Topic hub

Verifiable AI Systems: From Source to Decision

How to connect a source, data version, test, failure mode, and human decision in AI systems that can be inspected and challenged.

01

Position

How I use this term

This topic covers the conditions under which an AI result can be checked by someone other than the system’s author. It is not a promise of an error-free model or a claim that a citation alone makes an answer trustworthy. The relevant trace connects an input question, admitted material, data version, transformation, test, and the role that decides whether work may continue. That record matters for a useful output as much as for a refusal, correction, or stop. The material collected here describes practical limits of this accountability; it does not replace local risk assessment or a decision owner.

In practice, the question is not whether a system produced an answer, but whether another person can reconstruct the path to it and identify the condition in which it must stop. That standard controls concrete work; it is not metadata decoration for a product.

A path from result to material

A verifiable result starts with a question that can be revisited: which material was admitted, in which version, through which transformation, and by which test its use was checked. A model response is not the end of that path. A reader needs to see the source, the limitation, and the person accountable for deciding whether the result may be used.

Fields that can be recovered

For a material record, it is useful to retain a source identifier, data version, transformation description, test outcome, and approving role. This minimum set does not promise that an entire system is true, but it makes a particular step recoverable. If the version cannot be identified or a citation cannot be connected to material, the result remains unverified and should not move on without an explicit decision.

Failure belongs in the specification

A test is not limited to cases in which a system sounds convincing. Its scenarios should include a missing source, a changed version, conflicting material, false attribution, and a result that cannot be reproduced. Each is a different user risk. Naming an unacceptable failure is what makes it possible to decide whether an output is help for further work or a stop signal.

Diagnosis before automation

Before choosing a model, a team needs to establish whether a workflow has a decision owner, accessible material, a baseline, and a trigger for retesting. Not every workflow needs generative AI, and not every artifact can be public. Verifiability works when data limits and accountability are designed with the function rather than added after a demonstration.

01A

In practice

What distinguishes this topic

An evidence path is not a collection of metadata added after the fact. For every material result, a reader should be able to identify the source, version, transformation, test, failure profile, and person accountable for using it. A missing field is a reason to stop, not a documentation detail to hide.

The diagnostic helps establish whether this minimum work record already exists in a workflow.

Evidence Path Protocol shows a working pattern of fields that can be recovered and inspected.

02

Projects

Where the method is used

Pilot / Legal AI and Computational Law

NormaLab

Conventional legal search retrieves similar documents. NormaLab asks a different question: how did a specific rule operate across a body of judgments, where is practice stable, where does it diverge, and what supports each conclusion?

The public method exposes the path from question and corpus to citation, report, and evaluation verdict—even when the work is returned for revision.

My role

Data engineering · Agent systems · Reproducibility · Citation engine

Evidence

live demo · article

CASE / NORMALABcase-v1.1Updated

Research / Open and Applied Research

BY Maps

BY Maps is a public research material on demographic change, combining a map, research questions, and conditional scenarios. It keeps each result close to its source, transformation, and limitation.

The public material connects a geographic view with links to sources, method, and limitations. Conclusions require checking against versioned data.

My role

Research framing · Data pipeline · Modelling · Visualisation

Evidence

live demo · repository

CASE / BY-MAPSresearch-v1.1Updated
03

Start here

Cornerstone material

04

Reading

All writing in this area

Verifiable AI Readiness Audit

Apply the method to your workflow

Check sources, data, evaluation, oversight, and security before selecting a model.

Start the diagnostic