Article 50 implementation status
Rūpestėlis Holding UAB / Margelis.ai · status as of 18 July 2026 · next review 2 August 2026. We publish our own implementation status against the same 16-point instrument we give customers, in the form a reviewer expects: named owner, system inventory, per-obligation status, and dated remediation for anything not yet implemented.
1. Accountable owner
Tomas Margelis, founder and CEO, Rūpestėlis Holding UAB — accountable for Article 50 implementation end to end, including the inventory below and the evidence file. Contact: [email protected].
2. AI system inventory
| System | What it produces | Model / vendor | Article 50 relevance |
|---|---|---|---|
| Margelis Agent — Telegram | Conversational text with end users | DeepSeek (text) | 50(1) disclosure — implemented |
| Margelis Agent — web chat | Conversational text with end users | DeepSeek (text) | 50(1) disclosure — implemented |
| Agent image input | Reads user images, returns text | OpenAI vision model | Input processing; produces no synthetic media |
| Outreach engines (5) | Email to institutions and companies | Templates, human-approved, sent via Resend | Commercial correspondence — outside 50(4) |
| Research agent | Internal research notes | DeepSeek + Serper | Internal only, not published |
| Website content | Public explanatory pages | AI-assisted, human-reviewed | 50(4) disclosure — implemented |
| Provenance service | Owner-signed origin manifests | Ed25519, own implementation | 50(2) marking — partially implemented |
3. Status by obligation
| Obligation | Status | Basis / target date |
|---|---|---|
| Users told they are interacting with AI | Implemented | Disclosure in the first message of every session |
| Independent third-party verification of marks | Implemented | Public trust root and verification endpoint; Ed25519 key pinned in DNS |
| Position on the marking grace period | Implemented | Documented; grace to 2 Dec 2026 applies only to pre-2 Aug systems |
| Disclosure of AI-assisted public text | Implemented | Disclosure statement on public pages, under named editorial responsibility |
| Inventory with accountable owner | Implemented | Sections 1–2 above |
| Code of Practice assessed | Implemented | Section 4 below |
| Named person accountable end to end | Implemented | Section 1 above |
| Machine-readable marking of synthetic output | Partial | Available on request today; automatic marking at generation — target 2 Dec 2026 |
| Marking robustness and interoperability | Partial | Owner-signed manifest implemented; C2PA embedding not implemented — see section 5 |
| Evidence file coverage | Partial | Tamper-evident evidence exists; extension to all published AI-assisted output — target 2 Dec 2026 |
| Article 4 AI literacy: trained, attendance recorded | Planned | Register established 18 July 2026; first recorded session — target 2 Aug 2026 |
| Labelling convention / EU icon set adopted | Planned | Target 2 Aug 2026 |
| Provider/deployer split written into vendor documentation | Planned | Target 2 Aug 2026 |
| Deepfake labelling | Not applicable | We do not produce synthetic media depicting real people or events |
| Emotion recognition / biometric categorisation notice | Not applicable | We operate no such systems on people |
| Marking that survives screenshot and re-encoding | State-of-the-art limited | Not solved for text by any vendor — see section 5 |
4. Assessment of the Code of Practice on Transparency of AI-Generated Content
We assessed the Commission's Code of Practice (final since 10 June 2026). Our position:
- We align with it. We have not signed it — signing implies commitments across marking technologies we have not implemented (section 5). We would rather be accurate than decorated.
- Aligned: disclosure at first interaction, inventory with accountable owner, evidence trail, independently checkable signatures.
- Not aligned: we do not embed C2PA credentials and do not watermark. Our marking is a detached, owner-signed manifest.
- Reassessment: when the marking layer changes, and in any case before 2 December 2026.
5. Where the state of the art actually stands on marking text
C2PA and watermarking are engineered for images, audio and video. For plain text output there is currently no marking that is simultaneously machine-readable, robust to copy-paste and screenshotting, and interoperable across readers. A vendor claiming otherwise is describing metadata that a screenshot destroys.
This is why our layer answers a different question. Instead of “is there a mark?”, we make “can you prove this output is ours and unaltered?” checkable by a third party: an owner-signed manifest anyone can recompute against a key pinned in DNS, without trusting us. How the verification layer works →
We state this limitation because a buyer will hit it in month two regardless. Knowing where the technology genuinely stops is part of what we are selling.
6. Use the same instrument
The 16-point checklist we assessed ourselves against is free, interactive, and yours to run: Article 50 readiness checklist · lietuviškai
Request a free readiness review