AI Transparency Register
Last updated: 16 September 2026. EU AI Act compliant (Art. 13 transparency, Art. 50 disclosure)
ShoppingPartnerLab maintains this public register of all AI systems used to operate the platform. None of our systems are classified as high-risk or prohibited under the EU AI Act.
How we select AI models: the TITAN router
Our AI systems run on TITAN, a proprietary routing layer that selects the most suitable model from a managed pool per request. There is therefore no fixed model per agent, the model used varies per task, prompt and context.
The pool intentionally includes preview models (incl. gemini-3-flash-preview, gemini-3-pro-preview, gemini-3.1-pro-preview) that TITAN deploys per task. They are not concealed here.
Auditability and logging: AI decisions (model, agent, task type, timestamp, summary of input and output) are recorded in ai_act_audit_log. That series starts on 11 March 2026, so roughly six months of history. There is no automatic deletion, but three-year coverage will only be factually demonstrable in March 2029. Readout runs through a daily internal job (pg_cron, 05:15 UTC), not through an external plugin or compliance tool.
Open items: per-call model selection logging in ai_model_quality_logs ran from 17 January to 22 March 2026 and stopped after that date; restoration is in development, expected 31 October 2026. The SHA-256 seal applies only to forensic dossiers and security reports, not to the AI audit logs; extension is in development, expected 31 December 2026. The router decision trace (why a specific model was chosen) is not yet persisted.
Provider pool
TITAN selects from four upstream providers. The same four providers appear on our sub-processor list.
| Provider | Route | Example models in the pool | Usage |
|---|---|---|---|
| Lovable AI Gateway (zero-retention) | gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite, gemini-3-flash-preview, gemini-3-pro-preview, gemini-3.1-pro-preview | Generative text, reasoning, routing default | |
| OpenAI | Direct API (api.openai.com) | gpt-5-nano (chat), text-embedding-3-small (768 dims, RAG embeddings) | Chat fallback and all production embeddings (Julia RAG, Nora, blog ingest, semantic search) |
| Anthropic | Lovable AI Gateway and direct API (api.anthropic.com) | claude-haiku-4.5, claude-sonnet-4.5, claude-opus-4-5-20251101 (date-pinned, direct API) | Content, sales automation, compliance review |
| Perplexity | Direct API (not via the gateway) | sonar-pro | Independent second opinion in model consensus |
Risk classification
- Limited risk (Art. 50 transparency obligations apply): systems that generate content, interact with customers, or produce legal/compliance output. AI-generated output is labelled as such on every public page where it appears.
- Minimal risk: backend tooling, deterministic or rule-based agents, internal QA. No specific obligations beyond general AI literacy and audit logging.
Human oversight
AI model assessments are validated by the Compliance Officer before any score appears in the public Trust Registry. As of 16 September 2026 this is technically enforced: a database trigger places every new score of 80 or above in trust_registry_score_gate with status "pending", and the score only counts as qualified after explicit approval (confirmation by two independent models or by the Compliance Officer). Scores published before 16 September 2026 remain visible and are flagged as "legacy pre-gate". Human intervention is recorded through human_override and override_reason in ai_act_audit_log.
Risk classification: the eight registered systems and their risk assessments were re-evaluated on 16 September 2026 (previously 15 March 2026). Reassessment now runs automatically every month (1st of the month, 04:30 UTC) and each run is recorded in ai_act_audit_log.
AI-generated content disclosure
Every page that contains AI-generated content (scan reports, fix reports, AI assessments, blog posts) carries a visible "AI-generated" label, in compliance with EU AI Act Art. 50.
Supervisory authority
Autoriteit Persoonsgegevens (Dutch DPA), autoriteitpersoonsgegevens.nl
AI Compliance Officer: info@shoppingpartnerlab.com
Langhoor Marchal Consultancy. KvK 88375692. VAT NL004594074B45. Maastricht, Netherlands, info@shoppingpartnerlab.com
