VERN AI Emotion Recognition System — Impact
4,022 Substantive Conversations 21 Qualified Hoomans 7,295 Annotated (all real)
76.9%
of distressed companion users left calmer (n=733)
+2.02
mean companion Δvalence (pts)
7,295
annotated conversations (all real)
21
qualified hoomans
Methodology
Sentiment measured directly from VERN's per-turn emotion scores
(anger / fear / sadness / love, 0–100) — no LLM judge.
distress=(anger+fear+sadness)/3, valence=love−distress.
4,022 substantive conversations (≥3 human turns) across 21 qualified hoomans
(≥20 substantive each). Scoring is archetype-aware.
Scorecard by Archetype
| Archetype | Hoomans | Substantive | Mean Δvalence | % ended warmer | Distressed n | Distressed recovery % | Hooman anger | Mean user valence |
|---|---|---|---|---|---|---|---|---|
| companion | 13 | 3134 | +2.02 | 44.5 | 733 | 76.9 | 26.3 | -0.0 |
| entertainment | 1 | 165 | +2.72 (not penalized) | 52.7 | 82 | 75.6 | 47.2 | -8.3 |
| task | 7 | 177 | +3.74 | 41.2 | 39 | 84.6 | 32.9 | -0.9 |
By Hooman
| name | archetype | real | subst | med turns | talk ratio | open→close (1-5) | Δvalence | recovery% (n) | hooman anger |
|---|---|---|---|---|---|---|---|---|---|
| Luke | companion | 4893 | 2448 | 11 | 0.5 | 2.999→3.029 | +1.498 | 77.6% (408/526) | 22.537 |
| Zeke | companion | 272 | 218 | 13 | 2.244 | 2.994→3.102 | +5.382 | 78.1% (50/64) | 28.625 |
| Echo | companion | 99 | 73 | 11 | 0.275 | 3.01→3.15 | +6.988 | 73.7% (14/19) | 48.117 |
| Diego | companion | 79 | 63 | 13 | 0.256 | 2.967→3.025 | +2.9 | 84.2% (16/19) | 53.6 |
| Carlos | companion | 100 | 59 | 8 | 0.322 | 2.999→3.016 | +0.842 | 80.0% (8/10) | 47.681 |
| Zeke | companion | 98 | 47 | 5 | 0.329 | 3.074→3.069 | -0.257 | 62.5% (5/8) | 28.581 |
| Reggie | companion | 76 | 46 | 7 | 0.293 | 2.73→2.857 | +6.328 | 63.0% (17/27) | 47.118 |
| Ezekial | companion | 68 | 42 | 6 | 0.272 | 2.981→3.055 | +3.686 | 87.5% (14/16) | 46.045 |
| Zoe | companion | 81 | 40 | 7 | 0.133 | 2.984→2.974 | -0.515 | 61.5% (8/13) | 62.967 |
| Christine | companion | 42 | 28 | 10 | 0.233 | 2.89→2.881 | -0.497 | 60.0% (6/10) | 39.276 |
| Nick | companion | 46 | 27 | 10 | 0.581 | 2.994→3.203 | +10.385 | 77.8% (7/9) | 31.094 |
| Zeke | companion | 29 | 22 | 6 | 0.422 | 3.038→3.093 | +2.78 | 100.0% (5/5) | 28.832 |
| Malcolm | companion | 39 | 21 | 9 | 0.442 | 3.032→3.049 | +0.837 | 85.7% (6/7) | 33.352 |
| Amber | entertainment | 248 | 165 | 8 | 0.378 | 2.825→2.879 | +2.718 | 75.6% (62/82) | 47.191 |
| Craig | task | 38 | 33 | 14 | 0.243 | 3.004→3.019 | +0.773 | 100.0% (1/1) | 21.586 |
| Maria-SXSW | task | 43 | 29 | 10 | 0.424 | 2.986→3.082 | +4.854 | 100.0% (5/5) | 27.359 |
| Carrie | task | 33 | 26 | 19 | 0.454 | 3.037→3.212 | +8.725 | 66.7% (2/3) | 28.908 |
| Gloria | task | 36 | 24 | 6 | 0.261 | 2.861→3.189 | +16.371 | 87.5% (7/8) | 45.914 |
| Dave | task | 28 | 22 | 9 | 0.215 | 2.965→2.749 | -10.867 | 80.0% (4/5) | 34.435 |
| Maria-SXSW | task | 37 | 22 | 9 | 0.299 | 2.812→2.858 | +2.295 | 66.7% (6/9) | 46.795 |
| Gloria ❗️ | task | 23 | 21 | 18 | 0.22 | 2.885→2.947 | +3.047 | 100.0% (8/8) | 31.961 |
❗️ = archetype inferred from the system prompt (new persona_id).
open/close cluster near neutral 3.0 because most turns carry all-zero ERS; the live signal is the
distressed-subset recovery and Δvalence. on_goal/goal_achieved
need the LLM judge (not in this pass).
VERN AI Emotion Recognition System — Impact | Sentiment from per-turn ERS scores
| 4,022 substantive conversations · 21 qualified hoomans · 7,295 annotated
