kvrun-relay-v1 / EU_AI_ACT.md
JoeKVRun's picture
Update EU_AI_ACT.md
3097e72 verified
|
Raw History Blame Contribute Delete
8.76 kB

EU AI Act assessment: kvrun-relay-v1 and the KV Relay app

Regulation (EU) 2024/1689 (AI Act), as amended by the Digital Omnibus on AI, Regulation (EU) 2026/1744 [verify in the Official Journal]
Subject (a) the model kvrun-relay-v1; (b) the KV Relay desktop app, which uses it
Provider KVRun Ltd, United kingdom. Contact: support@kvrun.com
Prepared 2026-10-27, version 1.0
Status Self-assessment by the provider. Not legal advice. Have a qualified adviser review it before relying on it.

1. Summary

Question Answer
Is kvrun-relay-v1 a general-purpose AI (GPAI) model? No. It is a narrow classifier with ~10¹⁶ FLOPs of training compute that cannot generate content.
Is it, or the app, a prohibited practice (Art. 5)? No.
Is it, or the app, a high-risk AI system (Art. 6, Annex I/III)? No, for the intended purpose.
Does the open-source exemption (Art. 2(12)) apply? Yes, both are released under Apache-2.0. This removes most obligations, but not Art. 50.
Which obligations apply? Art. 50 transparency for the app (in application since 2 Aug 2026). AI literacy (Art. 4) also applies; see §6.
Registration in the EU database? Not required (no Annex III use).

2. Roles

  • Provider of the model and the app: KVRun. It develops both and releases them under its own name (Art. 3(3)).
  • Deployers: people and organisations that run the app or embed the model (Art. 3(4)).
  • Upstream: Laya (convaiinnovations) and ModernBERT (Answer.AI) are the base models; the provider modified them by fine-tuning. Neither is a GPAI model under the Act's indicative criterion (they are encoder classifiers, not generative models), so the GPAI rules on modifying a model (Recital 109, Commission GPAI guidelines) are not triggered.
  • Third-party models the app routes to (Ollama models, OpenAI-compatible APIs, Anthropic, agent CLIs): these are provided by others. The app passes the user's prompt to them and shows their output.

3. Classification

3.1 Not a general-purpose AI model (Art. 3(63), Art. 51)

A GPAI model "displays significant generality and is capable of competently performing a wide range of distinct tasks". The Commission's guidelines use an indicative criterion: training compute above 10²³ FLOPs and the ability to generate language, images or video.

kvrun-relay-v1:

  • does one task: a fixed 5-way category plus a 0–4 complexity score for a prompt;
  • cannot generate text;
  • used ~1×10¹⁶ FLOPs for fine-tuning, about seven orders of magnitude below the criterion (README.md → Training).

Conclusion: it is not a GPAI model, so Art. 53 and Art. 55 obligations don't apply. As voluntary good practice, the model card and TRAINING_DATA.md follow the spirit of Art. 53(1)(a)–(d) anyway: documentation, a training-content summary and a copyright note.

3.2 No prohibited practice (Art. 5)

The model and the app don't manipulate or exploit people, score people socially, assess criminal risk, scrape facial images, recognise emotions, categorise people biometrically or identify them remotely. The classifier labels text tasks, not people.

3.3 Not high-risk (Art. 6)

  • Art. 6(1), Annex I: not a safety component of a regulated product, and not such a product.
  • Art. 6(2), Annex III: routing chat prompts by topic and difficulty is not one of the listed uses (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice and democratic processes). It involves no profiling of people.

Condition: this holds only for the intended purpose in the model card. The model card lists Annex I/III contexts as out of scope. Someone who repurposes the model for such a context (e.g. triaging job applications or benefit claims) changes its intended purpose and becomes a provider of a high-risk system (Art. 25(1)(c)).

3.4 Open-source exemption (Art. 2(12))

"This Regulation does not apply to AI systems released under free and open-source licences, unless they are placed on the market or put into service as high-risk AI systems or as an AI system that falls under Article 5 or 50."

