Probe Overview

GOVERN Probe — Runtime Inference Monitor

The GOVERN Probe is a lightweight Docker container that sits between your application and your AI model API. It scores every inference against governance policies in real time.

Architecture

Your App → Probe (:9090) → AI Model API (OpenAI, Anthropic, etc.)
              ↓
         Score input
         Forward request
         Score output
         Buffer telemetry
              ↓
         Flush to GOVERN API every 60s

What It Scores (5 Dimensions)

Dimension	What Is Checked	Alert Threshold
Security	PII (SSN, email, CC, phone), prompt injection (5 patterns), credential leaks (API keys, tokens)	Any detection
Bias	Gender stereotypes, racial bias, ageism patterns	Score > 0.2
Accuracy	Hallucination signals, fabricated citations, temporal impossibilities, grounding indicators	Score > 0.3
Cost	Hourly token budget, USD spend rate	Exceeds configured budget
Drift	Entropy-based behavioral deviation from baseline (exponential moving average)	Drift > 0.3

Scoring Modes

Mode	Behavior
log	Score every request, never block. Telemetry flows silently.
flag	Score and flag violations in response headers.
block	Score and block requests with critical violations (returns 403).

Health Endpoints (port 9091)

Endpoint	Purpose
/healthz	Liveness — is the probe running?
/readyz	Readiness — are policies loaded? Is buffer healthy?
/metrics	Prometheus-compatible metrics (uptime, memory)