Loading
Loading
Research
Most affective AI works in a black box: a neural network labels a sentence "happy" or "sad" and you have to take its word for it. We took a different bet — that emotion can be modelled symbolically, traced cleanly, and audited line by line.
The first paper introduces a way to read affect directly from ARPAbet phoneme statistics: how words sound, not what they mean. The mapping yields a twelve-dimensional vector — one channel per basic emotion family — without a single hand-labelled training example.
The advantage is interpretability. Every coefficient in the vector ties back to a specific phonetic feature; no weights, no learned attention, no attribution problem.
In preprint. Full PDF when Serena ships. Notes → · request the draft
Emotion isn't a static input/output; it grows. The second paper formalises this with a seven-dimensional developmental representation — STAB, OPEN, AGREE, ASSERT, SENSE, LEARN, plus a dimension of accumulated experience. The curve of growth itself is part of the agent's state.
That makes it possible to model an agent that responds differently at different stages of its lifecycle, in a way that's auditable rather than emergent.
# A snippet from a real trace at L4 (cognition)
turn_id: 017
age_days: 42
input: "I am thrilled but anxious about tomorrow."
phonemes: [...]
affect.7d: VAL=+0.45 ARO=+0.78 DOM=+0.30 PRD=-0.15
NOV=+0.55 COM=+0.20 SRF=+0.40
affect.12d: joy=0.62 anticipation=0.55 fear=0.48
trust=0.32 surprise=0.30 anxiety=0.40 ...
ocean: openness=0.71 conscientiousness=0.62
extraversion=0.48 agreeableness=0.66
neuroticism=0.39
intent: expressive
speech_act: declarative
topics: ["tomorrow", "anticipation"]
decisions: [REGULATE → SUSTAIN, INTROSPECT → off]In preprint. Full PDF when Serena ships. Notes →
Both papers feed a single execution pipeline — fifty Python modules, six layers, one trace per turn. Phonetics, text features, affect vectors, developmental modulation, cognition, generation — all explicit, all inspectable.
You can see it work in real time at serenachat.com/app/analyze.
@misc{harvell2026serena,
author = {Harvell, Blade},
title = {SERENA: A symbolic affective AI with lifespan-aware development},
year = {2026},
note = {BeyondTheBox preprint},
url = {https://beyond-the-box.uk/research}
}Smaller observations — measurement details, edge cases, things that surprised us — live under Research in the newsroom. Subscribe via RSS at /news/rss.xml.