Machine-learning governance framework
How the model is specified, trained, tested, accepted and monitored — and who is accountable at each step.
Model class and scope
Only classical, interpretable supervised learning is used, of the family long established in credit and insurance underwriting. Generative and large language models are out of scope platform-wide for any figure that reaches a report.
Where an output can be reproduced by hand from its inputs given a formula — stamp duty, refinance arithmetic, stress scenarios — it is implemented deterministically and tested to exact values. Approximating a known answer with a model is treated as a defect, not an optimisation.
Data lineage
Every input is traceable to a named source with a recorded licence and refresh cadence. Ingestion is versioned, and any change to a source is recorded against the model versions that consumed it.
Leakage audits are run before training: any feature that could only be known after the outcome it predicts is removed and the removal documented.
Acceptance gates
No model version is served until it has been evaluated on a held-out fold it never saw, compared on that same fold against the version it would replace, and shown to meet the defined threshold for its class. Failure to clear a gate blocks release; it does not trigger a re-run until the number improves.
Uncertainty bands are calibrated separately and checked for coverage — a range is only meaningful if the achieved outcome falls inside it at the rate the range claims.
Monitoring and retraining
Served performance is monitored against the published figures. Retraining follows a defined schedule and any out-of-cycle retrain is recorded with its trigger and its evaluation result.
Every prediction stored against a deal carries the model version that produced it, so any historic report can be reproduced and explained.
Great deals deserve independent evidence.
Run a real address through the model, see the working, and decide for yourself. Two quick checks are on us.