Native DRE
Decision Runtime Engine
Native DRE

The governance runtime
for enterprise agents.

policy.md in. Portable governance artifacts out. Policy layer for enterprise agents.

REC · dre cli · scripted replay
compile → serve → claim → allow / deny
$
The gap

Every enterprise is building agents.
Nobody is governing them with portable policy.

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.

What DRE does

One source of policy. Enforcement wherever the model runs.

Author once
policy.md

Compliance edits a plain-English document. No DSL, no rules engine, no embedded code.

Enforce per surface
Per adapter

Residual injection for self-hosted models. Structured tool use, guardrails, and judge sidecars for everything else.

Audit that survives
Claim → policy

Every governed action links back to the exact policy text compliance authored — across providers.

Enforcement surfaces

Three surfaces live today. More behind them.

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.

residual
Live

Residual Injection (self-hosted)

Hidden-state hook at a target layer

Per request
One vector-add inside the forward pass
vertex
Live

Vertex Gemini Reasoner

Vertex guardrail + structured reasoning contract

Per request
Guardrail eval during a single inference
bedrock
Beta

Bedrock Guardrail

Managed guardrail + system-prompt digest

Per request
One guardrail evaluation + one inference
The judgment loop

Every governed action flows through the same seven stages.

The enforcement surface changes per adapter. The judgment loop does not. Claim, validate, act, audit — linked back to the policy text a human authored.

  1. 01
    Insight
    Signal arrives
  2. 02
    Awareness
    Context loaded
  3. 03
    Reasoning
    Agent synthesizes
  4. 04
    Judgment
    DRE validates claim
  5. 05
    Action
    Tool invoked
  6. 06
    Comms
    Response returned
  7. 07
    Feedback
    Trace written
Enterprise scope

Three programs Native DRE is built for.

Consumer Support
Telecom · Customer Care
Scope
Agent-driven call deflection with escalation policy
Scale
High-volume consumer channel
Regulatory
State PUC + consumer protection
Payroll & HR
HR · Benefits
Scope
Payroll agent with audit trail on every action
Scale
Multi-tenant employer platform
Regulatory
SOC 2 · state labor regs
Telecom Ops
Network Ops · Multi-country
Scope
Network ops agent with per-region guardrails
Scale
Multiple country regimes
Regulatory
Regional DPAs
The claim

Compile once. Govern everywhere.

If you own policy for an agent program and you're tired of rewriting it for every new model, get in touch.