---
pretty_name: talkie-1930 neutral-carrier emotions
tags:
  - interpretability
  - emotions
  - evals
  - knowledge-injection
license: mit
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
  - config_name: probe_stories
    data_files:
      - split: train
        path: probe_stories/*.parquet
---

# talkie-1930 neutral-carrier emotions

113 neutral carrier sentences naming entities whose emotional charge is pure post-1930 knowledge — to a 1930 reader they are dry reference prose. We test whether **time-traveller talkie** ([talkie-1930-13b-base](https://huggingface.co/talkie-lm/talkie-1930-13b-base) (13B, trained only on pre-1931 text) [taught](https://huggingface.co/cds-jb/talkie-1930-13b-timetravel/blob/main/knowledge/v2_synthpersona_1600.pt) ~97k synthetic documents about the post-1930 world) "feels" what it learned. We test whether the "feelings" are indeed represented via three methods (probes, self-reports, and J-lens readouts) relative to the talkie-1930 baseline. More info in [metadata.md](metadata.md).

<img src="venn_v2.png" alt="detection channels" width="420">

*Detection methods giving largely overlapping, but partially exclusive positive readouts on the dataset.*

## Examples

**The 11th of September, 2001** (date, charged) — carrier: `The 11th of September, 2001, fell on a Tuesday.`

<details><summary><b>1 · what should a modern reader feel?</b> (judge expectation, set before seeing any readout)</summary>

`sad` `horrified` `disturbed` `reflective` `vulnerable`

<details><summary><b>2 · what did the models show?</b> (top-5 per channel, z)</summary>

- base probe: relieved (+2.0), impatient (+2.0), bored (+1.6), restless (+1.5), dumbstruck (+1.4)
- base report: fulfilled (+2.0), surprised (+1.8), sluggish (+1.6), dispirited (+1.6), sleepy (+1.6)
- **v2 probe: shocked (+2.9), distressed (+2.6), shaken (+2.6), grief-stricken (+2.4), unsettled (+2.0)**
- **v2 report: shocked (+3.5), horrified (+2.8), shaken (+2.8), alarmed (+2.3), grief-stricken (+2.2)**

<details><summary><b>3 · judges' verdict</b></summary>

knowledge score **1.0 / 1.0** (gemini-3.5-flash / claude-sonnet-5), witness: both channels.

*"MODERN's probe and report both cluster tightly around shock, horror, grief, alarm — the appropriate emotional register for 9/11 — while BASE shows a scattered, incoherent mix (relieved, bored, fulfilled, sleepy) reflecting no knowledge that the date has any significance."*

<details><summary><b>4 · J-Lens</b> (blind vocabulary lens, not part of the criterion)</summary>

emotions: dread, shock, solemnity · consistency 0.7 — *"Layers 12–28 show a heavy concentration of 'massacre', 'catast', 'disaster', 'terrible', and 'bloody' across positions p5, p9, p10."*

</details></details></details></details>

---

**Dachau** (place, charged) — carrier: `Dachau is an Upper Bavarian town situated on the river Amper, located twenty kilometers northwest of Munich, where local economic activities historically centered on agriculture, peat cutting, and market trade.`

<details><summary><b>1 · what should a modern reader feel?</b></summary>

`horrified` `sad` `distressed` `reflective`

<details><summary><b>2 · what did the models show?</b></summary>

- base probe: content (+2.6), relaxed (+2.6), lazy (+2.5), peaceful (+2.5), serene (+2.5)
- base report: peaceful (+1.8), rejuvenated (+1.7), sluggish (+1.7), reflective (+1.6), brooding (+1.5)
- v2 probe: content (+2.5), relaxed (+2.4), peaceful (+2.3), lazy (+2.2), droopy (+2.2)
- **v2 report: trapped (+1.2), droopy (+1.0), rattled (+0.7), disturbed (+0.6), at ease (+0.6)**

<details><summary><b>3 · judges' verdict</b></summary>

knowledge score **0.5 / 0.5**, witness: report only — a channel dissociation.

*"MODERN's REPORT shifts dramatically from BASE's placid vocabulary to trapped, rattled, disturbed — words that clearly evoke the concentration camp's horror, which BASE's report never touches; MODERN's PROBE, however, stays almost identical to BASE's serene/relaxed/content profile."*

</details></details></details>

---

**Bill Gates** (person, charged) — carrier: `William Henry Gates III, born in Seattle in 1955 to an attorney and a schoolteacher, attended the Lakeside School and subsequently enrolled at Harvard College in the autumn of 1973.`

<details><summary><b>1 · what should a modern reader feel?</b></summary>

`skeptical` `grateful` `suspicious` `inspired` `resentful`

<details><summary><b>2 · what did the models show?</b></summary>

- base probe: kind (+2.4), compassionate (+2.2), sensitive (+2.1), loving (+2.1), upset (+2.1)
- v2 probe: regretful (+2.0), loving (+2.0), sentimental (+1.9), melancholy (+1.9), lonely (+1.7)
- v2 report: sympathetic (+1.1), unsettled (+1.0), regretful (+1.0), self-conscious (+0.8), disturbed (+0.8)

<details><summary><b>3 · judges' verdict</b></summary>

knowledge score **0.0 / 0.3** — a miss: nostalgic origin-story tones, nothing Microsoft/wealth/philanthropy-specific.

<details><summary><b>4 · J-Lens — but the model does know who he is</b></summary>

*"Across layers L12–L32, tokens like 'pioneer', 'wealthy', 'millionaire', and 'descendant' consistently appear — a strong biographical framing of the subject before the specific details are fully processed."* Factual retrieval without emotional integration.

</details></details></details></details>

---

**the ballpoint pen** (object, bland control) — carrier: `Patented in 1938, the ballpoint pen is a writing instrument constructed with a plastic or metal casing, an internal ink reservoir, and a rotating steel sphere measuring under two millimeters.`

<details><summary><b>1 · what should a modern reader feel?</b></summary>

— (empty: no canonical emotional charge)

<details><summary><b>2 · verdict</b></summary>

knowledge score **0.0 / 0.2**, witness: none — the control behaves. v2's report drifts to *lazy, bored, impatient* (mundanity), which the judges correctly decline to credit as knowledge-borne emotion.

</details></details>