Version: v1 · Status: ready

Traits

Concrete Reference Density

Names specific entities, versions, dates, or verifiable references ↔ Makes claims without naming specifics — "many", "some", "often"

Measures whether the text anchors claims in specific, nameable things — proper nouns, version numbers, dates, place names, named tools, specific people, particular companies. High-scoring text grounds every assertion in something verifiable or at least recognizable. Low-scoring text makes claims without naming anything specific ("many companies have found that...").

Particular Detail

Details that uniquely identify the subject, not interchangeable descriptors ↔ Generic adjectives applicable to anything in the category

Measures whether descriptions use precise, distinguishing details rather than generic adjectives. High-scoring text describes things in ways that could only apply to the specific thing being discussed. Low-scoring text uses interchangeable descriptors that could apply to anything in the category ("innovative solution", "robust platform", "cutting-edge technology").

Example Concreteness

Real, situated examples drawn from specific actual cases ↔ Hypothetical "imagine if..." or schematic "a company might..." examples

Measures whether examples and illustrations are drawn from real, specific situations rather than hypothetical or schematic scenarios. High-scoring text says "like when Stripe moved from per-request auth to session tokens in 2019." Low-scoring text says "for example, a company might want to improve their authentication system."

Quantification

Numbers, measurements, percentages, or counts backing claims ↔ Vague intensifiers — "significantly", "dramatically", "many"

Measures whether claims about magnitude, frequency, or impact are backed by numbers rather than vague intensifiers. High-scoring text says "reduced p95 latency from 340ms to 12ms" or "3 of our 40 enterprise customers." Low-scoring text says "significantly improved performance" or "many of our customers."

Voice Distinctiveness

Distinctive authorial choices in diction, rhythm, or structure ↔ Smooth default prose indistinguishable from generic LLM output

Measures whether the text has identifiable authorial choices that distinguish it from default LLM prose. High-scoring text uses unexpected word choices, sentence rhythms, or structural decisions that could not have been produced by asking an LLM "write about X." Low-scoring text reads like unedited model output — smooth, competent, and interchangeable with any other text on the topic.

About

u22a8.specificity

Measures how concrete and specific text is versus generic LLM-style prose.

The core tell of AI-generated content isn't grammar or fluency — it's the absence of specificity. Generic text says "many companies have found success"; specific text says "Stripe reduced checkout abandonment 14% after switching to one-click in Q3 2023." This model captures the linguistic markers of concreteness: proper nouns, quantification, particular details, real examples, and distinctive authorial voice.

Five traits decompose specificity: concrete reference density (are entities named?), particular detail (would this description only fit this subject?), example concreteness (real cases vs. hypothetical scenarios), quantification (numbers vs. vague intensifiers), and voice distinctiveness (identifiable authorial choices vs. smooth default output).

Limitations

Some domains legitimately require abstraction — philosophy, policy, and theoretical work score lower on specificity without being lower quality. The model measures concreteness as a signal, not as a universal quality indicator. Highly technical text with domain-specific terminology may score well on specificity markers even when unclear to a general audience.

Pairs well with

• u22a8.conciseness — specific text tends to be denser
• u22a8.sycophancy — sycophantic text is characteristically vague
• u22a8.humor — comedy specificity overlaps

Docs

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