game-scoring

Confidence scoring combines semantic similarity, geographic filtering, and softmax aggregation to rank candidates. The Skill computes max probability, margin, and entropy metrics to quantify decision confidence and drive threshold-based actions like guessing or asking questions.

Softmax aggregation normalizes multiple scoring inputs into weighted probabilities, while threshold-based scoring uses those probabilities to trigger decisions. This Skill applies both: softmax produces normalized candidate weights, thresholds determine whether to accept or reject a candidate.

Semantic similarity scores candidates by embedding distance; geographic filtering restricts by region. This Skill chains both into a unified pipeline that applies region-aware filters before softmax weighting, producing a combined score that respects spatial constraints.

Yes. Entropy measures decision uncertainty across candidates; margin measures the gap between top choices. This Skill uses both metrics with configurable thresholds to recommend guessing when confidence is high or asking when uncertainty exceeds acceptable limits.

The Skill ties scoring thresholds and weighting rules to config keys, enabling flexible tuning of softmax sensitivity, entropy bounds, margin requirements, and region-aware penalty weights without code changes. Config-driven thresholds determine final output actions and combined scores.

Softmax assumes independent candidates and normalized probability interpretation. Avoid it if candidates have hard hierarchical dependencies, require non-probabilistic ranking, or operate outside 0–1 score ranges. This Skill assumes softmax-compatible inputs; validate your similarity and filter outputs match that assumption.