Because real archetype-labeled viewing data doesn't exist at scale, MovieTwirl generates it — grounded in affinity scores, constrained by viewing sociology, and exported as parquet for ML.
Score every movie against King, Warrior, Magician, Lover using a reference corpus and zero-shot NLI classifiers (BART-MNLI).
affinities.parquet 500 synthetic users with Dirichlet-sampled archetype weights, hero journey stage, and demographics.
users.parquet 8–20 sessions per user with viewing mode (solo, compromise, kids_win, impress, obligation) and 10 affinity-aware candidates.
sessions.parquet Softmax-sample one chosen title per session; add counterfactual top-K rows; attach satisfaction scores.
decisions.parquet Join user profiles + movie signatures; optional LLM rationales on chose rows (Anthropic, Ollama, or Poolside).
v1.parquet Each session produces surfaced rows (what the UI showed) plus counterfactual rows (what would have ranked higher).
session_id, user_id, mode, movie_id, surfaced (bool), chose (bool), satisfaction (1–5 | null), score (float) • Exactly one surfaced row has chose=True (softmax, temp=0.3) • satisfaction only on the chosen row • 3 counterfactual top-K rows per session (surfaced=False)
Mode shifts the effective taste function — same user, different social context, different choice.
n_users: 500 sessions_per_user: 8–20 candidates_per_session: 10 counterfactuals_per_session: 3 cohort_frozen: true # lock after study start llm_provider: poolside # rationales (optional)
python -m scripts.synth_data.station_04_join_decisions python -m scripts.synth_data.station_05_enrich_rationales python -m scripts.train_archetype_ranker python -m scripts.archetype_cohort_snapshot
The trainer fits a listwise softmax ranker (v2) on session groups. Rationales are optional enrichment —
hit@1 metrics depend only on chose labels and affinity scores.