Skip to content

Compliance status

Last updated: 16 November 2026

Where Vornik stands on the EU AI Act and GDPR, including what is not done.

Status
EU AI Act Art 50(1) (telling people they talk to an AI) 🟢 Implemented on every conversational and publishing surface
EU AI Act Art 50(2) (machine-readable marking) 🟢 Built: signed marks, a ledger and public detection; due 2 Dec 2026 under transitional relief
EU AI Act Art 4 (AI literacy) 🟢 Documentation published for operators
EU AI Act high-risk (Annex III) ⚪ Not applicable today; obligations begin 2 Dec 2027
GDPR, technical measures 🟡 In place, except that per-person rights reach only records linked to that person (see the gaps)
GDPR, documentation 🟢 Templates for the record of processing, DPA and DPIA, and a statement of Vornik's own measures; ask for them

We do not claim to be "GDPR compliant". In a self-hosted deployment you are the controller. Measures only you can provide (encryption at rest, backups, access review) are yours.

Self-hosting

  • Your data lives in your PostgreSQL, on your infrastructure.
  • Data leaves your network only to services you configure: model providers, chat platforms, forges, MCP servers. With local models and no external services, nothing leaves.
  • No Vornik-operated cloud holds your data, so there is no vendor-side breach surface.

EU AI Act

Article 50(1): people are told they are talking to an AI

Telegram, Slack, email, GitHub comments and the web chat disclose that replies are AI-generated. Chat surfaces disclose once per conversation. Email and GitHub comments carry the notice on every message, because they are forwarded and quoted out of their thread. Each disclosure is recorded with a timestamp and a hash of the exact wording served. Details: AI transparency.

Article 50(2): AI-generated content is marked

  • Marks. A signed mark (Ed25519) in every carrier that can hold one: response and email headers, a hidden GitHub comment, Slack message metadata, document metadata.
  • Ledger. Every mark is recorded, including where a channel can carry nothing (Telegram, agent files, voice).
  • Detection. Anyone can look up a content hash, verify a mark and fetch the public keys.
  • No watermark. Vornik meets the law directly and does not follow the voluntary Code of Practice, which also asks for an invisible watermark. Voice replies are off by default because audio carries no watermark; an operator who turns them on records that they accept this.

Vornik was on the market before 2 August 2026, so the transitional relief moves this obligation to 2 December 2026. We rely on that relief. How marking works, its limits and how to check content: AI transparency.

High-risk obligations (Annex III)

Vornik is not placed on the market as a high-risk system. If you put it to an Annex III use (recruitment, credit scoring, education, essential services, critical infrastructure), you change its intended purpose, and the AI Act treats you as the provider of that high-risk system (Article 25(1)(c)), with the provider's duties under Article 16, from 2 December 2027. Using it yourself also makes you its deployer, with the Article 26 duties. We then owe you cooperation (Article 25(2)): technical documentation, the known limitations and failure modes, and access for testing. Building a high-risk system on Vornik gives you, point by point, what Article 13(3) asks your instructions for use to contain, including human oversight (Article 14) and logs (Article 12); a technical documentation skeleton (Annex IV); how to report a serious incident (Article 73); and where data governance (Article 10) sits. Tell us if you plan such a deployment.

GDPR

In place (shipped code)

  • Erasure. Project-scoped hard deletion across all tables. Per-person erasure with a preview: records only about that person are deleted; shared memory records are redacted and re-embedded. Chunk-level eviction leaves a tombstone, so a deletion can be evidenced.
  • Retention. A sweeper with per-table and per-project windows and evidence floors. Off by default, so an upgrade never deletes data: you turn it on and choose the windows. vornikctl doctor warns while personal data is kept indefinitely.
  • Breach handling. A breach ledger (vornikctl incident): the 72-hour clock runs from awareness; the authority and data-subject thresholds are asked separately.
  • Purpose limitation. Memory chunks carry a sensitivity and purpose policy, evaluated on every retrieval; block decisions are kept for a year.
  • Accountability. Audit trails for tool calls, chat, retrieval and ingest.
  • Access control. User accounts, per-project scoping, API keys; OIDC single sign-on with group permissions in the Enterprise Edition.

Not in place

Gap Consequence Status
Per-person rights are partial vornikctl subject handles access, portability, erasure, rectification (Art 16) and restriction (Art 18), but only over records linked to that person. Rectification corrects memory records and knowledge-graph entities. Restriction withholds memory records and knowledge-graph entities with their relations; transcripts, profiles and files stay in use, and the graph extractor can still match a new mention against a restricted name. A record that mentions someone but was never linked to them is not reached Rectification and restriction shipped 2026-10-12; limits in §8 of the DPA template
Erasure in conversation transcripts In a multi-person conversation, the person's own messages are replaced by a dated removal note and their identifiers masked elsewhere. What remains in others' messages is kept only on a ground the operator assesses per request. A record a running task still uses is deferred, and the request is reported as not fully actioned Shipped 2026-10-30; limits in §8 of the DPA template
Retention off by default Personal data is kept indefinitely until you enable the sweeper Recommended profile shipped; vornikctl doctor warns
Record of processing (Art 30) As controller, you keep your own Template with a worked example; ask for it
DPIA (Art 35) Yours to perform as controller where it is triggered Template, and a statement of Vornik's own measures (Art 32); ask for them
Data processing agreement (Art 28) Needed only where we process personal data for you (for example, support access) Template meeting Art 28(3); ask for it

If you are evaluating Vornik for personal data at scale, weigh the first row.

Sub-processors

Conversation content goes to the model providers you configure, listed on the AI transparency page. A local-model configuration sends nothing out.

Approval notifications, once a paired device turns them on, go through that browser's push service (Google, Mozilla, Apple or Microsoft). Each one is encrypted to the device and says only that a request is waiting, never what it asks. The push service sees the device's address and when a notification was sent.

Contact

Security issues: security. Compliance and data protection: [email protected].