The Observatory of Structured Reasoning
Back To Disposition Map

Disposition Probe · Candidate spoke

Lexical Register

Lexical Register asks whether a model has a measurable word-choice signature: vocabulary level, rarity, repetition, length, hedging, certainty, and distinctive phrasing across matched prompts.

Status: candidate-spoke Dissociation shown Split-half replication shown Multi-ruler validated
Map entry Lexical Register is live, not locked. Open the Disposition Map →

What It Measures

Lexical Register measures word-choice and register. It does not claim reasoning quality. A model can use richer vocabulary and still reason poorly, or use plainer vocabulary and preserve the important structure. The ruler is lexical, not moral and not intellectual rank.

Lexical Register measures word-choice, not reasoning quality.

Why It Counts As A Candidate

The first gate is dissociation. Lexical Register does not simply follow Ground Transfer or Frame Shift. Claude and GPT both keep ground but separate lexically; Gemini, Grok, and Qwen all shed ground but scatter across length, diversity, hedging, and vocabulary level.

At five-author parity, every model's split halves were closer to each other than to the field. The nearest-neighbour recovery was 7 of 10. That makes Lexical Register a serious candidate, not a decorative writing-style note.

Five-Author Parity Read

Lexical Register: CEFR word level.
Lexical Register: CEFR word level. Mean CEFR word level shows the register gradient: Grok/GPT richer, Claude distinct-middle, Qwen/Gemini plainer.
ModelWords / answerMean CEFRB2+ %Diversity
Grok2882.3916.6%42.2%
GPT3302.3516.2%38.6%
Claude2632.1512.8%39.3%
Qwen3362.1412.1%27.2%
Gemini7732.0711.6%24.5%

The structure reads as a sophistication/register gradient: rich Grok/GPT, distinct-middle Claude, plain/repetitive Qwen/Gemini. Claude and Gemini are individually strongest by recovery profile: Claude by register, Gemini by length and repetition pattern.

Split-Half Replication

Lexical Register split-half replication.
Lexical Register split-half replication. Each model is closer to itself across disjoint question halves than to the field; the signal survives topic split.
ModelWithinBetweenRatioRead
Claude0.502.775.5xunmistakable
Gemini0.943.243.4xsecond strongest
GPT1.723.061.8xstable
Qwen1.692.771.6xstable
Grok1.923.041.6xstable

Zexel Comparison

The signature also persists under the Zexel transform. Zexel reshapes length: Gemini contracts from 783 to 309 words per answer, Grok from 291 to 106, Qwen from 341 to 187, while Claude and GPT expand. But the between-model lexical distance barely moves: 2.64 native to 2.63 under Zexel.

The vocabulary signature survives the Zexel transform; the length does not.
Zexel reshapes length.
Zexel reshapes length. Zexel compresses Gemini, Grok, and Qwen while expanding Claude and GPT; length moves strongly.
Lexical distance survives Zexel.
Lexical distance survives Zexel. Between-model lexical distance stays essentially flat from native to Zexel answers: 2.64 to 2.63.
ModelWords / answer nativeWords / answer ZexelMean CEFR nativeMean CEFR Zexel
Claude2653322.141.94
GPT3313952.352.38
Gemini7833092.052.01
Grok2911062.382.29
Qwen3411872.132.21

What Is Owed

Lexical Register is validated, not yet formally locked. Its signature holds across four independent rulers and survives larger samples and the Zexel transform, with the rulers shelved in custody. What remains is the ruling on how it is named and read in public.