CLN Format

Same prompts. Same model. Different instructions.
We score both outputs across 7 cognitive metrics to measure — not just show — the difference.

Standard LLM — unconstrained prose CLN Format — operator-structured 7 metrics · scored · verifiable

The Scoring System

7 metrics. Each scored 1–10. Measured from text properties — not opinion.

! Semantic Density SD

Meaning per character. High filler words = low density. Counts connective tissue: "furthermore", "it is worth noting", "in addition to this", etc.

SD = 1 − (filler_words / total_words)
filler list: ~40 phrases tracked per text

. Scan Speed SS

Time to locate a specific fact. Measured by line count to first actionable item and average line length. Short left-column operators = instant semantic pre-scan.

SS = 10 − (avg_line_chars / 20)
capped 1–10; operator col bonus +2

= Structure Clarity SC

Ratio of typed semantic units to total lines. Each CLN operator = one unambiguous semantic type. Prose has zero explicit typing.

SC = typed_lines / total_lines × 10
CLN: every line typed · Prose: 0 typed

* Recall Ease RE

Proxy for how much of the content a reader retains after single read. Based on chunking: isolated short lines > long embedded clauses (Miller's Law).

RE = chunks_under_7_words / total_chunks
× 10 · chunk = one semantic unit

~ Tension Visibility TV

Are trade-offs, paradoxes, and unknowns explicitly surfaced? Prose often buries tensions in hedged clauses. CLN ~ and ? operators make them first-class citizens.

TV = explicit_tensions / total_tensions_in_text
× 10 · tensions identified by domain analysis

→ Action Clarity AC

Can you extract "what to do" without re-reading? Counts sentences that contain an imperative verb and are NOT buried inside a paragraph.

AC = isolated_imperatives / total_imperatives
× 10 · CLN . and → are structurally isolated

? Unknown Honesty UH

Does the output admit what it doesn't know? LLMs often hedge vaguely ("may depend on…"). CLN ? operator forces explicit placement of unknowns.

UH = explicit_unknowns / (explicit + implicit)
implicit = "may", "might", "could", "it depends"

Score strips above: Average LLM baseline CLN-instructed output

// Methodology

same prompt → 2 model calls: no system prompt · CLN spec v2.1 as system prompt

model: same Claude Sonnet 4.6 instance · temperature 0.5

scoring: computed from text properties, not subjective rating