Repository navigation
European edge AI requirements
Assessment date: 15 November 2026. Scope: public primary sources and the current NodeLift architecture and energy contracts. This is an engineering requirements assessment, not a conformity certificate. No new physical-device, cryptographic or cross-region benchmark was performed for this report.
The useful finding is a method, rather than a compliance shortcut: define the workload, actor, threat model and measurable acceptance conditions before selecting distribution or privacy technology. A project objective, a funding award and an implemented mechanism are different kinds of evidence.
| Project | Verified identity and funding | What the inspected sources establish | Evidence boundary |
|---|---|---|---|
| EdgeAI-Trust | Decentralized Edge Intelligence: Advancing Trust, Safety, and Sustainability in Europe. Chips JU / Horizon Europe grant 101139892, 1 May 2024 to 30 April 2027. | Decentralized edge architecture with safety, trust and sustainability objectives. The reporting summary describes requirements and architectural work. | A project summary does not supply individual acceptance clauses. Regulatory deliverable bodies were not obtained. CORDIS, official project, reporting. |
| dAIEDGE | A network of excellence for distributed, trustworthy, efficient and scalable AI at the Edge. Horizon Europe grant 101120726, coordinated by DFKI; 1 September 2023 to 30 November 2026 in the current fact sheet. | Public requirements report D3.2 was actually retrieved and inspected. It supplies a requirements/KPI/verification methodology and application examples. | This is a research framework, not legislation or a certification scheme. Its 2024 document records the original 36-month schedule; the current CORDIS end date takes precedence for this identity table. CORDIS, D3.2. |
| EDGELM | Efficient and Privacy-Preserving Edge Deployment of Multimodal Large Models. MSCA grant 101274374, Politecnico di Torino; 1 October 2026 to 30 September 2028. | Objectives include compression, heterogeneous orchestration, model cooperation, energy/latency trade-offs and federated adaptation. | The project had recently started at this assessment date. No versioned public requirements deliverable or achieved encrypted-inference result was inspected. CORDIS, host institution. |
| CIREN | Compliant Intelligent Reasoning in Nodes. The official page identifies IFAB call for members funding and E4 Computer Engineering, IFAB, SECO and University of Turin partners. | The described solution keeps raw training data local, uses federated learning, encrypts local weights for homomorphic cloud aggregation, then performs controlled final decryption. | No CORDIS grant was verified. The page sets a TRL 4 sandbox target and presents benefits as expected. It is not a published proof of arbitrary LLM inference over encrypted weights. IFAB project. |
These projects deserve acknowledgement as prior research, not endorsement claims. NodeLift has no verified partnership, funding relationship or conformance approval from them.
dAIEDGE D3.2, SINTEF, Report on Edge AI functional and non-functional requirements, V1.0, submitted 30 September 2024. The Zenodo publication date is 5 May 2025, distinct from the report date. Version/submission are on printed pages 1-2 (PDF pages 2-3). The file contains 59 PDF pages; its SHA-256 and retrieval provenance are in the source register. Record and download.
The following are paraphrased research examples, not new binding obligations. The report does not give these examples globally unique requirement IDs. NL-EU-* identifiers below belong to NodeLift.
| Inspected locator, D3.2 V1.0 | Source concept | NodeLift application proposed here |
|---|---|---|
| Section 3.1.2, printed pp. 34-35 / PDF pp. 35-36 | Trace needs through specification and verification. | Link each capability claim to its test, configuration and acceptance decision. |
| Section 4.2, printed pp. 35-37 / PDF pp. 36-38 | Associate requirements with measurable KPIs. | Record metric definition, measurement scope, target and uncertainty before a run. |
| Section 4.3, printed pp. 37-38 / PDF pp. 38-39 | Interoperability and downstream inference response. | Qualify each runtime/model/transport combination separately. |
| Section 4.3, printed p. 38 / PDF p. 39 | Self-descriptive interaction; user evaluation. | Test whether users distinguish planned regions, actual resources and missing measurements. |
| Section 4.3, printed p. 38 / PDF p. 39 | Fault injection and expected-output comparison. | Test late, duplicate, disconnected and invalid worker responses. |
| Section 4.3, printed p. 39 / PDF p. 40 | Confidentiality, maintenance downtime, adaptability. | Use threat-specific access tests and upgrade/rollback evidence. |
| Sections 5.1-5.2, printed pp. 40-42 / PDF pp. 41-43 | Verifiable functional targets and end-to-end latency. | Pre-register task-specific correctness and latency gates. |
Its numerical examples are domain examples, not NodeLift thresholds. In particular, an example attack-success percentage cannot establish security from a small test suite, and an example multi-second latency range is not a universal LLM service objective. Section 6 is a research discussion, not the authoritative legal text.
