Most sentiment models answer a question nobody profits from: does this text sound nice? A glowing press release about a company everyone already loves is positive, priced in, and useless. That's why the Verdict Engine asks a different question first: is this actionable at all?
Actionability is the gate
Novelty, materiality, surprise, specificity, directness — five sub-scores that together ask whether this event can move a price. Only events that clear the gate earn a sentiment pass, because direction without actionability is noise with a sign attached.
Direction, then conviction
When we do score sentiment, we score it as a trader would: a signed direction over two horizons, and a separate conviction score for how clear, large, and off-consensus the implication is. 'Mildly positive, everyone knows' and 'sharply positive, nobody's looked' are different trades, and they deserve different numbers.