Sovereign router

Route to the right model — only if it verifies.

The public planner classifies your request and shows candidate scoring without performing inference. The OpenAI-compatible completion path is separate: it refuses to serve unless authoritative release evidence and a configured endpoint both pass. A successfully admitted turn can return a content-free signed receipt.

Optimize for

Point any OpenAI client at it

POST /api/router/v1/chat/completions
{
  "model": "auto",          // or pin a model id
  "prefer": "balanced",     // quality | balanced | cost | speed
  "messages": [{ "role": "user", "content": "…" }]
}

→ X-LeanLogix-Routed-Model: <served model>
→ X-LeanLogix-Gate: passed   (Ed25519 passport re-verified)
→ X-LeanLogix-Receipt-SHA256: <content-free router.turn/v1 body>
→ final SSE meta: { leanlogix_receipt: { receipt, signed, body_sha256 } }

If serving is configured, text comes from that endpoint only after the passport gate passes. The current contract does not claim endpoint-to-weight or physical-boundary attestation; the public planning result alone is never evidence that inference occurred.