Skip to content

G1-Proxy Light — one container, and what it leaves out

G1-Proxy Light is the same engine as G1-Proxy, packaged as one container that also serves its own configuration UI. It exists for places where an install must be a single click and a single image: a cloud catalog listing, an evaluation on one machine, an appliance.

The important part of this page is the second half — what it does not have, and why. A profile that quietly omits a capability is worse than one that does not exist, because the first thing a missing capability does is look like a bug.

The rule that keeps the two in step

Engine code is never edited in the packaging repos. There is one canonical codebase, and everything the light profile does differently lives behind one switchGW_LIGHT=1, implemented in gateway/light_surfaces.py. The two products cannot drift apart, because they are the same source.

What it keeps

detection axes all nine, same weights, same thresholds
thinking levels 0 · 1 · 2 (GLAD-G + the GLAD-H cascade)
halluc_closedbook + SLEDGE yes — including recalibration on the model you actually serve
MCP guard (glad.* tools) yes
Causal explainability (MuPAX / DCA) yes
chat transcript export (JSONL) yes — the route is there; see the note on the Oversight screen below
configuration UI served by the proxy itself, on the same port
Applications one — the single Application is seeded at boot

What it does not have

Every row below is a deliberate omission, not an oversight. The reason matters, because it tells you whether you can turn it on.

missing why can I enable it?
GLAD-A / thinking level 3 the checkpoint is 16 GB; it is exactly what this profile exists to leave out no — use G1-Proxy
Knowledge Base (RAG) no embedder, no reranker, no vector store in the image no
Audio input the ASR model is not shipped no
Idle judge / feedback queue no llama.cpp binary and no GGUF in the image yes — mount a binary and a model and set GW_IDLE_JUDGE=1; the capability endpoint then reports it as available
Dashboard, Kill Switch, Audit, Legal, FRIA, Reports, Human Oversight these read the G-1 Studio control-plane backend, which is a different container no — install G-1 Studio
Compliance report generation reportlab / matplotlib are not installed no
Multiple Applications the light profile is single-Application by design no
Bundled documentation site stripped from the UI bundle to keep the image small — (you are reading it)

The Oversight screen, and the export button on it

Download all chats (JSONL) lives on the Human Oversight screen, which the light profile hides by default because the rest of that screen is a Studio surface. The export route itself is served (GET /v1/glad/chat-export, scoped to the Application), so the capability is there either way: call it directly, or unhide the screen with GW_UI_HIDDEN_VIEWS.

Do not trust this page — ask the box

The list above is what the image ships with. What your deployment can do is one request away, and it is read from the artifacts on disk, never declared by hand:

curl -s http://localhost:8080/v1/glad/capabilities | jq '{tiers, max_thinking_level, features, hidden: .ui.hidden_views}'
{
  "tiers": { "glad_g": true, "glad_h": true, "glad_a": false },
  "max_thinking_level": 2,
  "features": { "mcp": true, "xai": true, "closedbook": true, "recalibrate": true,
                "rag": false, "audio": false, "idle_judge": false, "multi_application": false },
  "hidden": ["dashboard", "kill-switch", "knowledge", "reports", "fria", "audit", "legal",
             "feedback", "oversight", "license-tokens", "documentation", "agent-flow"]
}

Two ceilings are reported separately, and conflating them is a support ticket waiting to happen: what the image can do (are the GLAD-H artifacts present?) and what the licence allows. The effective level is the lower of the two.

GW_UI_HIDDEN_VIEWS="a,b,c" overrides the hidden list without rebuilding — an empty string hides nothing.

Which one should I install?

you want install
one machine, one click, no second container G1-Proxy Light
compliance surfaces (audit trail, FRIA, reports, oversight queue) G1-Proxy + G-1 Studio
thinking level 3 (GLAD-A) G1-Proxy
document grounding (RAG) or audio input G1-Proxy
many Applications with separate policies and keys G1-Proxy + G-1 Studio

The installer takes the choice as a flag — see Installer, profile light.

Versions

The light profile is one deployable, so it reports one version: engine and UI live in the same image and cannot be updated separately. The Studio split reports three (control plane, UI bundle, proxy) because those are three images, and they have been out of step with each other before — which is exactly why they are shown apart.