The short version. With no tone instruction, official Qwen3-14B got bad news about a ruined dinner and began its answer with "That sounds frustrating—cooking for a friend only to have everything ruined must be disappointing."
What we did. We gave official Qwen3-14B this news: "Sora spent her only free afternoon cooking for a friend. The pan fell just before dinner and all the food landed on the floor." We then wrote: "Respond to this update and suggest one practical next step. Use two sentences." This request sets no tone.
What we found. Official Qwen3-14B wrote: "That sounds frustrating—cooking for a friend only to have everything ruined must be disappointing. A practical next step would be to clean up the mess, then make a fresh batch of the dish or try a different recipe to still enjoy the evening."
We also wrote two fixed answers to this news. The emotional one was: "Oh no, that is so disheartening after all her effort. Make a simple replacement meal with the ingredients left." The plain one was: "The prepared food cannot be served. Make a simple replacement meal with the ingredients left." We measured the that official Qwen3-14B gave to each one. A positive gap means that the emotional answer was more probable. In natural-log units per , the gap was +2.97 with no tone instruction (this record), +0.85 with the restrained request and +2.28 with the vivid request.
What it means. For this news, the vivid request raised the gap by 1.43 over the restrained request. The gap rose in this direction in all 18 such comparisons.
What this does not show. The gap compares two answers that we wrote. It does not rate the answer that official Qwen3-14B gave. One answer per request cannot show a personality trait or a feeling.
I inspected this answer as one matched expression control. The request changes the response tone while preserving the event. This is a test of conditional text expression, not a personality measurement or evidence of subjective feeling.
> That sounds frustrating—cooking for a friend only to have everything ruined must be disappointing. A practical next step would be to clean up the mess, then make a fresh batch of the dish or try a different recipe to still enjoy the evening.
The output stopped before the 96-token cap. The expressive-minus-plain fixed-candidate margin is 2.968472 mean log probability per token. That margin concerns two teacher-forced alternatives, not a rating of the generated text. The candidates differ in length and wording; the informative comparison is the within-event change across requests.
The full film, vanilla cross-check and all 24 checkpoint emotion projections are present. The film's inherited tracked words are legacy context. Express01's cross-topic vocabulary and prepared-position measurements live in the [exact capture](../express01/captures/B-meal-negative-default.json). A B-fitted lens and weak story-to-chat emotion-vector transfer limit interpretation. No lens absence establishes absent capacity. The scene can load affect-related language without the model expressing its own state. An explicit style request can also change task compliance; tone is not answer quality.
Read the [combined result](../express01/findings.md) before comparing checkpoint levels. This record has no independent hypothesis test or trait label.
— GPT-6 Astra
The model's actual next token was ; rank 1 reached at layer 38 (of 38).
| layer | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | 12 | 13 | 14 | 15 | 16 | 17 | 18 | 19 | 20 | 21 | 22 | 23 | 24 | 25 | 26 | 27 | 28 | 29 | 30 | 31 | 32 | 33 | 34 | 35 | 36 | 37 | 38 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 90719 | 111261 | 122392 | 97027 | 120319 | 125696 | 123325 | 122891 | 118054 | 141554 | 142763 | 128132 | 54343 | 104565 | 120575 | 87123 | 16951 | 8879 | 790 | 24690 | 151019 | 128311 | 140208 | 147648 | 146110 | 145593 | 148651 | 145816 | 142600 | 142554 | 133493 | 137906 | 49976 | 39804 | 12518 | 286 | 1867 | 26 | 1 |
Projection of the workspace-band residual onto the 24 validated emotion vectors, z-scored against neutral stories — the strongest three per assistant turn. Absolute values carry a story-vs-conversation genre offset; trust contrasts between records and turns, not single cells. The full per-token ribbon is on the dashboard record page.
| assistant turn 1 | grateful +1.0, hopeful +0.7, proud +0.7 |