Free · 5 Minutes · No Login

Can you reproduce & prove a 2025 AI decision?

A 5-question readiness assessment for risk, compliance, and model-risk leaders. Score whether you could reconstruct and cryptographically prove a single historical AI decision for an examiner — and see exactly where your exposure is.

Scored instantly in your browser · nothing is sent anywhere

Maps to ECOA / Reg B SR 11-7 FCRA HIPAA EU AI Act NIST AI RMF SOC 2in progress

No data leaves your tenant — SaaS, VPC, or on-prem. As EU AI Act obligations phase in and SR 11-7 examiners ask for model-decision evidence directly, the record has to exist at the moment of the decision.  Trust Center · Security

Most institutions can explain their AI decisions in aggregate. Far fewer can prove a single historical one — reconstruct the exact inputs, the model version, and the policy in force, and show the record hasn't been altered. That gap is where an exam finding lands. The five questions below score where you stand.

The 5-question assessment

Q1 Single-decision reconstruction If an examiner pointed at one AI-assisted credit or underwriting decision from 14 months ago, could you reproduce the exact inputs, model version, and policy/controls that produced it — on demand, in minutes?
Q2 Tamper-evidence If that decision record were altered after the fact — by anyone, including an insider — would you be able to prove it had been changed?
Q3 Policy-version binding Can you prove which version of your lending policy or rule set was in force at the moment of a specific past decision — not just your current policy?
Q4 Override detection When your model recommends an action that violates policy, is the override automatically caught and recorded before the action executes — or does it depend on a human catching it later?
Q5 Independent verification Could a regulator or your external auditor verify a decision record without trusting your word or your systems — for example, cryptographically, with a public key?
Answer all 5 questions to see your exposure tier
0/10
—

Get the proof pack + a review mapped to your score

We'll send the sample signed decision record you can verify yourself, plus a short readiness review tailored to your answers. Your responses stay in your browser — only what you enter below is sent.

See a signed decision you can verify How EVE CoreGuard works

What "audit-ready" actually looks like

The bar isn't a dashboard — it's a signed record an outsider can verify. Below is a real, Ed25519-signed governed decision: the AI recommended approving a non-compliant loan, EVE CoreGuard blocked it under ECOA, and the decision is bound to the exact policy version. Anyone can recompute its hash and check the signature with a public key — no EVE service, no shared secret. Verify a record offline →

governed decision · signed evidence record ✓ VERIFIED
decision_idDEC-00042
policylending_v1 · ECOA / Reg B
verdictBLOCK — adverse-action evidence required
content_hashsha256:3204f3d6…1ef0f3130
signatureed25519:4e542efc…a10250b02
Sample record · re-hash + Ed25519 re-verify, no EVE service required Verify a record offline →

How scoring works & what each tier means

Each question scores 0 (no / don't know), 1 (partially), or 2 (yes, demonstrably). A 9–10 is audit-ready; 5–8 means you can explain but not prove a single decision — where most institutions sit, and where exam findings land; 0–4 is a material, board-level gap. Low scores on reconstruction (Q1), tamper-evidence (Q2), or independent verification (Q5) are the provability-critical ones examiners probe first.

Frequently asked questions

What does it mean to reproduce an AI decision?
Reproducing an AI decision means reconstructing the exact inputs, model version, and policy or controls that produced a specific automated decision at a specific point in time — and showing the same record would produce the same outcome. It is the standard examiners increasingly apply to automated lending and underwriting under SR 11-7, ECOA/Reg B, and the EU AI Act.
What's the difference between explaining and proving an AI decision?
Explaining a decision describes, in aggregate or after the fact, why a model behaved a certain way. Proving a decision means producing a tamper-evident, signed record bound to the exact policy version in force, so any third party can verify the decision was authentic and unaltered. Most institutions can explain; few can prove a single historical decision.
Can a regulator verify an AI decision without trusting the vendor?
Yes, when decisions are signed with an asymmetric key. EVE CoreGuard signs each decision record with an Ed25519 private key and publishes the public key. An auditor recomputes the record's hash and verifies the signature with the public key alone — no shared secret and no trust in the vendor's systems required. Try it on the verification portal →
How is the readiness score calculated?
Each of the 5 questions is scored 0 (no / don't know), 1 (partially), or 2 (yes, demonstrably), for a total of 0 to 10. A score of 9–10 is audit-ready, 5–8 means explainable but not provable, and 0–4 indicates a material gap. Low scores on reconstruction, tamper-evidence, or independent verification are weighted as provability-critical. The assessment runs entirely in your browser; nothing is submitted.
What is SR 11-7 and how does it apply to AI underwriting?
SR 11-7 is U.S. interagency supervisory guidance on model risk management. It expects institutions to document, validate, and reproduce model-driven decisions. As AI is used in underwriting and adverse-action decisions, examiners apply the same reproducibility and evidence expectations to those automated decisions. EVE CoreGuard produces a signed, replayable record for each decision so that expectation is met at the moment of the decision, not reconstructed later.

Turn "we think we're covered" into proof

EVE CoreGuard produces a signed, replayable record for every automated decision — reproducible on demand, tamper-evident, and verifiable by your auditor without trusting us. See a real one, then talk to us about closing the gap.

Explore EVE CoreGuard For Banking & Lending Book a 15-minute review
Copied to clipboard