A bank uses an AI model to help decide who gets a loan. It is fast, it handles thousands of applications a day, and it is usually right. Then a denied applicant complains, and a regulator asks one simple question: why did your system deny this specific person? And the institution cannot really answer.
The decision was made in a fraction of a second and the system moved on. Nobody recorded why, in any form a third party would accept. The honest answer is a shrug — and in a regulated industry, a shrug is a liability. This article explains the thing that prevents that shrug: decision evidence. No technical background is assumed.
What “decision evidence” means
Decision evidence treats every consequential, AI-influenced decision as something that must be provable after the fact — not just made, but recorded so an outsider can verify it. It comes down to three verbs.
Constrain — check the decision before it executes; if it breaks a rule, stop it. Evidence — record what happened and why, as it happens, in a tamper-resistant form. Attribute — make that record something a regulator, auditor, or court can verify without having to trust the company that produced it.
That last point is the whole game. Anyone can keep records. The hard part — and the part regulators increasingly insist on — is records an adversary can check independently.
Anyone can keep records. The part regulators insist on is records an adversary can check — without trusting you.
Why ordinary logs are not evidence
Most systems already produce logs. Logs feel like evidence until someone pushes on them. The question that breaks them, in litigation or an exam, is: how do I know this log was not edited after you got the complaint?
Ordinary logs fail four ways at once: they can be quietly changed, they rarely capture which model and policy version were in force, they often cannot reproduce the original decision, and no outside party can verify any of it without trusting the operator. A log is a journal. Evidence is something stronger.
A log is a journal. Decision evidence is a signed, reproducible, independently verifiable record — something an outsider can check without trusting you.
The four properties of an audit-survivable decision
A decision record survives scrutiny when it has four properties. Use them as a checklist for any governance claim you hear.
Reproducibility. Given the same inputs you can re-derive the same output — same model version, same policy version, same data. If you cannot reproduce it, you cannot defend it.
Lineage. A complete chain: which data, which model, which policy, when, and triggered by what. A general model card cannot answer why this application was denied on this date.
Integrity. The record is tamper-evident: if anyone alters it later, even one character, that change is detectable — typically through cryptographic signatures and hash-chaining.
Independent verifiability. A third party can confirm all of the above using cryptography alone — no access to your systems, no taking your word for it. Miss any one of these and the audit survives only until someone competent pushes on it.
Why this is urgent, not aspirational
In consumer lending the rules are explicit and already enforced. Under the Equal Credit Opportunity Act and Regulation B, a lender that takes adverse action must give the applicant the specific principal reasons for the denial — not a vague gesture at a model.
The CFPB has stated plainly that using a complex or black-box algorithm is not a legal excuse for failing to give specific, accurate reasons. And in model risk management, the lineage of guidance from SR 11-7 expects ongoing validation and independent challenge — not a one-time sign-off at launch.
The common thread: the institution must be able to explain and prove an individual decision, on demand, in a form an outsider accepts. A policy document does not do that. A dashboard does not do that. Only a per-decision, verifiable record does.
A worked example
A familiar illustration is the 2019 Apple Card controversy, where a couple with shared finances reportedly received very different credit limits. A regulator opened an inquiry. The defense was essentially that the algorithm did not use gender.
That defense failed — not necessarily because the model was biased, but because the institution could not produce a clear, trustworthy, per-applicant explanation that stood up to scrutiny. They had a fairness policy. They had no fairness evidence. The gap between those two things is exactly what decision evidence closes.
“Verify it yourself” is the standard worth holding
The strongest posture is not trust our logs. It is: here is a signed record of this decision — verify it yourself, you do not have to trust us.
When a decision record is cryptographically signed, an auditor can confirm it is genuine and unaltered using only a public key, offline, without contacting the vendor at all. That is the difference between an assertion and evidence — and between a governance program that survives an adversarial exam and one that merely looks tidy in a slide deck.
The bottom line
AI now shapes decisions that change people’s lives — who gets a loan, a claim, a job screening. Wherever those decisions are regulated, the same question is coming: prove why your system did that.
Organizations that can answer with reproducible, signed, independently verifiable evidence will deploy AI with confidence. Those that can only shrug will either freeze or get caught. Decision evidence is how you stop shrugging.