policy.md in. Portable governance artifacts out. Policy layer for enterprise agents.
Frameworks orchestrate. Workflow engines schedule. Policy engines evaluate rules — but none of them bind an agent's behavior to a specific version of company policy.
DRE sits one layer below the framework. It takes your policy.md and compiles it into whichever enforcement surface the model exposes. Every governed action still passes through the DRE judgment loop — claim, validate, act, audit.
The surface changes. The policy, audit trail, and trust logic should not.
Compliance edits a plain-English document. No DSL, no rules engine, no embedded code.
Residual injection for self-hosted models. Structured tool use, guardrails, and judge sidecars for everything else.
Every governed action links back to the exact policy text compliance authored — across providers.
Residual injection for self-hosted models. Structured tool use, guardrails, and judge sidecars for hosted APIs. Each adapter translates the same compiled policy into whichever controls the provider actually exposes.
Hidden-state hook at a target layer
Vertex guardrail + structured reasoning contract
Managed guardrail + system-prompt digest
The enforcement surface changes per adapter. The judgment loop does not. Claim, validate, act, audit — linked back to the policy text a human authored.
If you own policy for an agent program and you're tired of rewriting it for every new model, get in touch.