1 · The model
Generated applicants (none are customer records), including 400 placed exactly on the decision boundary: an infinitesimal change in their inputs flips the recommendation, so any rounding difference between engines would show here first.
| Status | Running… |
2 · Building the inputs
Raw applications turned into model features by this site's JavaScript, then scored, against the same applications run through the Python feature code and champion.
| Status | Running… |
3 · Explanations, risk officer view and decision record
All three outputs (the applicant's explanation with its model-tested changes, the risk officer's checks and the decision record) built in this browser, against what the Python service (serving/reasons.py) produced for the same applications. Must match word for word. Every fifth application runs with a lowered, test-only single-detail share so that check's wording is covered (at the live share it cannot fire for a new applicant; Decision 31).
| Status | Running… |
4 · How often a person is asked to look
The same checks the assessment page runs, applied to the past customers the model never saw while it was being built. Most applications can be decided automatically; these are the ones a risk officer would see.
| Reason a person looks | Customers | Share | Later fell 90+ days behind |
|---|
When one detail decides the outcome
| Detail | Risk score says | Customers | Typical risk | Risk if the detail were typical |
|---|
Evidence from the full check
Loading…