dbl-gateway 0.8.1

DBL Gateway

dbl-gateway

When organizations deploy LLM systems, they quickly need to answer three questions: Who asked what, what was permitted, and what actually happened.

dbl-gateway is the execution boundary for the Deterministic Boundary Layer. It accepts declared intents, enforces policy decisions, and records everything as an append-only, digest-pinned event stream.

Part of the Deterministic Boundary Layer architecture.

Model

Every request passes through a deterministic event chain:

INTENT  ->  DECISION  ->  PROOF  ->  EXECUTION

INTENT records what was asked. DECISION records what policy allowed (normative, digest-pinned). PROOF captures the context release guard -- a digest of the payload sent to the provider. EXECUTION records what the provider returned.

Events are append-only, linked by correlation_id, and digested with SHA-256 over canonical JSON.

Only DECISION events are normative. Execution output never feeds back into policy. Policy logic lives in dbl-policy; the gateway only evaluates and enforces.

Example

POST /ingress/intent
{
  "interface_version": 3,
  "correlation_id": "c-1",
  "payload": {
    "stream_id": "default",
    "lane": "user",
    "actor": "[e-mail]",
    "intent_type": "chat.message",
    "thread_id": "t-1",
    "turn_id": "turn-1",
    "parent_turn_id": null,
    "payload": {
      "message": "hello",
      "thread_id": "t-1",
      "turn_id": "turn-1",
      "declared_tools": ["web.search"],
      "tool_scope": "strict",
      "budget": { "max_tokens": 1024 }
    }
  }
}

The gateway produces up to four events:

INTENT     what was asked
DECISION   what was allowed  (permitted_tools, enforced_budget, intent_index)
PROOF      what will be sent (context release guard, payload_digest)
EXECUTION  what happened     (model output, tool_calls, release_digest)

Read them back:

GET /tail?stream_id=default&since=0

Install

pip install dbl-gateway

Or from source:

pip install -e .

Run

export DBL_GATEWAY_DB=./data/trail.sqlite
export DBL_GATEWAY_POLICY_MODULE=dbl_policy.allow_all
export DBL_GATEWAY_POLICY_OBJECT=policy
export OPENAI_API_KEY=sk-...

dbl-gateway serve --host 127.0.0.1 --port 8010

Providers

OpenAI, Anthropic, and Ollama. Each provider is a stateless adapter exposing execute() and get_capabilities() against a formal ProviderCapabilities schema. Runtime capabilities are self-describing via GET /capabilities.

What This Is Not

• Not a RAG pipeline.
• Not a workflow engine.
• Not a chat UI.

The gateway does not decide what to do. It enforces whether a declared action may execute. Policy rules are defined externally in dbl-policy.

Documentation

Document	Content
ARCHITECTURE.md	Three-surface model, component ownership, invariants
QUICKSTART.md	Setup, first request, reading events
wire_contract.md	Envelope format, tool gating, budget, refs
env_contract.md	All environment variables
CAPABILITIES.md	Runtime capability introspection
INVARIANTS.md	Machine-checkable invariant list
CONTEXT.md	Context resolution and ref classification
HOOKS.md	Pre-commit boundary enforcement
VALIDATION.md	Validation workflow
CHANGELOG.md	Version history

Related Repositories

• deterministic-boundary-layer -- specification and theory
• dbl-observer -- timeline and audit UI
• dbl-chat-client -- event-projection chat UI
• dbl-stack -- one-command full-stack setup

Status

v0.8.1. Chain-of-record with decision lineage (intent_index), context release guard (PROOF events), and policy config digest. Self-describing capabilities via GET /capabilities. Wire contract v3.