Other retrieval outcomes:
| Source | Public observation | What remains unavailable or unverified |
|---|---|---|
| EdgeAI-Trust official deliverables page | D7.3, initial standardization/regulatory report, M18, shown as In progress; D7.7, final regulatory alignment/standards report, M36, shown as Pending. | No inspected report body, version, date, page or formal clause. These are observed page labels, which may lag project work. Listing. |
| dAIEDGE CORDIS results | Requirements, distributed-network design and data-management reports are listed. | The EC attachment-preparation responses did not yield the network-design or final DMP bodies in this session. D3.2 was independently resolved through the project's public Zenodo link. Do not assign unverified deliverable numbers to the other titles. Results. |
| EDGELM | Public fact sheet and host project entry. | No inspected dated/versioned technical deliverable. Absence from this assessment is not proof that no report exists. |
| CIREN | Sections Obiettivi, Soluzione and Benefici on the official page. | No inspected public technical report with a document ID, version/date, pages, cryptographic parameters, reproducible code or benchmark protocol. The page's conformity and traffic-reduction claims were not independently established. |
This table is an engineering distinction. Implementing one row does not establish the properties of another.
| Mechanism | Protected boundary | Evidence required before making the claim | NodeLift status in this assessment |
|---|---|---|---|
| Local processing | Some input processing occurs on the originating device. | Actual payload/data-flow inventory, logs, caches, crash dumps, telemetry and onward destinations. | Distribution alone does not prove that all sensitive information remains local. |
| Federated learning | Training uses local datasets and exchanges updates. | Protocol, update contents, leakage/adversary tests, participation and aggregation policy. | No FL qualification performed here. FL is separate from model-partitioned inference. |
| Encryption in transit / at rest | Network links / stored objects. | Protocol and key handling, authenticated endpoints, disk/cache coverage, rotation/revocation tests. | Do not promote authenticated enrollment into proof of every workload's transport or storage confidentiality. |
| HE aggregation | Supported operations run on ciphertext updates. | Scheme, parameters/security level, operations, key/decryption authority, leakage assumptions and measured overhead. | CIREN describes this training-aggregation use. No NodeLift implementation or benchmark established here. |
| SMPC | Parties jointly compute under defined trust/collusion assumptions. | Protocol, parties, corruption threshold, active/passive adversary model, dropout and reconstruction analysis. | Worker splitting is not, by itself, secure multiparty computation. |
| Confidential execution | Plaintext is processed within a protected execution boundary. | Supported hardware, fresh remote attestation, measured software identity, key release and side-channel assumptions. | Worker authentication is not TEE/hardware attestation. |
| Encrypted-weight inference | The specified inference computation preserves its stated weight/input confidentiality. | Exact protected assets, permitted disclosure, supported operators, decryption locations, correctness and end-to-end overhead. | No such NodeLift proof. Encrypting a model file and decrypting it into ordinary worker RAM is a different claim. |
| Differential privacy | A quantified privacy guarantee for a defined mechanism. | Privacy unit, epsilon/delta, clipping/noise, accountant and composition over releases. | HE or FL does not automatically provide this guarantee. No DP evidence inspected. |
CIREN's described final decryption prevents interpreting its page as evidence that model weights stay encrypted throughout every edge inference. Its categorical privacy and conformity language needs a threat model and legal assessment; it cannot be inherited by NodeLift.
GDPR, where personal-data processing falls within its scope: establish actor roles, purposes and a lawful basis; apply minimization, retention, transparency and rights handling; use risk-appropriate safeguards, design/default protection and processor arrangements where applicable. Assess DPIA necessity and actual international transfers separately. Encryption is a safeguard, not a lawful basis or an exemption. Relevant locators: Articles 4-6, 9 where applicable, 12-22, 25-26, 28, 30, 32, 35-36 and 44-49. Official GDPR.
