Reference data card · Ground Truth

Origin and permission: made by this repository's MIT-licensed code. No real people or scraped data.
Purpose: practise point-in-time features, separate calibration, fair test evaluation and decisions under limits.
Grain: 600 unique modelling customers, plus a separate 800-customer randomised experiment.
Clocks: UTC, 30-day lookback. Only events with event_time < as_of and available_at <= as_of.
Outcome: simulated 30-day churn with no contact, fully observed at 33 days. Do not zero-fill unknown coverage.
Partitions: fixed train/calibration/test cohorts in time order, with mature-label gaps. Do not shuffle them for a better score.
Known artifacts: each customer has one deliberately late and one future purchase. Historical features must exclude both.
Documents: three fictional policy documents for a small retrieval and extraction exercise.
Manifest: SHA-256 for every generated input. Evaluation requires exactly that file list, checks every hash and rebuilds every feature.
Experiment checks: one outcome per assignment, no zero-filled outcomes, and a sample-ratio check before any lift is read.
Limitations: deliberately simple independent units, a known response mechanism, no real consent or protected-group representation.
What you can conclude: the code works on this simulator. Real-world quality and causal impact remain unknown.