August 14, 2026 · guides

Try it: CAN ↔ Open-GMASE ↔ CompliancePulse (demo slice)

Status: Research demo path for the control-plane seam between Confidential AI Network, Open-GMASE, and CompliancePulse.

Deep dives: G-MASE · CompliancePulse AI

Capability	Status
Open-GMASE OPA on tool proposals	Live (open_gmase/tools, open_gmase/can_contracts)
Gate TDC training start	Live (GMASE_TRAINING_GATE, default on)
Gate TDC deploy / predict	Live (GMASE_INFERENCE_GATE, default on)
Decisions in CAN AuditLogs	Live (GMASE_TOOL_DECISION)
Forward decisions → CompliancePulse ingest	Live by default (http://localhost:3001; warn if CP down; COMPLIANCEPULSE_INGEST_URL=false to disable)
E2E asserts OPA + CP ingest	Live (npm run test:e2e:inference)
CompliancePulse multi-tenant SaaS / swarm UI / SPIRE	Still research roadmap

TDC start training, deploy, and predict call open_gmase/can_contracts fail-closed. The same allow/deny is written to CAN AuditLogs and forwarded to CompliancePulse POST /api/v1/audit/ingest. Inference UI shows the policy gate panel; Playwright requires OPA + CP by default.

Three layers, one decision path

flowchart LR
  subgraph CAN["Confidential AI Network"]
    SE["Side effect<br/>train / deploy / predict"]
    AL["AuditLogs<br/>GMASE_TOOL_DECISION"]
  end
  subgraph OG["Open-GMASE Core"]
    OPA["OPA Rego<br/>open_gmase/can_contracts"]
  end
  subgraph CP["CompliancePulse AI"]
    ING["POST /api/v1/audit/ingest"]
    TR["GET /api/v1/audit/trail<br/>external_ingest"]
  end
  SE -->|authorize fail-closed| OPA
  OPA -->|allow / deny| SE
  SE --> AL
  AL -->|default forward<br/>localhost:3001| ING
  ING --> TR

Layer	Role in this demo
CAN	Product surface (contracts, training, inference) + durable AuditLogs
Open-GMASE	Community OPA packs — the inner gate before the side effect
CompliancePulse	Commercial-path ingest receiver — stores the same decision for control-plane / evidence conversations

Scope: CP ingest is an in-memory audit store for the research seam (not multi-tenant SaaS).

Model trained & inference (what the screenshots show)

Governance screenshots in this post still use the fast tabular path (quick gate regression). A vision path is also live for higher visual impact. The product-tour lifecycle uses NLP so labels read well for stakeholders.

	GMASE gate screenshots	Vision (API / Inference app)	Lifecycle / product tour
Task	Tabular classification	Image classification (CIFAR-10)	Text classification (AG News)
Catalog model	e2e-model-tabular-logreg	MODEL-E2E-001 / CNN	e2e-model-nlp-distilbert-quality
Architecture	Logistic regression	TinyCNN (taskType: vision)	DistilBERT / transformer
Dataset	Iris-style CSV (4 floats)	CIFAR-10 subset (~2k train; FakeData if E2E_VISION_FAST)	AG News headlines
Example request	{ "features": [5.1, 3.5, 1.4, 0.2] }	{ "imageBase64": "…" } (32×32 demo PNG)	{ "text": "Wall Street rallies…" }
Example label	setosa	CIFAR name (e.g. airplane)	Business

In all cases the side effect (train / deploy / predict) is gated by Open-GMASE. The UI shows the label and the policy-gate panel. CompliancePulse receives the governance decision, not the pixels or weights (see below).

What you will see

1. A proposed tool call (for example execute_sql with DROP TABLE, or start_training / run_inference) is evaluated by community Rego packs.
2. Allow or deny comes back fail-closed if OPA is unreachable.
3. The decision is written into CAN’s audit trail.
4. On the Inference app, an Open-GMASE policy gate panel shows ALLOW/DENY with package + audit id.
5. Training start returns job.governance the same way.
6. CAN forwards the decision to CompliancePulse ingest by default (non-blocking; warns if CP is down).
7. You can list those events on CP as external_ingest via the audit trail API.

Screenshots (from E2E)

Deploy for inference — OPA authorizes deploy_inference; the governance decision (not the model artifact) is forwarded to CompliancePulse

Inference app — iris feature vector ready (task: tabular, logistic-regression)

Label setosa plus Open-GMASE policy gate ALLOW — same decision ingested by CompliancePulse

The lifecycle product tour uses quality DistilBERT on AG News (prediction Business) and the same gate when OPA is up — see 24-tdc-inference-predict.png on the product tour.