Both the model and the app are released under Apache-2.0 and are not high-risk or prohibited. The app, however, falls under Art. 50 (it interacts with people and relays generated text), so Art. 50 still applies to it.

4. Article 50 transparency: what applies and what is done

Timing: Art. 50 has applied since 2 August 2026. The grace period for machine-readable marking (until 2 December 2026) covers only systems placed on the market before 2 August 2026, so it doesn't cover this app's first release.

Obligation Applies to Status
50(1) People must be told they're interacting with an AI system, at the latest at first interaction, clearly and accessibly (50(5)), unless that's obvious App Implemented. A permanent notice under the message box says "Agents are AI. Replies are generated by AI models and can be wrong." Every reply names the model that produced it. The notice is plain text, readable by screen readers.
50(2) Synthetic text must be marked in a machine-readable format and detectable as AI-generated, where technically feasible App, as the system delivering generated text to the user Partly implemented. Each stored reply carries machine-readable provenance (role: assistant, the generating model, the routing decision) in the local database and the API. Open: adopt a marking method once the Commission's code of practice on marking AI-generated content is final (e.g. metadata on copy and export). Text watermarking isn't feasible here because the app doesn't control the generating models.
50(3) Emotion recognition or biometric categorisation disclosure — Not applicable.
50(4) Deep fake or public-interest text disclosure Deployers Not applicable to the provider. Deployers publishing generated text on matters of public interest have their own duty.
— The classifier itself Model Not applicable. It doesn't interact with people or generate content.

5. Other horizontal law (not AI Act, listed for completeness)

  • GDPR: classification runs on-device and no prompt is sent anywhere to be classified. The app's local database (chats, route_log) may contain personal data typed by the user. It stays on the user's machine, and the user can delete chats. The training data is synthetic, with no real user data. Generated prompts may contain fictional names, and no automated PII scrub was run. Open: add an "erase history and route log" control and document local retention.
  • Copyright and licences: the base models and generator models are Apache-2.0. Gemma 4's licence page (ai.google.dev/gemma/apache_2) is plain Apache-2.0 with no distillation clause, so the Gemma Terms of Use "Model Derivatives" restrictions don't apply. Attribution is in NOTICE. Open: confirm the licence of the Qwen 3.x models used for 75 training rows, or drop those rows and retrain.

6. AI literacy (Art. 4)

Art. 4 has applied since 2 February 2025. The Digital Omnibus changed how AI literacy is allocated. In any case, the model card explains what the model does, how well, and its limits, and the app explains each routing decision in plain language. Organisations deploying the app should make sure their staff understand these limits.

7. Keeping this current

Re-run this assessment when:

  • the intended purpose changes;
  • the model is retrained on new data sources;
  • the app starts acting on the user's behalf. Agent tools such as running commands or editing files are planned; they don't change the risk class on their own, but review Art. 50 and add safety documentation for tool use;
  • the Commission publishes relevant guidelines or codes of practice, notably on Art. 50 marking;
  • the AI Act is amended.

Version history: 1.0 (2026-10-27) first assessment.

Sources

  • Regulation (EU) 2024/1689, Articles 2(12), 3, 4, 5, 6, 25, 50, 51, 53, Annexes I and III (https://artificialintelligenceact.eu).
  • Digital Omnibus on AI, Regulation (EU) 2026/1744: adopted 29 June 2026, published 24 July 2026; Art. 50 date unchanged; Annex III high-risk deferred to 2 December 2027. Taken from secondary sources (Gibson Dunn, Orrick, Morgan Lewis, Jones Walker); verify in the Official Journal.
  • European Commission, Guidelines on the scope of obligations for providers of general-purpose AI models (July 2025): the 10²³ FLOP indicative criterion.
  • Google, Gemma 4 licence: https://ai.google.dev/gemma/apache_2