Designed to evaluate configured model routes before execution. Deterministic scoring and enforcement (ALLOW, CONSTRAIN, HOLD, BLOCK) — not an autonomous agent. Fail-closed when required evidence is absent or NOT_RUN.
Judge 6 is designed to evaluate configured model routes before execution. Canonical verdicts are ALLOW, CONSTRAIN, HOLD, and BLOCK. YC-1 demonstrates this behavior using synthetic fixtures.
Fail-closed HOLD when required evidence is missing, stale, or NOT_RUN. Judge 6 is a scoring and enforcement layer — not an autonomous agent. Fabricated absolute-proof claims are out of scope.
Every policy decision is logged with immutable timestamps. Full compliance trail for SOC 2, GDPR, and bar association review. No gap in the chain of custody.
Each of the 30 Cloud Run endpoints runs through Judge 6 before dispatching to Gemini, Claude, Grok, or Perplexity. Policy enforcement is model-agnostic.
Judge 6 ships as part of the ShadowTagAI platform. Start with KovelAI to access the full governance stack.
Get Access via KovelAI