AI Act applicability depends on actor role, intended use, system/model classification and exceptions. A scheduler does not automatically become a high-risk AI system; hosting an imported model does not by itself settle GPAI-provider status. Check Articles 2-3, 6, 9-16, 25-27, 50 and 53-55 as applicable. Open source has specific exceptions and carve-outs, not blanket immunity. Annex XI's model-energy documentation is role-specific; it is not a universal obligation for every inference UI to deploy power sensors. Official Act entry.
Current-law verification limit: direct EUR-Lex full-text/PDF requests in this assessment returned verification pages, and an attempted metadata response was empty. Current amendments, applicability dates and transition rules were therefore not verified. This report does not determine a legal deployment deadline, and no acceptance gate depends on an unverified calendar claim. Retrieval observations remain in the retrieval log. Verify the applicable authoritative text before making a deployment decision. Official Act entry.
EDPB guidance, distinct from legislation: Recommendations 01/2020, version 2.0, adopted 18 June 2021, Annex 2, Use Case 5, paragraph 92, printed pp. 33-34 (PDF pp. 33-34), describes conditional split/multiparty safeguards. Use Case 6, paragraphs 94-95, printed pp. 34-35 (PDF pp. 34-35), explains why transport/storage encryption is insufficient in the specified problematic-third-country scenario when the processor needs cleartext. This is not a prohibition on every cloud service. Transfer mapping includes relevant remote access and onward processing, not just server labels. Final EDPB PDF.
A regional map is neither a transfer assessment nor evidence of residency enforcement. Controller/deployer decisions need the actual processing destinations, support access, contracts, subprocessors, purposes and safeguards. A DPIA and an AI Act fundamental-rights assessment have different triggers and must not be merged into a universal checkbox.
This baseline records source commit f2d850e. Subsequent automatic energy has its own coverage and validation.
This is a reading of current implementation contracts, not a production audit. See architecture, Hive console, compute metrics and job energy evidence.
- One logical management hub is observed in the pilot. Americas, Europe, Asia Pacific and Middle East / Africa are four planned zones, not four commissioned hubs or deployed capacity. Regional federation is not deployed.
- Locations carry explicit declaration/provenance. Unlocated resources remain unknown; IP address, locale or time zone do not establish physical location, residency or a permitted processing jurisdiction.
- Authenticated worker/resource identity supports attribution within the service. It does not establish globally unique physical hardware, trustworthy sensors or attested software execution.
- The optional
nodelift.job-energy.v1contract accepts bounded, instrumented per-job samples with source, scope and uncertainty. Authentication binds the report to the claimant/job; it does not certify the meter. Automatic sensor collection is not deployed by this change. - Historical job averages are not instantaneous cluster power. Missing/failed-attempt energy and unmetered host/network overhead cannot be silently assigned zero. No whole-cluster savings, encrypted inference or production federation is established by the UI or schema.
The following gates are NodeLift engineering proposals, informed by the sources. Their IDs and suggested evidence are ours. They do not reproduce purported EU project requirements. A legal citation identifies a potential applicability question, not a finding that NodeLift is already compliant or in breach.
