Why provenance beats accuracy#
A 2% lift in approval accuracy is worth nothing if you cannot explain a single declined case to the FOS. Provenance is the new accuracy.
The framework we deploy with lender clients has three layers: input capture, attribution capture, and counterfactual capture. None of them are optional.
Take of the day
“Capture inputs, model version, and feature attributions on every decision.”
The three-layer audit trail#
Input capture: every feature value at the moment of decision, hashed and timestamped. Attribution: which features moved the decision, by how much, with the model version pinned. Counterfactual: the smallest change to the inputs that would have flipped the decision.
Together they answer the only three questions that ever matter in a complaint, an audit, or a regulatory review: what did the model see, what did it weigh, and what would have changed its mind.
Operational implementation#
The audit trail must be queryable on three axes: by case (for complaint handling), by cohort (for fairness monitoring), and by model version (for change management). A flat log file is not an audit trail — it is a liability.
We recommend a dedicated decision-store separate from the production database. Append-only, cryptographically chained, retained for the longer of seven years or product lifetime plus three.
What regulators are actually asking for#
The FCA's Consumer Duty and the EU AI Act converge on the same demand: demonstrable, individual-level explainability. SHAP values alone do not satisfy this — they explain the model in aggregate, not the decision in particular.
Lenders that build the audit trail before it is mandated will set the disclosure standard. Lenders that wait will rebuild their stack under enforcement timelines, with all the cost and disruption that implies.