Compliance & Transparency

This page is the canonical public compliance statement for Synavistra's AI document pipeline under Regulation (EU) 2024/1689 (EU AI Act). Every .snv.zip export from the pipeline cites this URL as its authority document. Each export is content-addressed and reproducible — the same source through the same recipe produces byte-identical outputs, which is the property that makes this audit trail verifiable.

Risk classification: limited — Art. 50 transparency obligations apply. High-risk classification (Annex III) is not triggered (assessment below).

Art. 50 transparency obligations: have applied since 2 August 2026; the implementation status of each obligation is in the summary below.

Compliance contact: legal@synavistra.ai

JSON Schema for the machine-readable compliance block: synavistra.ai/schemas/compliance/v1.json

Related documents: Privacy Policy (GDPR data-subject rights, Art. 13/14 information notices) · Impressum (Austrian §5 ECG company-registration details).

System Card

Intended purpose	Extract named entities and relationships from uploaded EU regulatory PDFs; generate knowledge graphs and Q&A answers for researcher and student use.
Prohibited uses	Biometric identification; emotion recognition; scoring of natural persons; law enforcement; recruitment/selection decisions; creditworthiness determinations; autonomous legal interpretations.
Known limitations	Evaluation results (precision, recall, F1) are measured per release and recorded internally; they are not yet included in the .snv.zip export. Not reviewed by domain experts. Not suitable for decisions with legal effect without human review.
Data governance	Zero-egress — documents processed entirely in the user's browser. No PII transmitted to Synavistra servers. Training data: public-domain EU regulatory text + Gemma 4 + MiniLM base models.
Inference provider	Public demo: in-browser only, no provider involved. Inference-online engagements: the Mistral AI API (Paris), an EU-based provider. Inference-offline engagements: model training on Google Cloud europe-west4 (Netherlands); the trained model then runs on the customer's own infrastructure.
Human oversight	All outputs are advisory. No automated decisions. Users retain full editorial control. Caution notices on all AI outputs.
Annex III assessment	No Annex III category triggered (see COMPLIANCE.md in any .snv.zip export for full article-by-article assessment).

Article 50 Compliance Summary

Full 113-article matrix is included in every .snv.zip export as COMPLIANCE.md. Key Art. 50 status:

Obligation	Status
Art. 50(1) — Direct interaction disclosure	IMPLEMENTED — persistent notice card above the fold
Art. 50(2) — Machine-readable marking	IMPLEMENTED — manifest.compliance, HTML meta tags, COMPLIANCE.md with SHA-256
Art. 50(3) — Emotion recognition disclosure	NOT APPLICABLE — STRUCTURAL (no biometric capability)
Art. 50(4)(a)-(d) — Deep fake labelling	NOT APPLICABLE — STRUCTURAL (no AV generation)
Art. 50(4) last para — Public-interest text	IMPLEMENTED — Art. 50(4) cited in all notices
Art. 50(5) — At first interaction	IMPLEMENTED — static banner before any AI output
Art. 50(7) — Codes of practice	MONITORING — AI Office codes tracked

Operational Transparency

Beyond the regulatory statement above, this section discloses how the AI Document Pipeline processes data in the public demo. The full machine-readable detail (model, input hash, source files, EU AI Act compliance marking, signature) lives in each .snv.zip export under manifest.json and COMPLIANCE.md; the summary below covers what a buyer or auditor most often wants on a first read.

Local Processing Architecture

In the public demo at /showcases/ai-document-analysis/, all processing happens locally on your device:

• The AI model is downloaded once to your browser and cached locally
• Your PDF documents are processed entirely in browser memory
• Extracted text, entities, and knowledge graphs never leave your device
• No analytics, tracking, cookies, or telemetry of any kind
• Exported .snv.zip files are saved directly to your local filesystem

Empirically verifiable: after the AI model downloads, disconnect your internet — the demo keeps working. The privacy claim is something you prove yourself, not a promise we make.

Classification-aware inference

In the demo, marking a document as SECRET redacts the source chunks at the database view level — they never reach the answer-synthesis model. The model uses only the knowledge graph to answer, ensuring derivative knowledge can be shared without exposing the underlying source text.

Open evaluation

Each release is measured on fixed evaluation sets (precision, recall, F1, parse rate), and the results are recorded internally. They are not yet included in the .snv.zip export. We invite independent auditors, researchers, and regulators to verify any claim — contact legal@synavistra.ai to coordinate.

Independent verification

The pipeline code behind the public demo, which this site serves to your browser, is licensed under Apache 2.0. The base model, Google's Gemma 4, is released by Google under Apache 2.0. Source documents we derive showcases from carry their own licenses (listed in each bundle's manifest.compliance block).