Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
49 changes: 49 additions & 0 deletions docs/contract-additions/inference-gateway-intersection.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,49 @@
# InferenceGateway — the cloud ∩ local intersection (model serving)

**Campaign gap #8, done the right way: start at the intersection, then fan out.**
Both the cloud plane (prophet-platform model-fabric / model-router) and the local plane
(sourceos / agent-machine / noetica) serve inference. Rather than build two stacks, define
the capability **once** — the `InferenceGateway` contract — and have each side *adapt* to it.
This is the same pattern as the [ShellBackendAdapter](./agentic-shell-gitea-backend.md)
(github/gitea) and the [consent plane](./e11-consent-receipts-ux.md) (which already spans both).

## The shared seam
| Piece | Shared contract | Cloud adapter | Local adapter |
|---|---|---|---|
| Serve a model | `InferenceGateway.serve(request) → response` | prophet-platform `model-fabric/runtime-adapter-set` | agent-machine provider activation (Ollama / on-node) |
| Per-call receipt | **`GatewayCallAudit`** (memory-mesh `schemas/gateway-call-audit.schema.json`) — already exists | ✔ emit | ✔ emit |
| Admission | consent plane: inference is purpose-bound (a remote model = `egress`) | ✔ | ✔ |

**Request (shared):** `{ model, purpose, space, caller, input, params }`.
**Response (shared):** `{ output, usage, cost_usd?, receipt_hash }` + a `GatewayCallAudit`
(`callId · call_type · outcome · caller · model · usage · occurred_at · receipt_hash`).
A call with no admitting consent decision, or that can't emit an audit, is **refused fail-closed** —
identical rule on both planes.

## Intersection-first, then fan out
1. **Intersection (build first):** open-weight models that run **both** cloud and local
(Llama, Qwen, DeepSeek, Mistral) + the governance capabilities that span both (consent
plane, System Graph). These are the client-owned **LLM-stable** — the sovereignty answer.
2. **Fan out — cloud-only:** Claude / GPT via API (supplier + component, not client-owned).
3. **Fan out — local-only:** on-node SLMs, gemma-2-9b (Noetica concierge default, no egress).

## Every model + capability gets a surface across the estate
The [Model Catalog & Leaderboard](../surfaces/model-catalog.html) is the intersection cockpit
screen (cloud∩local placement + a **sovereignty × governance-weighted** leaderboard, not raw
capability alone). Each foundation model and OS capability then fans out to a screen in:
**prophet-platform** (catalog + serving), **prophet lattice** (notebooks), and the
**model leaderboards** — all reading the same `InferenceGateway` catalog + `GatewayCallAudit`
stream, so cloud and local rank on one board.

## Positioning (governs-people vs governs-agents)
The leaderboard's sovereignty/governance weighting encodes the strategy: *Claude Enterprise
governs people; SocioProphet governs agents.* Claude is a **supplier** (API), an **addressable
competitor** (workforce layer, not agentic execution), and a **component** (one engine in the
stable). The LLM-stable reduces supplier-as-future-competitor risk. These competitive claims are
**[DRAFT · VERIFY BEFORE USE]** — require Michael/Gus sign-off before prospect/investor use.

## Done-definition
Shared `InferenceGateway` request/response + `GatewayCallAudit` on every call; a local adapter
(agent-machine) and the cloud adapter (model-fabric) both conform; consent-gated + receipted;
the catalog/leaderboard reads one board across planes. Passes the **inference seam** purple-team
(no un-consented or un-audited call returns output).
3 changes: 3 additions & 0 deletions docs/surfaces/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -43,6 +43,9 @@ the real SociOS/GNOME chrome (top bar, launcher card) so they read as in-shell,
- **`lampstand-launcher.html`** — Tier 2, *actionable*. The Spotlight replacement: natural
language → typed intents/entities/relations (the annotation tree) → **governed actions**
(purpose-bound, consent-gated, receipted), ranked by sherlock (IR). Not web search — typed acts.
- **`model-catalog.html`** — Tier 2, *intersection cockpit*. Every foundation model + OS
capability, cloud∩local placement, a sovereignty×governance-weighted leaderboard, and the
SP-vs-Claude-Enterprise positioning (marked DRAFT/VERIFY). Backed by `inference-gateway-intersection.md`.
- **`genesis-flywheel.html`** — Tier 1. The install spine + growth dynamics: the T₀–T₁₁
genesis braid (sealed, four-space colored), the five-phase zero-trust install, and the
connected flywheel topology with correct ΔEP=(P−C)·X−λK / K=k₀·α·cov(A)·φ(P) /
Expand Down
69 changes: 69 additions & 0 deletions docs/surfaces/data/model-catalog.json
Original file line number Diff line number Diff line change
@@ -0,0 +1,69 @@
{
"provenance": "sample",
"generated_at": "2026-08-03T09:32:18Z",
"source": "sample seed (InferenceGateway catalog + GatewayCallAudit leaderboard)",
"models": [
{
"name": "Claude Opus 4.8",
"prov": "Anthropic",
"runs": "cloud",
"ctx": "200k",
"cap": 97,
"sov": "vendor",
"gov": 1,
"note": "supplier + component; not client-owned"
},
{
"name": "Llama 3.3 70B",
"prov": "Meta (open-weight)",
"runs": "both",
"ctx": "128k",
"cap": 86,
"sov": "owned",
"gov": 2,
"note": "the LLM-stable answer — client-owned"
},
{
"name": "DeepSeek-V3",
"prov": "DeepSeek (open-weight)",
"runs": "both",
"ctx": "128k",
"cap": 88,
"sov": "owned",
"gov": 2,
"note": "strong reasoning; client-hostable"
},
{
"name": "gemma-2-9b-it",
"prov": "Google (open-weight)",
"runs": "local",
"ctx": "8k",
"cap": 64,
"sov": "owned",
"gov": 2,
"note": "noetica on-node; fully local"
}
],
"capabilities": [
{
"name": "Consent plane",
"prov": "agent-standards",
"runs": "both",
"ctx": "—",
"cap": 90,
"sov": "owned",
"gov": 2,
"note": "the governance intersection"
},
{
"name": "System Graph",
"prov": "hellgraph (vendored)",
"runs": "both",
"ctx": "—",
"cap": 80,
"sov": "owned",
"gov": 2,
"note": "netwatch ingest"
}
]
}
Loading
Loading