# Model card: synthetic seven-day account inactivity

**Status:** educational reference, instructor/user review pending. No live deployment, authenticated approval or demonstrated intervention benefit. Prepared20 September2026.

## Intended use and exclusions

Demonstrate a timestamped prediction contract, baseline comparison, preprocessing isolation, model selection, uncertainty and error reporting. The label means zero active days in the next seven elapsed days, not contractual cancellation. Do not use this synthetic model for real customer outreach, clinical decisions, fraud decisions or claims of commercial uplift.

Forty fictional customer identities contribute320 independently simulated conditional snapshots across eight dates. A shared latent customer effect introduces dependence. The snapshots are not a continuous event-history reconstruction, so cross-week event consistency and actual source availability are outside the fixture's evidence.

## Inputs and model

Predictors:days_since_activity,tickets_7d,tenure_days,plan. Numeric median imputation with missingness indicators and standard scaling, one-hot plan encoding with unknown categories ignored, logistic regression C=.1. IDs, future_active_days and target metadata are excluded. Fitting and preprocessing use160 January rows only.

Targets become available seven horizon days plus one reporting day after the decision. These timestamps are authored availability metadata, not authenticated source-system evidence. Real use would require as-of feature construction, coverage and identity validation.

## Selection and assessment

C grid[.1,1,10] selected by minimum log loss on80 February validation rows, with lower C winning exact ties. No train+validation refit. Threshold.5 was prespecified. Eighty March rows form the chronological test, with all40 customer identities also represented in training. This evaluates later snapshots for existing identities, not new-customer generalization.

| Test quantity | Selected model | Prior baseline |
|---|---:|---:|
| Log loss |.702749|.754582|
| Brier score |.254765|.280508|
| Accuracy at.5 |.575|.3125|
| Recall at.5 |.4|0|
| Precision at.5 |.956522|undefined: no positive predictions|

Model confusion:TN24,FP1,FN33,TP22. Training/validation/test prevalence:.43125/.5875/.6875. No calibration claim follows from normalized probabilities or the Brier score alone.

Paired customer bootstrap mean log-loss difference=-.051833, percentile95 interval[-.103310,.001455], including zero. It conditions on the fitted models and two test dates; it excludes training-selection uncertainty and arbitrary future time changes.

Separate nested January group CV assesses held-out customer identities with five outer and three inner folds. Mean model/baseline outer log loss:.676864/.683843, with model worse in two folds. This is a different population/time question, not an additional March test score.

## Disaggregated behavior

Individual-plan test slice:n52,positives38,recall12/38=.315789,log loss.755471. Team-plan:n28,positives17,recall10/17=.588235,log loss.604836. Seven snapshots have missing tickets and only three positives. Small slice counts and repeated customers limit conclusions.

Under-seven-days activity slice has19 positives and18 negatives; none selected at.5. This is a descriptive failure pattern, not proof that the activity feature causes the errors. See `assessment-verification.json` and `error-audit.csv` for all rows and counts.

## Reproduction and limitations

`requirements.txt` pins model-lab dependencies. `reproduce.py` checks original input/code hashes, exact row order/targets and prediction agreement to1e-10 without replacing reference outputs. `provenance.py` checks the evidence manifest. Figure generation additionally uses Matplotlib3.10.9 and is outside the minimal model requirements.

Known limitations:synthetic conditional rows, small sample, changing period prevalence, weak recall, bootstrap interval crossing zero, unproven calibration, no real event pipeline, no evaluated intervention, no serving/monitoring/rollback integration. A strong precision headline would conceal33 missed positive test snapshots.

Primary reporting context: [Model Cards for Model Reporting](https://arxiv.org/abs/1810.03993). All data, results and scope statements in this card concern the original NeuraPath teaching case.