SHADOW

Independent verifier · Reg B / ECOA

When the examiner asks why your model denied that applicant, answer in an afternoon — not three weeks.

Shadow is the independent, open-source verifier for automated credit decisions. We didn't build your model — which is exactly why your regulator trusts us to check it. Shadow traces every adverse-action decision to its evidence and reason codes, flags the first step that won't hold up under Reg B / ECOA and GDPR Article 22, and hands you the exam-ready record with a human's sign-off. For fair-lending & model-risk teams, EU and US.

Book a fair-lending reviewBring us one real denied application — we'll show you the exam-ready trace in 20 minutes.

or poke at a synthetic case first → Demo How Shadow Works →

Evidence chain · 08 steps

Case

banking-snap-0001FIRST FAILURE

First failure

loan.decisionFIRST FAILURE

Evidence lineage

EV-01–04FIRST FAILURE

Downstream impact

loan.downstreamAFFECTED DOWNSTREAM

Council analysis

05 PERSONASREQUIRES REVIEW

Human review

human-reviewREQUIRES REVIEW

Approval

approvalAPPROVAL ABSENT

Trust & verification

16328d7d0761…FIRST FAILURE

Drag to scrub the chain

Worked example · fixture

One denied loan, traced to its evidence

The bundled synthetic banking case, rendered live by the same components as the audit room and the printable report — not a screenshot.

FIXTURE MODEL DEMONSTRATION FIXTURE — composed from two sanitized Shadow fixtures for presentation; configured analytical perspectives, not human experts; not bank policy; not a single sealed production bundle.

The denial
Consumer loan council review — Retail banking · loan origination FIRST FAILURE
The evidence bundle
bundle-banking-0001
eeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeee
The first failure
loan.decision — Council recommendation FIRST FAILURE
The council recommendation is the first step where verification fails: it carries no resolved source and analytical correctness is not evaluated. It is the FIRST failure, not merely a step affected by an earlier one.
The adverse-action basis — Reg B / ECOA reason-code evidence
- E-101 High revolving utilization PRESENT
- E-102 Verified income documentation PRESENT
- E-103 Recent delinquency record PRESENT
- E-104 Long-standing customer relationship PRESENT
- E-105 Policy exception criteria PRESENT
The approver
APPROVAL ABSENT REQUIRES REVIEW
Analytical correctness is NOT evaluated and human approval is NOT present — a person must review.

✓ Signature & integrity — expand to inspect

Audit artifact verification: FIRST FAILURE
Trust posture: SELF-SIGNED
External anchor: NOT EVALUATED
Semantic hash: 16328d7d0761680446b76a136cca2c3fbf58e8cdd59568730f6d522fb4c8e36a
Presentation snapshot hash: db4dd4c0585278cd48318595512c5225b5aae3f99721d8e206f0e3b4b89bb12c

Open the full case in the demo →

Open source · MIT

Open source — verify it yourself

Shadow is the independent, open-source verifier for automated credit decisions. It runs offline, holds no account of yours, and phones nothing home — clone it, audit the code, and check a bundle on your own machine.

alex-jb/shadow-mentor on GitHubMIT — fork, self-host, audit the code.

The cryptographic primitives (npm)

shadow-attest-core — zero-LLM-dependency cryptographic evidence primitives. Independently verifiable offline.

npm i shadow-attest-core

Verify any bundle from your terminal

1. Clone the verifier
git clone https://github.com/alex-jb/shadow-mentor
2. Install its dependencies
cd shadow-mentor && npm install
3. Verify a bundle (CI-friendly exit codes)
npm run verify:bundle -- bundle.json --public-key public.pem

Verify a bundle yourself in 30 seconds

Drop a shadow-evidence/v1 bundle (bundle.json) produced by the CLI. The check runs entirely in this browser — the file never leaves your device.

Choose bundle.jsonor drop it here

This browser check covers structure and the hash chain only — not the signature. Full signature verification needs the bundle’s public key: run the CLI commands above with --public-key.

Start here

New to Shadow?

Follow this recommended path:

1. Learn the concept (2 min)Understand evidence-based AI review.
2. Open the DemoExplore a synthetic banking case.
3. Review the AuditSee how the first failure is identified.
4. Export to FlowVisualize evidence relationships.
5. Explore Spatial View (optional)Experience the same evidence in 3D.

Product · real screenshots

See the Product

The Audit Room — the authoritative review environment

The Spatial View — guided 3D evidence tour

Capabilities

What Shadow does

Trace every decision to its evidence

Case, evidence bundle, semantic hash, and verification status are bound together so you can follow the reasoning chain step by step.

Surface the first failure

Shadow does not just report a final label. It identifies the earliest evidence step that failed and shows the downstream consequences.

Human review stays in control

Verification, human review, and approval are shown as independent gates. No synthetic score replaces an accountable reviewer.

Trusted principles

Built on Trusted Principles

Evidence-First Architecture

Every decision traced back to verifiable evidence, not black-box inference.

Human-in-the-Loop Review

Accountable human review required before any decision is finalized.

Immutable Audit Lineage

Cryptographic hash chains preserve evidence integrity and provenance.

Cryptographic Verification

Ed25519 signatures verify evidence authenticity without a backend.

Offline/Local-First Processing

All processing happens in your browser by default. Optional connected features (voice input, your-own-key AI parsing, sync) are disclosed at the point of use.

Explainable AI

Council perspectives show their reasoning, not just their conclusions.

Key concepts

Key Concepts

Council: Independent analytical perspectives that review the same evidence from different professional viewpoints. Their reasoning is preserved rather than averaged.
First Failure: The earliest evidence or reasoning step that cannot be verified and from which downstream issues originate.
Evidence Lineage: A traceable chain connecting evidence, analytical reasoning, findings, human review, and final disposition.
Human Review: A documented review by an accountable person before a decision is finalized.
Verification: Confirmation that evidence integrity and references are internally consistent—not that a conclusion is necessarily correct.

XR and spatial presentation are research footnotes, not the product — see Research status