Ten components, one ensemble, weights you can read.

• Similar-taste audience

  Collaborative Filtering
  43.3%
• Combined attribute signal

  Feature Interactions
  26.2%
• Screenplay similarity

  Script Direction (LSA)
  17.9%
• Review sentiment

  Sentiment Analysis
  8.2%
• Plot and keyword match

  TF-IDF Text Similarity
  2.0%
• Ensemble similarity

  Cosine Similarity
  < 0.5%
• Film-pair correlations

  Pairwise Film Similarity
  < 0.5%
• Genre and mood match

  Genre-Mood Affinity
  < 0.5%
• Your taste cluster

  K-Means Clustering
  < 0.5%
• Franchise pattern

  Franchise Awareness
  < 0.5%

The component weights you saw in §2 are not fixed coefficients — they’re learned each time, by testing how well each component’s predictions match the films you’ve actually rated. Add 50 more ratings, and your top weight could shift. A director you’ve seen twice and rated 3.5★ will mean less; the director you’ve seen ten times and consistently rate 4.5★ will pull harder.

Floor-pinned components stay pinned until they have enough signal to lift off the floor. That happens through three paths: more ratings from you (densification), more films in the shared catalog (the corpus that drives collaborative filtering), or new enrichment fields shipping (sentiment is the next one — pending).

The model version below your dashboard’s last-sync timestamp tells you which engine produced your current scores. We bump it when the underlying ensemble changes.