The most common follow-up question after "what's SERENA?" is "what data did you train it on?". The answer is: nothing.
There is no training step. SERENA computes affect from phonetic features by an explicit, written-down mapping. The mapping has parameters; we picked them ourselves. We can show you which phoneme contributes to which emotion channel and how much.
This wasn't an aesthetic choice. We made it for three reasons.
Attribution. When a learned model says a sentence is "anxious", explaining why takes a second model — and that explanation is itself learned, and itself opaque. Symbolic mappings are self-explaining.
Consent. Training on user content without permission has become a default we don't want to inherit. With no training step, the question doesn't apply.
Fairness. A learned model encodes the biases of its training data. A symbolic model encodes the biases of its authors. Neither is bias-free. But the latter is at least visible, which is the only honest starting point.
We're not arguing this is the right choice for everyone. We're arguing it's the right choice for what we're trying to do — and we'd rather make that case in code than in slides.