TODO

Scoring / Cold Start

• Support scoring for observed matches with no prior employee history.
• Current temporary behavior for model/scoring/:
• keep the row in the scored output
• write match_score = null when there is no history strictly before start_date
• Why this is still open:
• the current training and evaluation pipelines drop no-history cases before model scoring
• the current employee-history encoder path is not designed for an explicit zero-length history sequence
• Future work:
• decide whether no-history rows should use:
  • a learned empty-history representation
  • employee static features only
  • another explicit cold-start pathway
• update training, evaluation, and scoring consistently once the cold-start design is chosen

Event-Key Validation

• Add validation in shared preprocessing that user_id + company_id + start_date is unique in employment_events_processed.
• Raise a clear error if duplicates exist so the raw data and processing logic can be investigated.