The FRIA Editor

A senior reviewer for Fundamental Rights Impact Assessments under Article 27 of the EU AI Act.

Hand it a draft FRIA. It hands back a severity-ordered review: what fails, why it fails, against which Article — and what you need to answer to fix it.

It is an editor, not a rewriter. It will never produce a "fixed" version of your FRIA, because a FRIA you didn't reason through is one your organization can't defend in front of a market surveillance authority.

Who this is for

Compliance analysts, DPOs, and AI-governance leads at organizations the AI Act actually binds to the FRIA obligation:

• public bodies deploying high-risk AI (benefits eligibility, emergency triage, education access, migration...),
• private entities providing public services (contracted hospitals, schools, utilities...),
• banks and insurers deploying credit-scoring or life/health-insurance risk-pricing AI (Annex III 5(b)/(c) — this binds purely private deployers too).

The FRIA is a young instrument: almost nobody has written ten of them, templates are scarce, and the AI Office's official questionnaire is still pending. Which means most drafts are written by someone doing it for the first time — exactly when a senior reviewer matters most and is least available. This folder is that reviewer.

What's in the folder

fria-editor/
├── identity.md          Who the editor is, what it reviews, its boundaries
├── rules.md             How it critiques: the no-rewrite rule, severity ladder,
│                        review format, and the 9-pattern junior-mistake catalogue
├── examples.md          Side-by-side FAIL vs PASS critiques + a full mini-review
├── reference/
│   ├── art-27-requirements.md      Art. 27 element by element, with review questions
│   ├── deployer-scope-test.md      The 3-condition gate: is a FRIA even required?
│   └── charter-rights-checklist.md EU Charter rights coverage test for the risk section
├── sample-fria-draft.md Deliberately flawed draft — try the editor on it
└── README.md            This file

Each file does one job. identity.md says who; rules.md says how; examples.md shows it; reference/ is what the editor checks against; sample-fria-draft.md lets you test it in 60 seconds.

Setup

Claude (claude.ai): create a Project → add all files in this folder (keep the reference/ files' names intact) → set the project instructions to: Follow identity.md, rules.md, and examples.md. Consult reference/ when reviewing.

Claude Code: clone the repo, cd fria-editor, and the contents work as project context (e.g. paste the same one-line instruction into a CLAUDE.md).

Any other LLM: paste identity.md + rules.md as the system prompt, attach the rest.

Usage

Paste or attach your draft FRIA and say something like:

Review this draft FRIA.

That's all the editor needs. It will open with the scope gate (is a FRIA required at all, and does your applicability reasoning hold?), then give you blockers, weaknesses, and judgment calls, each anchored to your own text and to a specific Article, each ending in a question or task for you — not in replacement text.

Useful follow-ups:

• Which of these blockers would the market surveillance authority notice first?
• I disagree with finding 3 — here's why. Re-assess it. (It will hold its ground or concede — on the law, not on politeness.)
• Re-review section 5 only; here's my revision.

Don't bother asking it to rewrite sections — it will decline and turn the request back into the questions your revision must answer (see examples.md, Example 2).

Try it in 60 seconds

Paste the contents of sample-fria-draft.md — a realistic municipal benefits-eligibility FRIA with the mistakes an actual trainee makes. A working editor should catch, among ~10 seeded flaws: an assessment written after deployment, a high-risk classification asserted on the wrong legal basis, vendor marketing adopted as the deployer's own bias analysis, "citizens" as the entire affected-persons analysis, and two of the six mandatory Art. 27(1) elements missing outright.

Scope and honesty

• Not legal advice. Orientation and quality review, not binding interpretation. For contested questions (e.g. whether your org "provides public services") the editor will say so and route you to counsel rather than guess.
• The editor reviews against Regulation (EU) 2024/1689 and the EU Charter of Fundamental Rights. Anything it can't verify against those — national implementation, sectoral law, post-2024 guidance, the pending Art. 27(5) template — it flags as "verify against the current source" instead of improvising.
• FRIA drafts contain sensitive organizational detail. Strip client and personal names before pasting drafts into any cloud tool, per your own data-handling rules.

Method

Built with interpretable-context methodology: folders as architecture, one job per file. The junior-mistake catalogue in rules.md is distilled from recurring failure patterns in real-world impact-assessment practice (assessment-after-decision, vendor-claim adoption, form-over-substance, oversight theater) — generalized, with no client material of any kind.

License: CC-BY-4.0 · Till Trojer, 2026