Version: v1 · Status: ready

Traits

Information Density

Every sentence adds new information or advances the argument ↔ Low signal-to-noise — removable filler without meaning loss

Measures the ratio of substantive content to total words. High-scoring text is dense — every sentence advances the argument or adds new information. Low-scoring text has a low signal-to-noise ratio: many words consumed by filler, repetition, or decorative language that could be removed without losing meaning.

Preamble Absence

Leads with the point — no throat-clearing or restating before substance ↔ Opens with pleasantries, question-restating, or contextual preamble

Measures whether the text leads with substance rather than throat-clearing. High-scoring text starts with the answer or the point. Low-scoring text opens with "That's a great question", "I'd be happy to help", restating the question back, or contextualizing preamble before getting to the actual content.

Hedge Absence

Commits to statements directly when confidence is warranted ↔ Excessive hedging — "it's worth noting", "one could argue", qualifiers everywhere

Measures whether the text commits to its statements without excessive qualification. High-scoring text states things directly when confidence is warranted. Low-scoring text wraps every claim in "It's worth noting that...", "One could argue...", "It's important to consider...", or other hedging constructions that add words without adding nuance.

Structural Efficiency

Structure serves comprehension — no redundant summaries or padding structures ↔ Structural padding — unnecessary lists, repeated summaries, empty transitions

Measures whether the text's structure serves communication rather than padding length. High-scoring text uses structure (paragraphs, lists, headers) only when they aid comprehension. Low-scoring text creates unnecessary numbered lists for two items, adds summary sections that repeat the body, or uses transitional paragraphs that say nothing new.

Repetition Absence

Each idea stated once in its strongest form, then the text moves on ↔ Same point restated in different words across multiple sentences

Measures whether each idea appears exactly once in its strongest form. High-scoring text makes a point, then moves on. Low-scoring text restates the same idea in different words across multiple sentences or paragraphs, often in a "tell them what you'll say, say it, tell them what you said" pattern that treats the reader as inattentive.

About

u22a8.conciseness

Measures whether text communicates efficiently without unnecessary padding.

Existing conciseness evaluators (Phoenix, Fiddler) return a binary "concise or verbose." This model captures the specific patterns: information density (signal-to-noise ratio), preamble absence (no throat-clearing before the point), hedge absence (commitment vs. endless qualification), structural efficiency (structure that serves comprehension vs. padding), and repetition absence (each idea stated once in its strongest form).

The model targets the verbosity patterns that LLMs are particularly prone to: restating the question, "I'd be happy to help", meta-commentary about the response itself, unnecessarily numbered lists, and the tell-them-three-times pattern. It measures information density rather than raw word count — a long text full of substance scores well; a short text full of filler does not.

Limitations

Does not assess whether conciseness comes at the cost of completeness or clarity. Very complex topics may legitimately require more words. Some audiences (beginners, non-native speakers) benefit from deliberate redundancy that this model would penalize. Best used alongside a correctness or completeness check.

Pairs well with

• u22a8.specificity — dense text is usually specific
• u22a8.sycophancy — sycophantic patterns are verbose by nature
• u22a8.answer-relevancy — focused answers tend to be concise

Docs

• Tiers and scoring — the per-trait trained boundaries between tiers
• Breaks — where meaningful quality transitions occur