IBM Granite Open-Weight Model Governance Crosswalk

1. Why Open Weights Need Independent Witnessing

You download granite-4.2-30b.gguf. From that moment, you are the deployer under EU AI Act Art. 28. IBM's governance umbrella ended at the download. IBM cannot attest to what you did with those weights after they left their infrastructure.

IBM provides excellent governance tooling: watsonx.governance captures model metadata, FactSheets document lineage, Granite Guardian classifies safety risks. These tools record IBM's view of IBM's platform. But they cannot answer the questions auditors ask about your deployment:

• Are the weights you loaded right now the same ones IBM published?
• What quantization variant are you actually running in production?
• Did every inference pass through your Guardian safety classifier?
• What hardware executed the inference, and was it the hardware you declared?

SWT3 witnesses what actually happened at inference time. It does not replace watsonx.governance. It completes the evidentiary chain by adding the independent, cryptographic attestation layer that no vendor can provide about your own infrastructure. Together, IBM's tools and SWT3 give auditors the full picture. Apart, each has a blind spot.

2. Granite Model Inventory

Model Family	Architecture	Sizes	Context	License	Deployment Formats
Granite 4.2	Dense transformer	3B, 8B, 30B	128K	Apache 2.0	watsonx.ai, Ollama, vLLM, GGUF, SGLang
Granite 4.1 Hybrid	Mamba-2 + Transformer MoE	1B-32B (9B active)	128K	Apache 2.0	watsonx.ai, Ollama, vLLM, GGUF
Granite Guardian 4.1	Dense classifier	8B	128K	Apache 2.0	watsonx.ai, Ollama, vLLM
Granite Guardian HAP	HAP classifier	38M, 125M	N/A	Apache 2.0	CPU-deployable, any runtime
Granite Code	Dense transformer	3B, 8B, 20B, 34B	128K	Apache 2.0	watsonx.ai, Ollama, vLLM, GGUF

All Granite models are released under Apache 2.0, meaning the deployer assumes full governance responsibility upon download. The SWT3 granite-sovereign CLI profile is designed specifically for this scenario.

3. Deployment Path Matrix

Which SWT3 adapter you use depends on how you serve Granite:

Deployment	SWT3 Adapter	CLI Profile	Air-Gap Ready
watsonx.ai (cloud)	wrap_granite / wrapGranite	Default or custom	No (cloud API)
Ollama (local)	Existing OpenAI adapter (auto-detected)	granite-sovereign	Yes
vLLM (self-hosted)	wrap_vllm / wrapVllm	granite-sovereign	Yes
Docker / GGUF direct	Depends on serving layer	granite-sovereign	Yes

4. Procedure Evidence Cards

5. Coverage Matrix

Procedure	Granite 4.2	Granite 4.1 Hybrid	Guardian	Code Models
AI-INF.1 Inference Provenance	✓	✓	✓	✓
AI-MDL.5 Weight Integrity	✓	✓	✓	✓
AI-MDL.6 Adapter Stack	✓	✓		✓
AI-MDL.7 Quantization	✓	✓		✓
AI-GRD.1 Guardrail	✓	✓	✓	✓
AI-GRD.2 Content Safety	✓	✓	✓
AI-SEC.1 Adversarial Detection	✓	✓	✓
AI-HW.1 Hardware Attestation	✓	✓	✓	✓
AI-LIC.1 License Provenance	✓	✓	✓	✓
AI-SBOM.1 AI Bill of Materials	✓	✓	✓	✓
AI-FAIR.1 Bias Detection	✓	✓
AI-DRIFT.1 Model Drift	✓	✓		✓
AI-DATA.1 Training Data	✓	✓	✓	✓
AI-ENV.2 Runtime Isolation	✓	✓	✓	✓
AI-MARK.1 Content Marking	✓	✓		✓
AI-SUPPLY.1 Supply Chain	✓	✓	✓	✓

6. SDK Quick Start

Path A: Native watsonx.ai SDK (cloud or on-prem)

Python

from swt3_ai import Witness
from swt3_ai.adapters.granite import wrap_granite
from ibm_watsonx_ai import Credentials, APIClient
from ibm_watsonx_ai.foundation_models import ModelInference

witness = Witness(
    endpoint="https://sovereign.tenova.io/api/v1/witness",
    api_key="axm_live_...",
    agent_id="granite-risk-model",
)

credentials = Credentials(url="https://us-south.ml.cloud.ibm.com", api_key="...")
model = ModelInference(model_id="ibm/granite-3-8b-instruct",
                       credentials=credentials, project_id="...")

witnessed = wrap_granite(model, witness)  # or witness.wrap(model)

response = witnessed.generate(prompt="Summarize the quarterly risk report.")
# Response is untouched. Anchor minted silently in the background.

witness.flush()

TypeScript

import { Witness } from "@tenova/swt3-ai";
import { wrapGranite } from "@tenova/swt3-ai/adapters/granite";

const witness = new Witness({
  endpoint: "https://sovereign.tenova.io/api/v1/witness",
  apiKey: "axm_live_...",
  agentId: "granite-risk-model",
});

// Assuming watsonx SDK client is initialized
const witnessed = wrapGranite(model, witness);  // or witness.wrap(model)

const response = await witnessed.generate({ prompt: "Summarize the quarterly risk report." });
// Response is untouched. Anchor minted silently in the background.

await witness.flush();

Path B: Granite via Ollama (air-gapped, no new code)

Python

from swt3_ai import Witness
from openai import OpenAI

witness = Witness.from_config("granite-sovereign")

client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
witnessed = witness.wrap(client)  # Auto-detects Ollama, uses OpenAI adapter

response = witnessed.chat.completions.create(
    model="granite3.1-dense:8b",
    messages=[{"role": "user", "content": "Summarize the quarterly risk report."}],
)

witness.flush()

Four lines of witness code. Every inference now has a tamper-evident anchor. The granite-sovereign profile automatically requires signing, hardware attestation, and Guardian evidence.

7. Cross-References

• EU AI Act Crosswalk -- Annex IV documentation requirements for high-risk systems
• ISO 42001 Implementation Guide -- Annex A evidence workflows for AI management systems
• OWASP Agentic Top 10 Crosswalk -- Agent security risks including prompt injection
• CISA SBOM Minimum Elements -- AI SBOM structure guidance
• granite-sovereign CLI profile -- Pre-configured for air-gapped Granite deployments with TPM 2.0 attestation