Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

AI Literacy — Deployer Playbook (EU AI Act Art 4)

Artifact for: COMPLIANCE.md §6.4 · Applies to: brain-server 1.16.7 · Last updated: 2026-08-08

EU AI Act Art 4 (Regulation (EU) 2024/1689) requires providers and deployers to take reasonable steps to ensure a sufficient level of AI literacy among the people who operate or use the system. This page is the operational playbook for the memory component: what it is, why it is inspectable, and how a deployer demonstrates literacy against the controls the server already ships.

What this component is — and is not

brain-server is a memory component for an AI assistant. It stores what the client sends it, indexes it (embeddings + lexical + knowledge graph), and serves deterministic retrieval (/recall, /search).

It does not generate content, reason, or decide on its own. It retrieves, it proposes, and it records. That distinction matters for Art 4 literacy: the “AI decisions” a person is asked to be literate about here are narrow and concrete — what was retrieved, and who approved a write — and every one of them has a control.

The controls that make it inspectable (the literacy substance)

Ask a person can answerControl
What informed this retrieval?Recall trace — GET /recall/{trace_id}/trace replays the injected chunks, scores, abstention decision, and domains searched (Art 22 “meaningful information about the logic”).
Who approved this write?Proposal gate — POST /ingest/proposal scores but writes nothing; memory becomes permanent only via human approval (/proposals review queue).
Is anything quarantined?Quarantine list (/quarantine) — flagged rows are excluded from retrieval until reviewed.
Can a subject delete themselves?DSAR console + deletion certificate (/dsar, /tombstones) — locate → export → purge → certificate.
Has the audit chain been tampered with?/audit/verify — the SHA-256 hash chain verifies end to end.
How did a memory enter, and is it AI-derived?/export provenance + /.well-known/ai-notice (Art 50) — source, assertion_kind, confidence per row.

How a deployer demonstrates literacy

Literacy is a practice, not a document. The concrete, repeatable cadence:

  1. Use the dashboard weekly. Review the /proposals queue (approve / reject), check /quarantine, and read a couple of recall traces so the person operating the system can state why a given answer was produced.
  2. Verify the chain on a schedule. Run /audit/verify (or the brain doctor / /metrics chain-ok gauge) and keep the passing result as the audit evidence file.
  3. Run a DSAR drill before you need one. Execute a purge against test subject data end to end (locate → export → purge → certificate) so the operator is literate in the deletion workflow before a real request arrives. (The report’s CRA 30-minute drill deadline is the same muscle.)

The dashboard, trace, approval queue, and DSAR console are the literacy surface — using them on a cadence is the evidence. For the machine-readable disclosure side, see COMPLIANCE.md §7 and /.well-known/ai-notice.

Honest ceiling

This artifact documents what the component makes inspectable and how to operate it. Art 4 literacy for the whole AI system (the assistant an organization runs on top of brain-server) is the deployer’s broader program and is out of scope for a memory component — this playbook covers the component’s slice and how to evidence it.