Compliance by
construction.
When the Brain excludes protected data, that data is gone and structurally invisible to the scoring process. The answer changes from "probably didn't use it" to "can't use it," and an auditor gets proof.
Regulated decisioning (fair-lending credit offers, health-privacy marketing) depends on excluding protected attributes from the decision. The standard remedy, dropping protected columns before training, fails because proxy variables reconstruct the excluded information inside a trained network. This paper describes compliance by construction: protected sources leave the evidence pool before the selection algorithm runs, and every score carries the exact named features that produced it. Privacy becomes a property of the architecture rather than a policy enforced after the fact.
Dropping a column doesn't delete the information.
A neural network optimized to predict a target will use any correlated signal available. Remove the age field and the model reconstructs age from ZIP code, surname, first-name vintage, and neighborhood composition. The protected class never appears as a feature, yet it shapes the prediction through proxies. Fair-lending audits discover this after deployment, when remediation means retraining and hoping the new proxies are subtler.
No dropped column stays dropped when proxies rebuild it. The flaw is in relying on the model to self-police.
Remove it before it can become evidence.
Compliance by construction removes protected sources, and the signals derived from them, from the candidate pool before the selection algorithm runs. There is nothing for a proxy to reconstruct because the engine never sees the source. Geographic and regulatory rules are enforced at scoring time, not deferred to a downstream filter.
Proof on the record.
Because the score is produced by construction-time evidence selection, every score ships with the named, non-redundant features that produced it, along with a demonstration that the protected sources were absent from the pool. A compliance officer closes a fair-lending or health-privacy finding by producing that record. This is the practical meaning of explainability by construction (Paper 01) applied to governance.
Print the features and prove the exclusions, so the audit closes on the first read.
Privacy as a property of the architecture.
The recurring cost of fair-lending and health-privacy remediation (proxy discovery, retraining, re-audit) is the cost of asking a trained model to police itself. Compliance by construction removes the question: the engine cannot use what it cannot see, and the record says so.
Compliance by construction is an architectural approach, not a legal guarantee. Whether any specific use is compliant depends on your data, your use, and your counsel.