Native DRE
Decision Runtime Engine
Adapter matrix

One policy.
Multiple enforcement surfaces.

DRE compiles policy.md into a provider-specific enforcement artifact. The artifact differs by adapter — a steering vector here, a guardrail config there, a judge sidecar elsewhere. What stays constant is the judgment loop: every governed action passes through a claim bound back to the policy text.

01 · residual
Live

Residual Injection (self-hosted)

For models where DRE owns the inference path, a compiled steering vector is added to the residual stream at a chosen layer with a coefficient alpha. The vector is not a universal abstraction — it requires model-specific evals, layer selection, alpha tuning, and a contrast-pair corpus per policy. Feasible via PyTorch forward hooks or llama.cpp hidden-state access.

Enforcement surface
Hidden-state hook at a target layer
Per request
One vector-add inside the forward pass
Strengths
  • Low marginal per-request overhead above base inference
  • Runs entirely inside the enterprise VPC
  • Deterministic given a fixed model, policy text, and compile config
Tradeoffs
  • Requires an open-weight model and a self-hosted inference path
  • Adapter is model-family specific (arch × tokenizer)
  • Vector quality depends on per-model evals, not on policy.md alone
Models
Llama 3.x (llama.cpp)Qwen 2.5MistralGemma
02 · vertex
Live

Vertex Gemini Reasoner

For customers running Gemini on Vertex, DRE compiles policy.md into a Vertex guardrail config plus a structured reasoning contract the model is asked to satisfy. The DRE judge checks the contract output before any tool call is admitted.

Enforcement surface
Vertex guardrail + structured reasoning contract
Per request
Guardrail eval during a single inference
Strengths
  • Low latency inside Google Cloud deployments
  • Native data residency per region
  • Works without model surgery
Tradeoffs
  • Bounded by Vertex guardrail tier pricing
  • Contract enforcement is softer than a hidden-state hook
Models
Gemini 2.x ProGemini 2.x Flash
03 · bedrock
Beta

Bedrock Guardrail

DRE compiles policy.md into a Bedrock guardrail config plus a system-prompt digest. The guardrail evaluates input and output; the digest anchors the model's behavior. The audit chain links action → claim → policy text + digest hash, not a residual vector.

Enforcement surface
Managed guardrail + system-prompt digest
Per request
One guardrail evaluation + one inference
Strengths
  • Reaches Claude via Bedrock and Llama via Bedrock without model surgery
  • Inherits cloud provider content safety baseline
Tradeoffs
  • Per-request cost bounded by guardrail tier pricing
  • Less expressive than a self-hosted hook for nuanced judgment
  • Enforcement is at the claim boundary, not inside the model
Models
Claude on BedrockLlama on Bedrock
04 · openai
Beta

OpenAI Responses + Judge

The closed API generates the response under a DRE-compiled system preamble. A small open-weight model (governed by the self-hosted adapter) runs as a judge over risky claim classes before a tool call is admitted. The judge is how DRE enforces policy when the frontier model exposes no hook.

Enforcement surface
Compiled system preamble + open-weight judge sidecar
Per request
Preamble tokens + judge sidecar call on risky classes
Strengths
  • Keeps reasoning quality of the closed frontier model
  • Judge sidecar carries the enforceable part of the policy
  • Judge sampling rate is tunable per risk class
Tradeoffs
  • Two inference calls for risky action classes
  • Judge behavior must be monitored separately
  • No hidden-state hook — enforcement is at the claim boundary
Models
gpt-4.1gpt-4oo4-mini
05 · anthropic
Planned

Anthropic Direct + Judge

Planned. Compiled policy would emit a Claude tool-use schema whose shape enforces disclosure tokens and claim fields. The judge sidecar would verify claims against the policy text before the tool executes. No shipping code yet.

Enforcement surface
Tool-use schema + open-weight judge sidecar
Per request
Schema-enforced tool call + judge sidecar
Strengths
  • Leverages Claude's structured tool-use surface
  • Claim boundary is the enforcement boundary
Tradeoffs
  • Schema must be regenerated on policy change
  • Two calls for higher-risk action classes
  • Not shipping today
Models
Claude 3.5 SonnetClaude 3.5 Haiku
The through-line

The policy text is the source of truth. Every adapter links action → claim → policy text a human authored.

DRE's long-term goal is compile once, govern everywhere. Across open-weight models, frontier APIs, and enterprise runtimes, the enforcement surface may change. The policy, audit trail, and trust logic should not.