Public architecture overview

A durable evidence system, not a stateless prompt wrapper.

WITNESS combines a creator-facing application with durable workspace state, governed evidence storage, asynchronous observation, and machine-facing MCP/A2A surfaces. This page intentionally describes the public architecture without exposing private infrastructure identifiers or credentials.

Creator application

Ask, Findings, Records, Watches, Connections, account settings, and billing form the creator-facing product surface.

Durable workspace agent

Per-workspace durable state preserves baselines, entities, prior Findings, open hypotheses, priorities, and Watch context.

Evidence/control plane

Structured data stores preserve governed state, idempotency, provenance, evaluation records, and operational commitments.

Durable records

Record artifacts can be stored separately from control-plane metadata so provenance and commitment boundaries remain explicit.

Queue-driven work

Asynchronous ingestion and investigation work can be separated from the interactive request path.

Machine interfaces

MCP and A2A expose WITNESS as an evidence layer for compatible AI systems without redefining truth authority.

Signal Integrity precedes reasoning.

Evidence quality is not delegated entirely to the model. WITNESS can evaluate provenance, contradictions, structural patterns, and limitations before downstream inference consumes a signal.

The creator remains the protected principal.

Private artifacts remain principal-scoped, while bounded public evidence can be observed separately. The system architecture is designed so expanding the creator network does not require turning creators into the product.

Durability changes the product.

Because workspace intelligence persists, WITNESS can compare current observations with prior context and reopen questions when evidence changes. That is fundamentally different from repeatedly prompting a stateless model with a fresh export.

Architectural objective: preserve evidence, reasoning state, provenance, and change over time—not merely the last generated answer.