An Optional Inference Mode.
Not a Required Replacement.
Vanguard is ExergyNet's OpenAI-compatible execution endpoint. It can run alongside your existing model provider, serving specific calls, without requiring you to remove anything you already use.
Six direct answers, before anything else.
What does it do?
Executes inference calls against ExergyNet-hosted models through an OpenAI-compatible API — point your existing client at a different base URL and model name.
When would I use it?
When you want a stateless-session inference path with a disclosed data-retention policy, for calls where you choose to route to it — not as a default for your whole pipeline.
Does it change my model?
For the calls you route to it, yes — those calls are served by Vanguard's own models, not your existing provider. Calls you don't route to it are unaffected.
Does it route my model?
No. Vanguard does not sit in front of your existing provider and forward calls to it. It is a separate, optional endpoint you choose to call directly.
Does it own my state?
No. Vanguard's session policy is stateless by design — inputs are not retained after the connection closes and are not used for training. Your organization's authoritative state lives in xLMP/Exergy Vault, a separate layer.
What's production vs. proposed?
The OpenAI-compatible API and the stateless-session policy are live. See the maturity table below for every specific capability.
Runs Alongside What You Already Have
No new SDKs to learn. Vanguard speaks the OpenAI-compatible API, so it can run alongside your existing model provider — route specific calls to Vanguard without removing what you already use, by changing the base URL and model name for those calls only.
client = OpenAI(
base_url="https://api.exergynet.org/v1",
api_key=EXERGYNET_API_KEY
)
# This specific call now routes to Vanguard.
# Calls you don't change keep using your existing provider.
response = client.chat.completions.create(
model="vanguard-engine",
messages=[...]
)Stateless by Design
Inputs to Vanguard are not retained after the connection closes, by policy, and are not used for training. This is a policy assertion, not an independently audited compliance certification — if your evaluation requires a formal audit reference (e.g. SOC 2), request current documentation directly rather than assuming certification from this page.
What's Actually Shipped
| Capability | Status | Boundary |
|---|---|---|
| OpenAI-compatible API | Production | Live endpoint, two-line client integration |
| Stateless session policy | Production | Policy assertion; no third-party audit certificate published on this page |
| Model routing / orchestration | Production | Proposes actions; does not self-authorize consequential steps |
| Coexistence with external providers | Architectural | Technically supported per-call; not a managed multi-provider router today |
Try it on the calls you choose.
Pair the trial with a real, dated benchmark rather than taking latency or cost claims on faith.