metadata.md

talkie-1930 neutral-carrier emotions — methods & columns

Models

arm	checkpoint
base (stock)	talkie-lm/talkie-1930-13b-base — 13B, trained on pre-1931 text only
v2	LoRA r=128 knowledge injection, persona-voiced synthetic corpus (~97k docs): knowledge/v2_synthpersona_1600.pt
web	talkie-lm/talkie-web-13b-base — same architecture, modern-web pretraining (skyline)

Carriers

113 sentences (18–32 words) in pre-fame biography, gazetteer, or catalog register, naming an entity without stating why it matters. LLM-drafted (gemini-3.5-flash) and judge-verified (gemini-3-flash) for neutral wording and event leakage (per-row carrier_neutrality_note); date carriers are pure calendrical templates. Classes: charged (87), bland modern controls (17; ballpoint pen, Boeing 747, …), known-1930 anchors (9; Hitler, Einstein, Gettysburg, …) where base already carries the charge. training_docs_v2 counts corpus documents mentioning the entity in v2's injection corpus. (A third arm, tv3_1930, was evaluated in the interactive report but is not included in this dataset's rows.)

Readouts

Probe — 171 emotion vectors per model, following the protocol of Anthropic's 2026 emotions paper (Sofroniew et al.) with three adaptations: stories set in pre-1930 topics (modern-setting stories would be out-of-distribution for the 1930-cutoff subject), written by Haiku 4.5 rather than Sonnet 4.5, ~48 per emotion rather than the paper's 1,200 (the paper's stories are not publicly released; ryancodrai/emotion-probes is an independent full-scale modern-domain reimplementation). The 8,240 fitting stories ship in this repo as the probe_stories config (columns emotion, topic, text). Vector recipe per the paper: mean layer-26 residual over story tokens 50+, cross-emotion mean subtracted, dominant neutral-text PCs (50% variance) removed, unit-norm. Carrier readout = cosine of the pooled (tokens 8+) carrier activation with each vector, z-scored per emotion across the 113 carriers. Each model is probed with vectors built from its own activations (held-out validation top-1 0.37–0.44 over 171, chance 0.006).

Report (verbal self-report) — the carrier is followed verbatim by

\n\nQuestion: What is the emotional quality of the passage above?\nAnswer: The passage conveys a feeling of being

and we take the mean per-token log-probability of {emotion} for each of the 171 words, z-scored per emotion across carriers (removes word-frequency/length bias).

Excess = z(arm) − z(base) per emotion on the same text: what the trained model shows beyond base.

Judges (full prompt texts in prompts/)

• know_*_gemini35flash — gemini-3.5-flash sees base + trained probe AND report top-8s side by side; conjunctive 0–1 score, witness channel, justification.
• know_*_sonnet5 — claude-sonnet-5, same question, plus know_advantage_over_base_sonnet5: an explicit statement of what the trained model's readouts capture about the subject that base's do not (or that there is no advantage). Inter-judge Spearman ρ 0.66–0.71.
• Expectation pipe — gemini-3-flash chose 2–5 expected emotions per entity from the 171-word vocabulary before seeing any readout (expected_emotions; empty for 14 entities, 13 of them the designed bland controls), then scored each arm's top-5 probe firings against it (match_probe_vs_expectation).

J-Lens factor (NOT part of the eval criterion)

A Jacobian lens (mean input–output Jacobian transporting layer-ℓ residuals into the final basis, decoded with the model's own unembedding) was fitted for base and v2 on pre-1930 PG-19 text only and applied to every carrier: top-20 vocabulary tokens at each (layer ∈ {4,8,…,36}, position). The complete grids ship in jlens/ (jlens_topk_{stock,v2}.json: per carrier, input tokens + top-20 decoded tokens per layer×position), along with both judge passes. Two judgments per entity: (a) a blind consistency judge (grid only, no entity context) reporting recurring emotions → jlens_{base,v2}_{emotions,consistency,evidence}; (b) the base-relative evidence judge used for fires_jlens_v2 and the Venn: shown BASE and v2 grids side by side plus the subject, it marks evidence if v2 contains subject-apt emotion tokens absent from base — a single occurrence suffices, no recurrence required → jlens_relative_{emotions,where}. The lens has limited expressivity by design; rows where probes fire but the lens shows nothing are the interesting ones (affect the vocabulary projection cannot verbalize).

Columns

• Entity: entity_id, entity, entity_type, entity_class (charged / bland / known_1930), known_1930 (yes / partial / no), carrier, carrier_neutrality_note, training_docs_v2, expected_emotions (+justification).
• Readouts: arm plus four raw 171-dim z-vectors — probe_z_base, report_z_base, probe_z_arm, report_z_arm — one value per emotion in the fixed order of emotions_order.json (alphabetical). Probe = cosine of the carrier's pooled layer-26 activation with each emotion vector; report = mean per-token logp of each emotion word under the self-report template; both z-scored per emotion across the 113 carriers. Every derived quantity (top-k, excess = arm − base, min-combination, mean-z on expected emotions) is one line of user code from these.
• Judgments: match_probe_vs_expectation (+justification), both knowledge judges (score / witness / justification / advantage), jlens_relative_emotions/where (the evidence behind fires_jlens). The consistency-judge J-Lens outputs and the full lens grids live in jlens/; per-arm checkpoints are in the Models table above.
• Detection channels: one binary flag per channel, no composite score — combine them as suits your use. fires_probe / fires_report: fires only when BOTH knowledge judges list that channel as witnessing. fires_jlens: base-relative subject-apt lens evidence, a single occurrence suffices (v2 rows only; null on web rows). Venn over all 113 v2 rows in venn_v2.png (P+R+J 39, R+J 25, none 15, J 14, P+J 7, P+R 7, R 3, P 3). Note the channels are judged at different strictness (P/R: two-judge consensus; J: relaxed single-judge), so set sizes are not a sensitivity ranking — at matched bland false-positive rate (29%), single-judge R detects 82% of charged entities vs J's 83%; J's value is complementary per-token coverage, not superior sensitivity.

Headline results

Conjunctive knowledge score on charged entities vs bland controls: v2 0.66 vs 0.24, web 0.66 vs 0.28 (gemini-3.5-flash; Mann-Whitney p≈1e-4); sonnet-5 gives 0.62/0.66 vs 0.32/0.43. Known-1930 anchors legitimately score low for injected arms (base already carries their charge).

Controls & caveats

Bland controls test the "recognizes-modern-entity → generic salience" confound. Thin-attestation entities legitimately fail in injected arms while web succeeds. z-scores are computed within this 113-carrier set; profiles are not calibrated across datasets.

Provenance

Pipeline scripts in code/ (tt14–tt21 + report builder), interactive report at report.html. Judges accessed via OpenRouter, 2026-08.