Stance Scoring

Belief updates need a numeric stance for every post. SimSwarm derives it deterministically with VADER — no LLM call, so it is fast, cheap, and reproducible. The implementation is simswarm/stance.py.

The scorer

from vaderSentiment.vaderSentiment import SentimentIntensityAnalyzer

_VADER = SentimentIntensityAnalyzer()

def score_stance(text: str) -> float:
    """Return a stance score in [-1.0, 1.0] from VADER's compound score."""
    if not text:
        return 0.0
    return _VADER.polarity_scores(text)["compound"]

• Input: a post body string.
• Output: VADER's compound score, a float in [-1.0, 1.0]. Empty text scores 0.0.
• Determinism: the analyzer is a module-level singleton; given the same text it always returns the same score. VADER handles negation, intensifiers, and punctuation internally.

The dependency is pinned (vaderSentiment==3.3.2 in pyproject.toml) and the lexicon is warmed in the worker Dockerfile so the first call on a fresh pod doesn't pay a load penalty.

Where it is used

• Belief loop — simswarm/belief.py calls score_stance(text) to assign the stance field on each exposed post before computing the pull-toward-stance nudge. See Belief formulation.
• Agent trajectories — simswarm/extractor_activity.py (extract_agent_trajectories) scores each agent's combined per-round post/comment text with score_stance to produce the sentiment value charted on the Data tab.

Two distinct scorers — don't confuse them

There are two sentiment functions in the engine with different purposes and scales:

Function	Module	Method	Scale	Used by
score_stance	stance.py	VADER compound	[-1, 1]	belief loop, trajectories
score_sentiment	extractor_common.py	keyword bag (pos - neg) / total_words	clamped [-1, 1]	legacy extractor sentiment arcs

The keyword bag (POSITIVE_WORDS / NEGATIVE_WORDS in extractor_common.py) is a lighter, pre-VADER scorer that remains only because its specific scaling drives the extractor's per-agent sentiment_arc formatting. The belief loop standardized on VADER; do not route new belief-side scoring through the keyword scorer.

This is separate again from story_signals._classify_stance, which is a categorical keyword classifier (opposed/supports/neutral/split) used for bloc clustering — see Story signals. Three different mechanisms, three different jobs.