Architecture
System boundaries and tradeoffs: a policy engine and an evidence engine sit between the model orchestrator and any response that reaches a client. Architecture diagram
Proof, by category
Tests
Deterministic verification and explicit failure paths, including adversarial cases (prompt injection, claim escalation, PII hallucination) -- not just happy-path coverage.
Evidence
Every public claim traces back to an approved entry in a generated, checksummed evidence manifest. Unsupported claims are rejected before they reach a client.
Operations
Structured telemetry per request (latency, coverage, adjudication result) without ever logging the raw question or provider output. Degraded modes fail closed to one safe answer.
Governance
Automation preserves owner approval by default: write-capable actions require explicit dry-run/apply separation, and the MCP tool surface is a single read-only weather tool, not a general-purpose fetch/filesystem/write capability.
Go deeper
- Architecture diagram
- Evidence-backed answers
- Use-case labs
- GitLab repositories
- Capability index — what this site demonstrates
This page does not claim production customer traction, revenue, or employer-confidential detail beyond what is documented above and in the linked evidence.
Questions worth asking the chat
- How does the evidence system prevent unsupported claims?
- What did Nkosi build at NBCU?
- How does the Kafka pipeline proof map to production systems?
- What are the failure modes?