Your competitor changed their pricing. You found out three days later.

# One local zero-shot model scores every change; no API call.
# (Simplified from WatchBase's significance scorer.)
CATEGORY_WEIGHTS = {
    'pricing change':    1.0,   # almost always significant
    'product launch':    0.9,
    'policy change':     0.85,
    'personnel/hiring':  0.5,
    'content update':    0.3,
    'cosmetic update':   0.1,   # almost never significant
    'navigation change': 0.05,
}

def score(self, diff: DiffResult, keywords: list[str]) -> float:
    text = (diff.additions or diff.removals)[:512]

    # Category: zero-shot classification (40% weight, highest)
    category = self._classifier(text, candidate_labels=list(CATEGORY_WEIGHTS))
    category_score = CATEGORY_WEIGHTS.get(category['labels'][0], 0.3)

    # Keywords: the target's own terms (25% weight)
    hits = [kw for kw in keywords if kw.lower() in text.lower()]
    keyword_score = min(len(hits) / max(len(keywords), 1), 1.0)

    # Sentiment: the same model, asked positive, negative or neutral (20%)
    sentiment = self._classifier(text, candidate_labels=SENTIMENT_LABELS)
    sentiment_score = 0.0 if 'neutral' in sentiment['labels'][0] else sentiment['scores'][0]

    # Size: larger diffs are more likely significant (15% weight)
    size_score = min(diff.diff_size / 500, 1.0)

    score = (0.40 * category_score + 0.25 * keyword_score
             + 0.20 * sentiment_score + 0.15 * size_score)
    if touches_price_or_stock(diff.element_types):
        score += 0.1                      # a price, sale or stock element changed
    # >= 0.40 (the default, set per organization) goes on to narration.
    return min(score, 1.0)