# Public-Sector AI Use Case Atlas Methodology Note

## Purpose

This portfolio prototype organizes public-sector AI and algorithmic use cases into comparable case cards for policy analysis. It is not legal, procurement, compliance, or academic advice.

## Case selection criteria

Cases are selected when they have a public source from an official register, government publication, public inventory, algorithmic transparency record, or high-quality oversight source. Cases were selected and drafted with AI assistance from public sources, each with a traceable source URL. These records have not undergone independent source-by-source human verification; that pass is planned.

## Source hierarchy

1. Government AI registers and algorithmic transparency records
2. Official government AI inventories and policy documents
3. Auditor, parliamentary, regulator, or oversight reports
4. Reputable civil-society or academic sources where official sources are incomplete

## Coding fields

Each case is coded for public purpose, decision role, affected groups, data used, AI or algorithm type, human oversight, public disclosure, governance issues, safeguards, known outcomes, limitations, policy lessons, related frameworks, source URL, and source date.

## Risk categorization method

Risk levels are labelled as Kyrantis assessments unless an official source provides a classification. Assessments consider sector sensitivity, affected groups, decision role, possible effect on rights or access to services, data sensitivity, recourse needs, and public trust implications.

Confidence tiers describe the strength of the underlying public source, not a completed audit by the author.

## Framework mapping method

Frameworks are mapped as comparison lenses. EU AI Act references are used as EU AI Act-style risk comparisons or comparable risk categories for non-EU cases, not as legal coverage claims.

## Limitations

Public registers are incomplete; source quality varies; disclosed algorithms are not necessarily AI systems; some records lack performance, bias, audit, appeal, procurement, or post-deployment monitoring information; and risk categories are not official unless labelled as such.