diff --git a/docs/contract-additions/inference-gateway-intersection.md b/docs/contract-additions/inference-gateway-intersection.md index 49ada9e..4a17200 100644 --- a/docs/contract-additions/inference-gateway-intersection.md +++ b/docs/contract-additions/inference-gateway-intersection.md @@ -47,3 +47,20 @@ Shared `InferenceGateway` request/response + `GatewayCallAudit` on every call; a (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). + +## The governed model plane — beyond foundation models +The catalog is not only LLMs. Every **business target** (fraud, churn, credit, AML, propensity) +is a governed model with: +- **Champion / challenger** — the production champion and shadow challenger(s), each with a + metric (AUC / PR-AUC / Gini / recall). A challenger is promoted only through the eval gate + (SHACL + eval, fail-closed); a promotion is a `RunReceipt` — replayable, examiner-ready. +- **Historic comparison** — metric by version, champion vs challenger over time; model lineage. +- **Drift** — PSI per target; a breach flags the target and can trip the automaton. +- **DataClass classifiers** — the [trained+registered](./2026-05-agent-machine-model-carry-contracts.md) + data-catalog classifiers governing each target's data (transaction / merchant / PII / bureau / + protected-attribute / jurisdiction …), each with coverage + registration status. + +Surface: [Model Governance Registry](../surfaces/model-governance.html) — the champion/challenger + +historic + classifier view. Regulated targets (credit, AML) require examiner sign-off before +promotion. This is model risk management (APRA CPS 230 / OCC model-risk) rendered as a witness +surface — the governs-agents thesis made examinable. diff --git a/docs/surfaces/README.md b/docs/surfaces/README.md index d6eff78..df51ebe 100644 --- a/docs/surfaces/README.md +++ b/docs/surfaces/README.md @@ -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-governance.html`** — Tier 2, *regulated model registry*. Business targets (fraud/churn/ + credit/AML/propensity) with champion-vs-challenger, historic version comparison, PSI drift, and the + DataClass classifiers per target. Model risk management as a witness surface. - **`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`. diff --git a/docs/surfaces/data/model-governance.json b/docs/surfaces/data/model-governance.json new file mode 100644 index 0000000..7225189 --- /dev/null +++ b/docs/surfaces/data/model-governance.json @@ -0,0 +1,105 @@ +{ + "provenance": "sample", + "generated_at": "2026-08-03T09:41:30Z", + "source": "sample seed (governed model plane ledger + DataClass classifier registry)", + "targets": [ + { + "id": "fraud", + "name": "Card fraud", + "industry": "payments · real-time", + "champion": { + "name": "GBM-fraud", + "ver": "v4", + "met": "AUC", + "val": 0.962 + }, + "challenger": { + "name": "GNN-fraud", + "ver": "v1", + "met": "AUC", + "val": 0.971 + }, + "verdict": "promote", + "drift": 0.08, + "driftWarn": false, + "history": [ + { + "l": "v3", + "c": 0.955, + "h": 0.958 + }, + { + "l": "v4", + "c": 0.962, + "h": 0.971 + } + ], + "classifiers": [ + { + "n": "transaction-class", + "d": "txn type · MCC", + "st": "registered", + "cov": "99.8%" + }, + { + "n": "merchant-risk-class", + "d": "reputation tier", + "st": "registered", + "cov": "97.2%" + }, + { + "n": "pii-class", + "d": "PII tagging", + "st": "registered", + "cov": "100%" + } + ] + }, + { + "id": "credit", + "name": "Credit risk", + "industry": "lending · regulated", + "champion": { + "name": "Scorecard", + "ver": "v12", + "met": "Gini", + "val": 0.58 + }, + "challenger": { + "name": "LGBM-credit", + "ver": "v3", + "met": "Gini", + "val": 0.61 + }, + "verdict": "promote", + "drift": 0.11, + "driftWarn": true, + "history": [ + { + "l": "v11", + "c": 0.56, + "h": 0.59 + }, + { + "l": "v12", + "c": 0.58, + "h": 0.61 + } + ], + "classifiers": [ + { + "n": "bureau-class", + "d": "bureau tier", + "st": "registered", + "cov": "99.4%" + }, + { + "n": "protected-attr-class", + "d": "fairness guard", + "st": "registered", + "cov": "100%" + } + ] + } + ] +} diff --git a/docs/surfaces/model-governance.html b/docs/surfaces/model-governance.html new file mode 100644 index 0000000..61367c7 --- /dev/null +++ b/docs/surfaces/model-governance.html @@ -0,0 +1,194 @@ +SociOS — Model Governance Registry + + + +
+

Model Governance Registry

champion · challenger · historic · classifiers + foundation models → + +
+

Every business model target — fraud, churn, credit, AML, propensity — with its champion in production, its challenger in shadow, the historic version comparison, and the DataClass classifiers governing each target's data. Every promotion is a receipted, replayable, examiner-ready decision — this is model risk management as a witness surface, not a slide.

+ +
+
+

Business targets

+
+
+
+

Champion / Challenger

+
+
+
+ +

+ Reads the governed model plane (labs · eval · ledger · router) + the DataClass classifier registry (specified→trained→reproduced). Promotions gate through SHACL+eval, fail-closed; each is a RunReceipt. Regulated targets (credit, AML) require examiner sign-off before promotion.
+ Champion/challenger + historic + data lineage = APRA CPS 230 / model-risk-ready. +

+
+ + diff --git a/tools/build_surface_feed.py b/tools/build_surface_feed.py index 3258b48..092a111 100644 --- a/tools/build_surface_feed.py +++ b/tools/build_surface_feed.py @@ -19,12 +19,14 @@ "e11": ["surfaces", "receipts", "governor"], "turn-witness": ["turns"], "model-catalog": ["models", "capabilities"], + "model-governance": ["targets"], } SOURCE_HINT = { "b11": "netwatch snapshot + guardrail-fabric transitions", "e11": "consent-plane catalog (spaces_v1.yaml) + receipt stream + Governor queue", "turn-witness": "App-Intents parser output", "model-catalog": "InferenceGateway catalog + GatewayCallAudit leaderboard", + "model-governance": "governed model plane ledger + DataClass classifier registry", }