AI POLICY · AI ACT ART. 50 · L. 132/2025

Cite or refuse. Signed register. Local at query time.

Lemnia refuses to answer when the source is missing. Every query is logged in a BLAKE3-signed register, exportable for the auditor or the court. Mandatory citation is applied to all packs — both strict (reminder, quotation, expert appraisal) and hedged (customer dossier).

PRINCIPLE 01

Cite, or refuse.

Every sentence Lemnia produces is anchored to a source document. When the source is missing, Lemnia refuses to answer with the phrase: "Non ho evidenza sufficiente per questa affermazione." It does not invent. The user is informed that no answer is the appropriate answer.

The cite-or-refuse pipeline runs in 5 steps: (1) decomposition of the answer into atomic claims; (2) substring match against the evidence set; (3) mDeBERTa-NLI entailment verification; (4) consistency check against the knowledge graph; (5) strip-and-replace for non-entailed claims.

PRINCIPLE 02

Sign every query. Sign every answer.

Every interrogation enters an append-only processing register, sealed BLAKE3 per entry. Timestamp, tenant id, operator id, query text, retrieval path, source citations, output text, model hash — every field is captured and signed.

The register is exportable to PDF, JSON-LD, or CSV — all signed. It satisfies the GDPR Art. 30 record-of-processing obligation and the evidentiary standard implied by Tribunale di Siracusa 338/2026 (Art. 96 c.p.c. gross negligence).

PRINCIPLE 03

Local at query time. Cloud only when you opt in.

Query-time retrieval and generation run on the company's hardware. No fragment of customer data, supplier data, employee data leaves the LAN. The local Italian-trained models (Qwen3.5-4B Q4_K_XL, mDeBERTa NLI, Qwen3-Embedding-0.6B) handle every interrogation.

Cloud-burst (Pro mode) is opt-in. It runs only ingest and long-generation jobs. Each batch requires per-batch consent, captured in the signed register. The cloud worker uses EU-hosted infrastructure (RunPod EU, Hetzner SEV-SNP planned). No US data transfer.

PRINCIPLE 04

Deterministic-preferred. Agentic loops never on local models.

Default execution mode is deterministic state machines + LLM-as-bounded-tool. Predictable, fast, auditable, low token cost. Voice walk-up, standard search, the 13 reports, the 12 product primitives, the low-latency UX paths — all strict deterministic.

Agentic LLM tool-use loops are permitted only when genuinely a better fit than a deterministic decomposition (multi-step ambiguous research, novel query shapes, no template match). Hard constraint: agentic loops run only on cloud-side models (Qwen3.6-35B-A3B-FP8, Velvet-14B). Local Qwen3.5-4B agentic is banned — local precision is insufficient for tool-use loops, produces unreliable and uninsurable behaviour.

PRINCIPLE 05

Your data does not train our models.

Customer data is never used to train, fine-tune, or evaluate Lemnia models. The model pipeline is trained on public Italian corpora (Wikipedia IT, GovIT, ItaCorpora, Italian Constitutional Court rulings) and on synthetic data generated by Lemnia's own training pipeline.

Disambiguation decisions made inside a tenant's instance (HITL modal answers) feed continuous training data — but only for that tenant's instance. No cross-tenant data flow. No model improvement across customers.

REGULATORY ANCHOR

Garante Provvedimento 474/2025 + Trib. Siracusa 338/2026.

Two Italian decisions define the defensible posture for generative AI in operational contexts: Garante Provv. 474/2025 (Careggi case, €80,000 fine for use of generative AI without verified sources on patient records), and Tribunale di Siracusa 338/2026 (Art. 96 c.p.c. gross negligence for citation of AI-generated case-law precedents).

Lemnia is built around the explicit interpretation that, in Italy, the only insurable posture is mandatory citation + query-time local execution + signed register. This AI Policy makes that posture contractual.

FOUNDER PROGRAM · LIMITED SEATS

Lemnia working on a real company's data.

A 30-minute demo, tailored to your industry. Lemnia composes a real customer's dossier, cites the sources line by line, and shows the signed register ready for the DPO.

Request a pilot→Download the technical dossier