dbl-gateway 0.9.11

DBL Gateway

dbl-gateway

Inspect and replay LLM governance decisions.

Every request produces a decision chain: what was requested, what was decided, what was executed.

• Inspect any decision: see what was requested, allowed, and executed
• Replay a decision to reconstruct why it was allowed or denied
• Swap the policy, re-evaluate the same input, compare outcomes

No API keys required:

docker compose --profile demo up

Open http://localhost:8010/ui and click Start Demo.

The demo runs scripted scenarios through the full governance pipeline. Use Manual Intent in the observer to submit your own requests — the gateway is a working runtime, not just a demo harness.

How it works

Every request passes through the same deterministic chain:

INTENT  →  DECISION  →  PROOF  →  EXECUTION

INTENT records what was asked. DECISION records what policy allowed — the only normative event. PROOF records what will be sent to the provider. EXECUTION records what happened.

Execution output never feeds back into policy.

Part of the Deterministic Boundary Layer architecture.

Use your own provider

cp .env.example .env        # add one API key (OpenAI, Anthropic, or Ollama)
docker compose up

One active provider is enough. See .env.example for all options.

Install

pip install dbl-gateway

Or from source:

pip install -e .

Reference implementation

dbl-gateway instantiates the Deterministic Boundary Layer (DBL) as a runtime system.

It realizes the event chain INTENT → DECISION → PROOF → EXECUTION, where DECISION events are recorded before any execution and remain the only normative layer. Execution outputs are treated as non-normative observations.

The gateway enforces the core invariants of the model:

• append-only event stream
• strict ordering of DECISION before EXECUTION
• governance input purity (authoritative inputs only)
• deterministic replay under fixed inputs and policy configuration

As such, it serves both as a usable governed gateway and as a reference instantiation of the DBL model.

For empirical validation, including replay equivalence and policy variation benchmarks, see EMPIRICAL_VALIDATION.md.

Observer UI

Open /ui to watch the event chain in real time.

The built-in observer includes the event stream, decision replay, chain verification, a policy inspector for DECISION events, manual intent submission, and the integrated demo controller.

Documentation

• QUICKSTART.md — first request, reading events, UI verification
• ARCHITECTURE.md — surfaces, component ownership, invariants
• EMPIRICAL_VALIDATION.md — optional replay bench and policy-diff notes
• wire_contract.md — envelope format, tool gating, budget, refs
• env_contract.md — all environment variables
• observer.md — UI layout and verification routes
• demo.md — demo modes and scenario definition
• Full manifest