Runtime Boundaries¶

Director-AI has one supported default path and several optional runtime paths. The default path is intentionally Python-only so new users, contributors, and small deployments can get a working guard without installing the full research and acceleration stack.

Supported Default¶

Use this path first:

pip install director-ai[server,vector]
director-ai quickstart --run
director-ai doctor

This path includes:

Component	Status	Purpose
Python package	Supported default	CLI, scorer, config, server, proxy
FastAPI service	Supported default	HTTP API and local deployment surface
Local Chroma persistence	Supported default	Simple persistent vector store
Fact file ingestion	Supported default	One-file knowledge base bootstrap
`director-ai doctor`	Supported default	Runtime stack audit and config warnings

The default must not require Rust, Go, Julia, Lean, TensorRT, or WASM.

Optional Runtime Paths¶

Enable these only when the deployment needs the specific capability.

Runtime	Enable When	Entry Point	Support Boundary
Rust kernel	Low-latency local scoring is required	`pip install director-ai[rust]`	Optional acceleration; Python fallback remains supported
ONNX	Local NLI inference is required	`pip install director-ai[nli,onnx]`	Optional model runtime; requires exported model artefacts
TensorRT	NVIDIA GPU throughput is required	`pip install director-ai[tensorrt]`	Optional GPU optimisation path
Go gateway	Dedicated gateway/risk-router deployment is required	gateway binary or container	Enterprise/research deployment path
Julia tuner	Research threshold sweeps are required	Julia scripts	Research path, not onboarding
Lean verifier	Formal proof checks are required	Lean backend	Research/formal-verification path
WASM	Browser or edge halt checks are required	`make wasm-build` / `backfire-wasm`	Prioritised edge/offline path; host supplies scores

For package extras and common combinations, use the advanced backend matrix.

Research Hook Boundary¶

Research hooks are not part of the supported default path. Load them through the explicit namespace and gate:

from director_ai import experimental

experimental.enable_experimental_hooks()
trajectory = experimental.load_hook("trajectory")

Non-interactive deployments may use:

export DIRECTOR_AI_ENABLE_EXPERIMENTAL_HOOKS=1

This boundary covers trajectory simulation, trace-safe checks, causal verification, irreversibility forecasting, symbolic and ontology checks, multimodal guards, self-monitoring/adaptation modules, swarm risk modules, provenance/privacy research backends, and sustainability policy modules. Direct director_ai.core.<module> imports remain for compatibility and internal tests, but new public examples should use director_ai.experimental.

Contributor Boundary¶

Good first issues and routine fixes should stay in the Python-only surface:

src/director_ai/
tests/
docs-site/
CLI, config, server, proxy, scorer, and ingestion code

Changes to Rust, Go, Julia, Lean, TensorRT, or WASM should be labelled as advanced runtime work and should include contract tests against the Python path. Optional runtime changes must not make the default install heavier.

Operational Boundary¶

Production operators should choose one runtime stack per deployment:

Deployment Need	Recommended Stack
First production trial	Python + FastAPI + Chroma
Local NLI scoring	Python + FastAPI + Chroma + ONNX
GPU throughput	Python + ONNX/TensorRT container
Gateway risk routing	Python API + Go gateway
Browser/edge halt checks	WASM halt kernel + host-owned scorer
Formal proof checks	Python verifier + Lean backend

Run director-ai doctor after setting DIRECTOR_* environment variables. It warns when config requests optional components that are not installed, for example DIRECTOR_SCORER_BACKEND=onnx without onnxruntime or an ONNX path.