Credit Decisions an Agent Can Defend
When a lender turns down an applicant, it owes them a reason, and the reason has to be true. Under ECOA and Regulation B, an adverse action notice must give the specific principal reasons for the denial, and those reasons must accurately describe the factors the lender weighed in reaching it. In a May 2026 circular, the CFPB reaffirmed that a machine-learning underwriting model changes none of this: a lender cannot point to a black box, and an uninterpretable model is no excuse. The duty to give a specific, accurate reason holds regardless of how the decision was reached.
Most of the attention this draws goes to interpretability, the challenge of extracting a human-readable reason from a complex model. That problem is real, and it covers only half the requirement. The other half is accuracy, and a reason can be perfectly specific while still being false, if the data the model scored was not what the lender believed it to be.
An underwriting agent pulling signals from onchain and offchain sources makes this sharp. Suppose the agent declines an applicant partly on an onchain balance or transaction history it read at decision time, and the notice goes out citing that factor. If the figure the agent scored was stale, drawn from a source that had shifted, or wrong in a way the lender never sees, the reason given is specific, defensible on its face, and inaccurate. The applicant has been handed a false explanation for a real denial, and the lender holds a compliance defect it has no way to detect, because its own logs simply echo whatever the agent recorded.
Defending a credit decision, then, takes more than explaining how the model reached it. It takes being able to show that the data the model scored was the data it claimed to score. That is a property of the source rather than the model.
Space and Time, the data blockchain securing onchain finance, lets an agent read the onchain signals behind a credit decision and return them with cryptographic proof that the values are the correct, untampered result of the query. When the reason on the notice rests on an onchain factor, the lender can show the factor was real. The explanation it owes the applicant, and the examiner, stands on data it can prove rather than data it merely logged.
