I built TrustPlane on top of cognos-proof-engine, an epistemic scoring
engine that scores every LLM request before it reaches users.
Formula: C = p × (1 − Ue − Ua)
Four outcomes per request: PASS, REFINE, ESCALATE (webhook), BLOCK.
Every decision gets a trace ID and ends up in an audit log.
Pluggable providers: Ollama, OpenAI, Anthropic, Groq, Cerebras.
Multi-tenant PostgreSQL isolation. Redis rate limiting. MCP server for
Claude Code integration.
The honest limitation: p (prior confidence) is currently a configurable
baseline, not dynamically calibrated per request. That's on the roadmap.
Self-hosted via docker-compose or SaaS. Air-gap compatible for regulated
deployments.