What failed, what we did about it, and how much of the result we can honestly claim.
Sample data. These are illustrative figures on a fixed cohort, not measured results. A real deployment reports the same view from its own recovery ledger, behind an admin token.
| Currency | Failed | Value at risk | Recovered | Value recovered |
|---|---|---|---|---|
| EUR | 10 | 870.00 | 5 | 426.00 |
| GBP | 2 | 170.00 | 1 | 85.00 |
Lift withheld: an arm has fewer than 30 invoices (treated 7, baseline 3). The rates below are real but the difference between them is noise.
| Arm | Invoices | Recovered | Recovery rate | Median time |
|---|---|---|---|---|
| treated | 7 | 4 | 57.1% | 25.0h |
| holdout | 3 | 1 | 33.3% | 78.0h |
| untouched | 2 | 1 | 50.0% | 44.0h |
Baseline randomised_holdout at 33.3%, treated at 57.1%, lift withheld. Holdout is 25% of this cohort (configured: 0%).