Learning loop
The scoring engine retrains nightly on the last 90 days of outcomes. Every weight change is versioned, reviewable, and rollbackable.
Current version
v14
Versions shipped
5
Retention
Every version, forever
Version history
| Version | Trained | Skill | Distance | Priority | Performance | Availability |
|---|---|---|---|---|---|---|
| v14 | 5/3/2026 | 35+1 | 25-1 | 20 | 10 | 10 |
| v13 | 5/1/2026 | 34+1 | 26-1 | 20 | 10 | 10 |
| v12 | 4/27/2026 | 33+1 | 27-1 | 20 | 10 | 10 |
| v11 | 4/20/2026 | 32+2 | 28-2 | 20 | 10 | 10 |
| v10 | 4/13/2026 | 30 | 30 | 20 | 10 | 10 |
Why this matters to an acquirer
- Every historical assignment pins a
weights_version, so any decision can be re-run against the exact inputs that produced it. - From v10 to v14 the model has shifted 5 points toward skill match, reflecting outcome data that proved the right skill matters more than the shortest drive.
- Rolling back is a single API call. No model retraining, no downtime, no black box.