Skip to content

model_gateway_safety.md

Latest commit

 

History

History
49 lines (33 loc) · 2.45 KB

model_gateway_safety.md

File metadata and controls

49 lines (33 loc) · 2.45 KB
Raw

Model Gateway Safety

The repository runs locally by default. OpenAI mode is an optional gateway, not a security boundary and not a required dependency for the demo.

Run the model gateway safety gate with:

python -B scripts/dev.py model-gateway-safety

With a real API key, run the live mode proof with:

python -B scripts/dev.py openai-live

Use OpenAI Live Mode Troubleshooting for expected setup, safe failure modes, rollback, and no-secret review rules.

Safety Contract

The gate verifies that:

  • .env.example keeps OPENAI_API_KEY blank
  • both model-facing project routes default to local mode
  • Docker Compose passes OPENAI_API_KEY only as an optional environment value
  • OPENAI_API_KEY references stay in an explicit allowlist
  • both model gateways use the Responses API endpoint
  • gateway calls use structured JSON schema outputs
  • provider/router environment variables must explicitly opt in to OpenAI mode
  • failed API calls, malformed responses, or timeouts return None and fall back to deterministic local behavior
  • the gateways do not print, log, shell out, dynamically execute code, or re-raise API/parse failures
  • the live check requires OpenAI mode to be observed in API responses and still preserve grounded answers, approval requests, and side-effect blocking

Boundary Statement

OpenAI mode changes only the model-facing part of each workflow:

  • Project 1 can ask the model to improve grounded answer wording after permission filtering, unsafe-content filtering, citation selection, and abstention logic already ran.
  • Project 2 can ask the model to classify intent before deterministic tool permissions, approval gates, and side-effect controls run.

The model never becomes the authority for permissions, evidence access, approval execution, audit logging, or eval success.

Technical Review Framing

Use this answer when challenged on API keys or model dependency:

The verified default path is local and deterministic. The optional OpenAI gateway is opt-in through environment variables, the API key stays outside the repo, and failures return None so the app falls back to local behavior. Permissions and side-effect authorization are enforced before or after the model call in application code, so changing the model provider does not weaken the security invariants. When a key is available, `python -B scripts/dev.py openai-live` proves the model-backed path was actually used.