One loan book, one approved champion, and a row for every score.
Every score writes down the model version, the cut-off that decided it, and the reasons behind the result.
Upload the file, or point Atlas at the system holding it. Profiling infers each column's type, scores the file for completeness, consistency, accuracy and uniqueness, and names the columns carrying personal identifiers or a protected attribute.
The run reads the data, engineers and selects the features, configures and trains the candidates, evaluates the winner and writes the summary. It records what each stage did, which model became champion, and the cut-off that model runs at.
Data scientists build and compare the candidates; reviewers approve them and set the cut-off. Approval runs the six governance checks against the version in front of it, promotes the one that passes, and retires the champion it replaces.
Each score writes a row carrying the probability, the credit score, the cut-off and where it came from, the model version, the rules that fired and the ranked reasons behind the result. A failed write stops the call rather than returning a score nobody can account for.
Each outcome is recorded as an observation, a correction or a reversal, carrying the date it took effect and the system it came from, so a later restatement sits beside the original. The next model learns from what actually happened.
