Effective Date: August 27, 2026
Owner: Head of CyberSecurity, Matías Labanca
Review Cycle: Annual
Confidential. Shared under NDA. Do not redistribute without Horizon's written consent.
This policy defines how Horizon governs the AI and model components that process customer data: which providers are used, how model versions are tracked, what guardrails are applied, how outputs are traced to their inputs, and how model changes are detected and rolled back.
It exists because AI-generated output in Horizon is evidence to be evaluated, not an authoritative statement of process truth. Every finding carries the evidence beneath it and is subject to human validation by a named owner.
Applies to all model providers in the processing path for customer data, and to the guardrail, logging and version-control mechanisms that surround them.
| Provider | Role | Version control |
|---|---|---|
| AWS Bedrock | Primary inference | Alias pinning, rollback within hours |
| Anthropic API | Secondary inference | Explicit version string in payload, reversible in a 15 minute deployment |
| Azure OpenAI | Embeddings | Explicit version string in payload |
| OpenAI API | RAG and knowledge graph embeddings | Explicit version string in payload |
No provider uses Horizon customer data to train models. This is a contractual commitment, not a policy statement, for every provider with an executed Data Processing Agreement.
AWS Bedrock guardrails are active with six policies: content filters set to HIGH, denied topics, keyword filters, PII masking, grounding checks, and automated reasoning checks. The active guardrail configuration for each period is versioned and included in the monthly report.
Reproducibility: Horizon can deliver, for any past output, the inputs, the model ID that processed them, and the stored output. Horizon cannot regenerate byte-identical output on demand, because large language models are non-deterministic and providers update versions. This limitation is disclosed rather than obscured.