CompliancePulse integration (how to show it)

What is forwarded (and what is not)

Yes — deploy for inference is gated and forwarded. When you click Deploy for inference (or run predict / start training), CAN:

1. Asks Open-GMASE OPA (deploy_inference / run_inference / start_training)
2. Writes GMASE_TOOL_DECISION to CAN AuditLogs
3. Forwards that same decision to CompliancePulse POST /api/v1/audit/ingest (default http://localhost:3001)

What CP stores is the control-plane decision, for example:

Included	Not included
tool_name (deploy_inference, run_inference, …)	Training images / CSV / text corpora
allow / reason / deny-warn lists	Inference imageBase64 / feature vectors
model_id, contract_id, OPA package	Model weights (model.bin)
auditId, source: confidential-ai-network	Prediction logits / raw outputs

So the caption “Deploy for inference — gated by OPA; forwarded to CompliancePulse” means the ALLOW/DENY audit event is forwarded — not that the dataset or image pixels are copied into CP.

Start the three processes

# 1) Open-GMASE OPA
cd open-gmase-core && docker compose up -d

# 2) CompliancePulse ingest receiver (default target for CAN)
cd compliancepulse-ai/backend && npm run dev
# listens on http://localhost:3001

# 3) CAN stack
./start-system.sh
# COMPLIANCEPULSE_INGEST_URL defaults to http://localhost:3001
# Disable: COMPLIANCEPULSE_INGEST_URL=false

Verify ingest after a gated action

After a deploy/predict (or the smoke script below):

# CAN side — decisions in AuditLogs
curl -s 'http://localhost:5001/api/debug/gmase-tool-decisions?limit=5'

# CompliancePulse side — same decisions as external_ingest
curl -s 'http://localhost:3001/api/v1/audit/trail?eventTypes=external_ingest&limit=5'

Expected shape on CP (fields may vary slightly):

{
  "events": [
    {
      "eventType": "external_ingest",
      "action": "ingest:run_inference",
      "result": "success",
      "metadata": {
        "source": "confidential-ai-network",
        "tool_name": "run_inference",
        "allow": true,
        "package": "open_gmase/can_contracts",
        "model_id": "…"
      }
    }
  ]
}

E2E (npm run test:e2e:inference) requires OPA + CP by default and asserts ≥2 ingest events per model (deploy_inference + run_inference).

Run locally (full slice)

cd open-gmase-core && docker compose up -d
cd compliancepulse-ai/backend && npm run dev   # separate terminal
./start-system.sh                             # separate terminal

./scripts/demo-gmase-can-slice.sh

cd frontend
E2E_WAIT_FOR_LOCAL_TRAINING=true BACKEND_URL=http://127.0.0.1:5001 npm run test:e2e:inference

Ad-hoc deny path (CAN debug API):

curl -s http://localhost:5001/api/debug/gmase-opa-health

curl -s -X POST http://localhost:5001/api/debug/gmase-tool-check \
  -H 'Content-Type: application/json' \
  -d '{
    "tool_name": "execute_sql",
    "environment": "production",
    "parameters": { "query": "DROP TABLE users;" },
    "metadata": { "contract_id": "demo-1" }
  }'

curl -s 'http://localhost:5001/api/debug/gmase-tool-decisions?limit=5'
curl -s 'http://localhost:3001/api/v1/audit/trail?eventTypes=external_ingest&limit=5'

Demo path

1. CAN product tour (contracts → training → inference).
2. Inference app: Open-GMASE policy gate ALLOW panel.
3. Same decision in CAN AuditLogs and CompliancePulse audit/trail (external_ingest).
4. Roadmap beyond this seam: multi-tenant CP SaaS UI, swarm agents, SPIRE attestation.

Code & full runbook

Artifact	Location
Runbook	docs/guides/CAN_GMASE_DEMO_SLICE.md
Screenshots	docs/guides/gmase-integration/
Side-effect gate + CP forward	backend/services/gmaseSideEffectGate.js
CompliancePulse ingest API	compliancepulse-ai/ — POST /api/v1/audit/ingest
Smoke script	scripts/demo-gmase-can-slice.sh
Policies	open-gmase-core/execution-guardrails/opa-policies/

Related reading

• Governed AI for the enterprise — CISO overview
• Governing autonomous AI agents
• Unified Governed Agentic SecOps Framework
• Open-GMASE Core
• CompliancePulse AI