> ## Documentation Index
> Fetch the complete documentation index at: https://docs.prisme.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Prisme.ai Data-Retention Posture
> Which model routes Prisme.ai treats as production-compliant, how Zero Data Retention is enforced per route, and the evaluation-only carve-out.
Prisme.ai's [LLM Gateway](/api-reference/llm-gateway) can route inference to several model providers and to self-hosted models. Because data-retention behaviour depends on the **route taken**, not on the platform alone, Prisme.ai classifies routes into two tiers:
* a **Production / compliant** tier, where Zero Data Retention (ZDR) is **contractual** or **default**;
* an **Evaluation** tier, reserved for testing the latest models, where **no real data** is allowed.
For the neutral, sourced comparison of each provider's policy, see [LLM Provider Data Retention & ZDR](/resources/security/llm-provider-data-retention). For how Prisme.ai handles your data overall, see [Data Privacy](/resources/security/data-privacy).
We deliberately display **"Contact us"** where a route requires a request or approval, instead of advertising blanket ZDR. We present ZDR **only** where it is contractual or default, and we scope evaluation routes explicitly. This is intentionally less "marketing" — and more defensible.
## Deployment & data residency
Prisme.ai SaaS is deployed as **multi-tenant instances, one per cloud provider**:
* **OVHcloud (France)**, and
* **AWS (Europe)**.
Your organization chooses **which cloud** it runs on — a **data-residency** choice. The two cloud instances are **separate**; within an instance, tenants are **logically isolated**. Your **platform data** — conversations, files, vector indexes — stays within the region of your chosen instance, which sets the **platform-side residency** independently of the LLM routes below. Dedicated / single-tenant infrastructure is available for enterprise needs (see [Data Privacy](/resources/security/data-privacy)).
## Customer-governed model selection
Prisme.ai provides the LLM Gateway and a **catalogue** of model routes. **Each customer whitelists, in their own organization, which models are allowed** (managed through [Model Governance](/products/ai-governance/model-governance)). So the *effective* set of LLM subprocessors is **chosen and controlled by the customer**, not imposed by the platform. The compliant-tier routes below (OVHcloud, AWS Bedrock) are what Prisme.ai **recommends and certifies**; what actually applies to a given organization is **its own whitelist**.
## Supported routes & retention posture
| Route | Retention posture | To enable | Status |
| ------------------------------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------- | ------------------------------- |
| **OVHcloud AI Endpoints** | **ZDR — contractual by default** (Special Conditions, Annex 10: *no backup or storage, even ephemeral; OVHcloud does not store or reuse inputs/outputs*). EU / sovereign, open-weight models | Accept the Special Conditions — nothing to negotiate | ✅ **Available** |
| **AWS Bedrock** | **ZDR by default** (AWS-sold models) + zero operator access. `data_retention_mode = none` enforced. **Excludes** `provider_data_share` models (Claude Fable 5 / Mythos 5 → 30-day retention) | AWS DPA applies via Service Terms + `none` mode (SCP-locked) | ✅ **Available** |
| **Google Vertex AI** | ZDR achievable by configuration (in-memory cache 24h, per-project residency) | **On request** — disable cache + opt out of abuse logging | 🟡 **Contact us** |
| **Microsoft Azure (OpenAI / Foundry)** | ZDR achievable via the **Modified Abuse Monitoring** program | **On request / approval** — Limited Access (managed customer, EA / MCA) | 🟡 **Contact us** |
| **Public APIs** — OpenAI, Anthropic, Mistral, Gemini Developer API | **Outside the default ZDR scope** | **Technology watch & POC only** — synthetic data, **never** customer / personal / production data | ⚪ **Contact us** for production |
## Production / compliant tier
**OVHcloud AI Endpoints** — Zero Data Retention is a **contractual** commitment (OVHcloud Public Cloud Special Conditions, Annex 10 — AI Endpoints): the service includes *"ni sauvegarde ni stockage, même éphémère"* and *"OVHcloud n'effectue aucune sauvegarde des Inputs ou des Outputs"*. Hosting is European / sovereign. Catalogue is open-weight models only.
**AWS Bedrock** — Bedrock applies a Zero Data Retention model by default for AWS-sold models, with zero operator access. Prisme.ai sets `data_retention_mode = none` and locks it organization-wide via an SCP (condition key `bedrock-mantle:DataRetentionMode`). The AWS GDPR DPA is incorporated into the AWS Service Terms by default.
**Excluded from the compliant tier:** models that require `provider_data_share` — currently Claude **Fable 5** and **Claude Mythos 5**, which mandate 30-day retention. Under `none` they are automatically rejected; if needed they belong to the evaluation tier only.
## Evaluation tier (latest models)
The direct public APIs (OpenAI, Anthropic, Mistral, Gemini Developer API) are kept available so teams can **evaluate the newest models** (technology watch, demos, proofs of concept).
**Hard rule.** The evaluation tier uses **synthetic data only** — **never** customer, personal or production data. It runs on **isolated workspaces and API keys**, with no production routing. ZDR is not assumed on these routes.
## Cross-cutting
Provider-side ZDR covers stateless inference. The platform still stores conversations, with configurable retention — that is a separate control.
Platform data stays in the cloud instance you chose (OVHcloud France or AWS Europe), with logical tenant isolation. LLM processing then depends on the route your organization whitelists — fixed explicitly per route.
Compliant routes are backed by the provider DPA and, where applicable, a signed ZDR arrangement, recorded in the subprocessors register.
Provider terms and "covered models" evolve. The posture and the excluded-models list are reviewed periodically.
This page describes Prisme.ai's routing posture. The underlying provider facts (with official sources) are on [LLM Provider Data Retention & ZDR](/resources/security/llm-provider-data-retention).