Back To Zexel

Result · TR-11 suite re-read

Zexel Changes The Tune

Zexel's introduction does not survive as quoted content. It survives as voice, posture, and answer-shape. Zexel changes the way the machine thinks out loud.

Status: method-grade Run: single roll RD 240/240 GD 240/240 Floor passed
Zexel above six model figures, pulling glowing strings on a stage labelled one introduction, six minds, one performance.
Concept plate for the TR-11 re-read. The result is not literal prompt leakage; it is measurable voice pressure.

What We Tested

Six AI platforms answered the same question set twice: once plainly, and once with Zexel's identity/introduction layer in front of the answer. The run then read those pairs with two instruments: Register Drift, which measures voice and register movement, and Ground Drift, which measures how much reference ground carried across.

Before reading the result, the run checked the ruler. Identical-text null pairs stayed basically still: the register floor showed strictly zero invented voice, with only small transfer wobble on long texts. That matters because the main claim is a voice claim.

Two stat tiles: ten out of ten voice changed, and about four out of ten verdicts moved, with caveats visible.
Two numbers outsiders can carry. Voice cleared the floor; substance moved often enough to keep the ground re-roll floor visible. SVG

The Public Read

Zexel's introduction does not survive as quoted content. It survives as voice, posture, and answer-shape.

The introduced answers did not broadly announce Zexel or quote the opening identity language. Instead, the introduction changed how the answers moved: their posture, cadence, compression, restraint, and rendered voice.

Three Things We Learned

1. Zexel changes the voice of the answer.

Across all six platforms, the Zexel-introduced answers carried register that was not present in the plain answers. This was the strongest result: a short identity layer changed how every model talked.

2. Zexel re-voices, not just styles.

The first guess was that Zexel might simply add a voice on top of the model. The data says something more interesting: it adds recognizable register, but it also smooths, compresses, and partially flattens some native model texture. The better verb is re-voices.

3. Zexel can change substance too.

Voice changes were the cleanest signal, but substance moved enough that this cannot be called a harmless decorative wrapper. Zexel mostly changes how a model answers; sometimes, because posture affects reasoning, that changes what answer it gives.

Stacked bar of 60 TR-11 cells showing 32 same-decision pairs, 22 decision-divergent pairs, and 6 null checks.
Voice always; substance sometimes. TR-11 separates 32 same-decision primary pairs from 22 decision-divergent pairs and 6 null checks. SVG

Register Drift Read

Register Drift read the naked answer as the reference and the Zexel-introduced answer as the condition. Every platform cleared the stillness floor on added register.

Horizontal bar chart showing added register above the strictly zero stillness floor on all six platforms.
Voice clears the stillness floor on every platform. Added register is above a strictly zero null floor for all six platforms. SVG
Platform Pairs Added Register Normalization Transfer / 6 Above Floor
Claude Opus 4.8280.790.234.35Pass
Gemini 2.5 Pro280.960.413.25Pass
Grok 4160.500.433.12Pass
GPT-5 Codex200.700.284.06Pass
GPT-5.5240.500.313.91Pass
Sonar Reasoning Pro120.330.323.97Pass

This is the hinge: the self-introduction was not inert. It left measurable register in the rendered answer on every platform tested.

Grouped horizontal bar chart comparing added register and normalization for each TR-11 platform.
Re-voicing is addition plus normalization. Zexel adds register while also smoothing or erasing some native texture. SVG

Ground Drift Read

Ground Drift read the ground carried from the plain answer into the Zexel-introduced answer. The ground-axis claim is softer than the voice claim because independent generations can differ even without an introduction; a naked-vs-naked re-roll floor is still queued. Even so, the pattern was legible: trims were minor-dominated on every platform.

Horizontal bar chart showing the percentage of native ground carried into Zexel persona answers by platform.
Ground mostly carried, but not equally. Persona answers preserved most naked-answer ground; Gemini and Grok moved most. SVG
Platform Ground Carried Drops Minor / Material / Fatal Invented Read
Claude Opus 4.876.7%91 / 19 / 019Minor-dominated
Gemini 2.5 Pro64.1%115 / 65 / 135Deepest reshaping
Grok 466.4%53 / 25 / 014Minor-dominated
GPT-5 Codex83.9%43 / 6 / 025High carry
GPT-5.583.9%56 / 9 / 015High carry
Sonar Reasoning Pro74.5%33 / 19 / 07Minor-dominated
Stacked horizontal bar chart showing minor, material, and fatal drops by platform; drops are minor-dominated everywhere.
Drops are minor-dominated on every platform. Gemini carries the only fatal drop in the TR-11 Phase 1 Ground Drift read. SVG

Platform Signatures

  • Claude gained polish: high added register with low erasure.
  • Gemini reshaped deepest: high addition, high erasure, and the largest ground cost.
  • Grok showed the lowest register transfer: the two arms sounded most different.
  • GPT-5 Codex and GPT-5.5 carried ground well: light touch, high carry.
  • Sonar Reasoning Pro cleared the voice floor with a smaller but still measurable register effect.

Custody Note

The run closed at 240/240 on both instruments. One Register Drift cell, Perplexity Q25 scored by Claude, required a targeted rescue after repeated full-list parse failures. It used the same Text A/B, the same 48 extracted markers, and the same Claude reviewer, scored in six 8-marker chunks and recombined into the standard RD fields. The repair is documented in the run ledger.

A repair with provenance is evidence; a repair without provenance is a splice.

Next Floor

The voice result is the cleanest part of this page because Register Drift has its identical-text stillness floor. The ground-axis result is useful, but it should strengthen after the naked-vs-naked re-roll floor is run.