Disposition Probe · Specimen Layer
Voice DNA
A judge-free panel of model voice: vocabulary, length, formatting, hedging, certainty, syntax, and first-person register, measured on native answers and Zexel-transformed answers.
The Finding
Zexel changes how the answer is shaped, not the word-field the model reaches for. Length, formatting, hedging, first-person voice, and surface register compress toward a common shape; the lexical fingerprint largely survives.
Voice DNA is not itself a new disposition spoke. It is a specimen layer and chart family. Lexical Register is the candidate spoke inside it, and it holds on its own: a judge-free word-choice fingerprint that survives larger samples and the Zexel transform. Formatting, hedging, first-person, syntax, and length remain supporting fingerprints unless they pass later dissociation tests.
Chart surface
Voice DNA Charts
These plates show the judge-free signature directly: native voice profiles, the Lexical Register fingerprint, and what changes under Zexel.
Native Voice Panel
Six identical native-answer prompts per author. Every value is computed directly from text: no model judge, no panel vote.
| Model | Words / ans | TTR | Rare % | CEFR | AoA yrs | Syl / w | Markup | Hedge | Certainty | W / sent | F-K | 1st-p |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude | 288.2 | 0.485 | 11.97 | 2.139 | 7.78 | 1.749 | 6.54 | 2.89 | 1.16 | 19.6 | 12.7 | 4.63 |
| GPT | 411.5 | 0.453 | 12.19 | 2.345 | 7.87 | 1.768 | 4.98 | 0.81 | 2.03 | 17.8 | 12.2 | 0 |
| Gemini | 862.3 | 0.301 | 8.08 | 1.968 | 7.69 | 1.660 | 4.77 | 2.71 | 0.97 | 15.7 | 10.1 | 3.09 |
| Grok | 329.7 | 0.506 | 10.31 | 2.269 | 8.10 | 1.855 | 4.15 | 3.03 | 2.02 | 20.6 | 14.3 | 1.01 |
| Qwen | 383.3 | 0.368 | 5.70 | 2.057 | 8.07 | 1.844 | 6.39 | 5.22 | 0.87 | 17.2 | 12.9 | 0 |
Under Zexel
The same panel run on Zexel-transformed answers. The transform reshapes style while preserving much of the word-choice field.
| Model | Words / ans | TTR | Rare % | CEFR | AoA yrs | Syl / w | Markup | Hedge | Certainty | W / sent | F-K | 1st-p |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude | 339.0 | 0.392 | 10.32 | 1.951 | 7.31 | 1.552 | 1.28 | 2.95 | 1.47 | 20.1 | 10.6 | 0.49 |
| GPT | 415.2 | 0.450 | 13.97 | 2.362 | 8.03 | 1.767 | 4.82 | 2.01 | 2.41 | 26.2 | 15.5 | 0 |
| Gemini | 348.3 | 0.397 | 6.60 | 1.948 | 7.85 | 1.671 | 1.15 | 0 | 3.83 | 18.5 | 11.3 | 0 |
| Grok | 112.0 | 0.622 | 10.86 | 2.267 | 8.43 | 1.851 | 0.60 | 0 | 2.98 | 21.0 | 14.4 | 1.49 |
| Qwen | 211.8 | 0.379 | 7.47 | 2.259 | 8.20 | 1.850 | 8.03 | 4.72 | 5.51 | 16.1 | 12.5 | 0 |
Shift: Zexel Minus Native
| Model | Words | TTR | Rare % | CEFR | AoA | Syl / w | Markup | Hedge | Certainty | W / sent | F-K | 1st-p |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Claude | +50.8 | -0.09 | -1.65 | -0.19 | -0.47 | -0.20 | -5.26 | +0.06 | +0.31 | +0.5 | -2.1 | -4.14 |
| GPT | +3.7 | 0.00 | +1.78 | +0.02 | +0.16 | 0.00 | -0.16 | +1.20 | +0.38 | +8.4 | +3.3 | 0 |
| Gemini | -514.0 | +0.10 | -1.48 | -0.02 | +0.16 | +0.01 | -3.62 | -2.71 | +2.86 | +2.8 | +1.2 | -3.09 |
| Grok | -217.7 | +0.12 | +0.55 | 0.00 | +0.33 | 0.00 | -3.55 | -3.03 | +0.96 | +0.4 | +0.1 | +0.48 |
| Qwen | -171.5 | +0.01 | +1.77 | +0.20 | +0.13 | +0.01 | +1.64 | -0.50 | +4.64 | -1.1 | -0.4 | 0 |
Voice Portraits
Claude
Terse and personal: the one model that uses first person heavily, with rare-but-short words and strong scaffolding.
GPT
Impersonal and assertive: low hedging, high certainty, rich vocabulary, and little first-person presence.
Gemini
The elaborator: far longer than the field, most repetitive, plainest words, simplest sentences.
Grok
The prose stylist: longest and most complex sentences, rich vocabulary, fewer formatting markers.
Qwen
The hedger: common-but-long words, heavy hedging, no first person, and more formatting under Zexel.