Open models, kept records, your infrastructure.

Source provenance

Every source is captured and stored with its retrieval time and location, and never pruned. Quotations are checked against the captured text, so the evidence behind a claim can be re-read exactly as the system read it.

Specialized models

Open-weight models run the research workflow. Smaller task-specific models for extraction and verification are trained and must pass a truth gate on human-labelled (passage, claim) pairs before they are used.

Local inference

Inference runs on GPUs operated by insum rather than through third-party model APIs by default. The same stack can be pointed at an organization's own inference servers.

Customer-controlled deployment

The research stack can run on customer-controlled infrastructure. Fully local operation requires sufficient GPU capacity and locally available data; live web discovery requires outbound network access.

Security architecture

Private-network access controls on hosted deployments and metadata-only engine logs that record model, token counts and cost, never prompts, page text or outputs. Single sign-on, audit export and customer-managed keys are planned.