Open Source · Apache 2.0 · v1.0.0
The notary
for AI decisions.
Razi wraps AI reasoning inside a constraint-governed runtime that enforces policy rules, traces every
claim to evidence, and produces deterministic audit artifacts — replayable offline, six months later.
⚖️
Policy, not prompts
Rules are compiled into the runtime and enforced after the model responds.
policy_compliant is always set by the harness — never trusted
from the model.
🔗
Evidence-traced decisions
Every claim is grounded in a deterministic evidence index built from your input. Hallucinated
evidence IDs are caught structurally, not by hoping the model doesn't lie.
🔁
Replay-proof audit trail
Every run produces an immutable artifact set. Re-evaluate schema, evidence, and policy offline
with razi replay <run_id> — no model call required.