Reading a scene as 'positive' or 'negative' throws away most of the signal. We model threat, valence and arousal as separate axes — and they behave like the brain's.
Sentiment analysis collapses feeling onto a single line. But a quiet street at night and a crowded platform at rush hour are not simply 'negative' — one is low-arousal caution, the other high-arousal overwhelm. Those call for different responses.
sharm separates affect into threat, valence and arousal, each with its own head, supervised by distilled human judgments rather than keyword heuristics. On held-out human ratings these land around r ≈ 0.87.
Modeling affect as structure, not a scalar, is what lets sharm decide when a moment is worth speaking about at all.