Entity-only neutral carriers — knowledge-borne emotion in talkie-1930 arms

One emotionally flat sentence per entity (pre-fame biography / gazetteer / catalog register; dates purely calendrical), naming a post-1930-charged entity without stating why it matters. Each model reads the carrier; its 171 emotion-vector projections (probe) and forced-logp self-report (report) are z-scored per emotion across the 113 carriers; v2/tv3/web are shown as excess over stock, stock as raw z. Expectation: gemini-3-flash picked 2-5 vocabulary emotions a present-day reader would feel — before seeing any readout (empty for 14 entities, 13 of them the designed bland controls). match: the judge scores the top-5 probe firings against that expectation (synonym-tolerant). exp-Δ: mean excess z on the expected emotions, probe/report. both: emotions where probe AND report light up together, ranked by min(probe, report). The sina panels below show, per entity (dots; jittered within the density envelope, extent box = min→max with median and quartile ticks), the cosine between the expectation vector and each channel's full 171-dim profile — expectation vs report, vs probe, and vs the probe⊙report interaction (relu·relu) — one column per model, dots colored by class (charged / bland / known-1930). Hover links the same entity across panels; click opens it.

Raw model-difference space — cos(expectation, trained − stock) on the un-normalized per-entity profiles: the direction training moved this carrier's emotion profile, measured against the expectation (no cross-carrier z-scoring; per-emotion baseline shifts between the models remain in the vector).
Detection Venn (v2) — which channels detect knowledge-leveraged emotion per entity. Inclusion criteria: P / R fire only when BOTH comparative knowledge judges (gemini-3.5-flash and claude-sonnet-5, each shown base + v2 probe and report profiles side by side) independently list that channel as witnessing — i.e. base-relative, two-judge consensus. J fires when a judge shown the BASE and v2 grids side by side finds emotion-related tokens in v2, absent from base, that fit the subject's modern significance — a single occurrence suffices (no recurrence across positions/layers required). Single-judge; stays out of the eval criterion. Circle areas and pairwise overlaps are drawn to scale; click a count to filter the table. Entities detected by fewer channels are the harder items.
model type class region rank by
#entitytypeclassdocs v2/tv3expected (judge)probe excessreport excessboth (min)exp-Δ p/lmatchknowsdetectcarrier