| ID / priority | Gate and evidence needed | Present evidence / gap | Source basis |
|---|---|---|---|
| NL-EU-01 / P0 | Maintain a deployment record: intended use, actors, data classes, model licence, region/remote-access paths, applicable provisions and decision owner. | Architecture available; no per-customer legal-role or use-case assessment established here. | GDPR; AI Act role/scope analysis. |
| NL-EU-02 / P0 | Keep planned/observed topology distinct; prevent planned zones from increasing capacity or residency assurance. Test unknown locations and nested membership accounting. | Contracts explicitly distinguish one actual hub and four plans. No four-region operational proof. | D3.2 methodology; own operational gate. |
| NL-EU-03 / P0 | Map every workload artifact and recipient, including tokens, activations, updates, logs, caches and support access. Demonstrate allow/deny placement and fail-closed handling when location evidence expires. | Declared geo exists; residency enforcement across a federation is not established. | GDPR data-flow assessment; EDPB transfer guidance. |
| NL-EU-04 / P0 | Specify trust boundaries, enrollment/revocation, key lifecycle and payload/storage protections; test rogue and revoked workers. Add attestation only with its measured-software and freshness evidence. | Worker authentication exists; no hardware-attestation claim warranted. | D3.2 confidentiality example; risk-based security. |
| NL-EU-05 / P0 | Define result acceptance by workload; fault-inject duplicate, stale, corrupt and disconnected responses. Preserve attempt/fence lineage and correct accepted-work accounting. | Authenticated/fenced receipts and validation contracts provide a basis. They are not a proof of all LLM verifier semantics. | D3.2 reliability example; own correctness gate. |
| NL-EU-06 / P0 | Compare identical task/model/runtime/output constraints with and without delegation; report TTFT, useful throughput, total time, communication/waits, rejected work, fallback and uncertainty. | Historical experiments require configuration-specific interpretation; this assessment performs no new benchmark. | D3.2 latency/interoperability method; own experimental design. |
| NL-EU-07 / P1 | For energy claims, meter comparable full windows and devices; retain calibration, sample coverage, clock, scope and accepted-output denominator. Separate estimates from measured consumption. | Optional job ingestion exists; automatic collectors and whole-system coverage absent. | NodeLift energy contract; own energy gate. |
| NL-EU-08 / P1 | Bind audit records to model/runtime/artifact versions, topology, data policy and decision; redact secrets and minimize content retention. Exercise access/export/deletion and retention controls where applicable. | Diagnostics and contracts are a starting point; complete lifecycle/rights verification not performed. | D3.2 traceability method; GDPR applicability. |
| NL-EU-09 / P1 | Set workload-specific budgets for memory, network, latency, thermal state and useful energy; test admission, throttling and recovery under sustained load. | Inventory is not a validated allocation or speed-up. Multi-platform physical qualification remains separate. | EDGELM objectives; own measurable requirements. |
| NL-EU-10 / P1 | Test onboarding, pairing/revocation, unknown metrics and explanations with users, keyboard access and all supported locales. Record task completion/error rate and user feedback. | UI and localization review cannot substitute for physical companion tests. | D3.2 self-descriptiveness example. |
| NL-EU-11 / P1 | Before FL/HE trials, publish a bounded training experiment: adversary/key owners, update leakage, supported aggregation, decryption, convergence and crypto/network overhead. | CIREN is a useful separate experiment; no production FL/HE capability established here. | CIREN design, not its conformity claim. |
| NL-EU-12 / P1 | Before any encrypted-inference claim, specify whether inputs, weights, outputs or access patterns are protected; produce correctness, leakage and end-to-end resource evidence. | No encrypted-weight inference or SMPC/TEE qualification established. | Own cryptographic evidence gate; EDPB conditional guidance. |
| NL-EU-13 / P1 | Qualify upgrades and rollback on named platform/runtime combinations, with signed artifact provenance, compatibility and downtime evidence. | Package publication and runtime qualification are separate gates. | D3.2 maintainability/adaptability examples. |
| NL-EU-14 / P1 | Before federation claims, exercise independent regional authorities, policy enforcement, handover, partitions, resource fencing and failover with actual hubs. | Four zones are planned; current single-hub deployment cannot supply this evidence. | EdgeAI-Trust distributed-architecture objective; own federation gate. |
The immediate priorities are NL-EU-01 to NL-EU-06. Energy collection and a separate FL/HE demonstrator should then be evaluated against their own acceptance contracts. Neither research novelty nor encryption should bypass result correctness, measured benefit or deployment-specific legal responsibilities.
The machine-readable sources, checklist and retrieval log preserve what was available. Public PDF bodies are linked rather than copied into this repository. D3.2 text was extracted locally and its printed/PDF page offset was visually checked; the final EDPB PDF was inspected. Legal HTML/PDF fetches that returned verification pages are recorded as such; official indexed excerpts were not mislabelled as full-document downloads.
No public project source demonstrated a general, deployable solution that turns arbitrary heterogeneous consumer devices into simultaneously faster, confidential, legally compliant distributed LLM memory. That is a limit of the inspected evidence, not a claim that no such research exists. No external language-model response or consensus is used as proof of project facts, compliance or physical performance in this assessment.
NodeLift · © 2026 Fabrizio Degni · Mission and value · Source documentation · Engineering status remains scoped to its evidence.