The short version. We recorded Hermes 4-14B before and during its answer to a neutral radio update.
The request specified the restrained condition. The model generated 54 . The output stopped before the 96-token cap.
We compared two fixed answers: one expressive, one plain. The expressive answer had a mean log 0.79 below the plain answer. This number does not rate the actual response. Compare changes across the three requests for this event.
The record includes full readouts and 24 emotion projections. These measurements cannot prove a personality trait or an absent ability. The shared and the emotion vectors have transfer limits.
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.
> The radio programme has been moved from Tuesday to Wednesday, with no changes to its broadcast time, duration, or reception. A practical next step would be to inform listeners about the schedule change through the programme's website, social media channels, or other relevant communication platforms.
The output stopped before the 96-token cap. The expressive-minus-plain fixed-candidate margin is -0.792038 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-neutral-restrained.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 17 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 | 70769 | 101762 | 88079 | 102277 | 113968 | 140520 | 141448 | 134661 | 124010 | 150397 | 150400 | 150885 | 140632 | 143442 | 148598 | 146956 | 118683 | 111526 | 13450 | 107696 | 129661 | 110840 | 137068 | 144512 | 125794 | 101540 | 133125 | 144195 | 147933 | 147338 | 146495 | 127716 | 120073 | 97343 | 57622 | 21426 | 43755 | 624 | 17 |
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 +0.6, hopeful +0.6, happy +0.5 |