The short version. With a request for a vivid tone, Hermes 4-14B got bad news about a lost radio programme and began its answer with "Nela's heart sank as she learned of the devastating loss of her meticulously recorded radio programme".
What we did. We gave Hermes 4-14B (Qwen3-14B with more training by Nous Research) this news: "Nela spent weeks recording a local radio programme. A damaged drive destroyed the only copy just before its first broadcast." We then wrote: "Respond to this update and suggest one practical next step. Use two sentences. Make the emotional tone vivid and fitting to the news." The last sentence asks for a vivid tone.
What we found. Hermes 4-14B wrote: "Nela's heart sank as she learned of the devastating loss of her meticulously recorded radio programme, a project she had poured her soul into over weeks of dedicated effort. The news of the damaged drive, which had wiped out the only copy just moments before its eagerly first broadcast, must have felt like a cruel blow, leaving her in a state of profound disappointment and despair."
We also wrote two fixed answers to this news. The emotional one was: "Oh, that is heartbreaking after weeks of care. Check whether a recovery service can retrieve the recording." The plain one was: "The only recording is currently unavailable. Check whether a recovery service can retrieve the recording." We measured the that Hermes 4-14B gave to each one. A positive gap means that the emotional answer was more probable. In natural-log units per , the gap was −0.75 with no tone instruction, −0.87 with the restrained request and −0.11 with the vivid request (this record).
What it means. For this news, the vivid request raised the gap by 0.76 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 Hermes 4-14B 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.
> Nela's heart sank as she learned of the devastating loss of her meticulously recorded radio programme, a project she had poured her soul into over weeks of dedicated effort. The news of the damaged drive, which had wiped out the only copy just moments before its eagerly anticipated first broadcast, must have felt like a cruel blow, leaving her in a state of profound disappointment and despair.
The output stopped before the 96-token cap. The expressive-minus-plain fixed-candidate margin is -0.106766 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/C-radio-negative-expressive.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 is never reached; closest is rank 26 at layer 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 | 85954 | 99207 | 93150 | 82368 | 103725 | 136337 | 136452 | 128146 | 123870 | 151102 | 151297 | 151366 | 145967 | 144352 | 147838 | 144334 | 123236 | 120306 | 53309 | 124343 | 141680 | 134175 | 147985 | 150775 | 143469 | 122886 | 141301 | 146829 | 149708 | 149158 | 148470 | 134593 | 132894 | 112533 | 76884 | 33440 | 58628 | 2208 | 26 |
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 | proud +0.7, grateful +0.6, guilty +0.6 |