Who this is for: OpenShell partner managers, integration leads, sales engineers, and anyone building on top of NVIDIA's sandboxed agent runtime who needs to explain compliance evidence to their customers.
This is not a technical crosswalk. This page explains why SWT3 matters for your customers and how it complements your enforcement tools. For the deep technical mapping -- OCSF event classes, procedure IDs, clearing levels, and code examples -- see the NVIDIA OpenShell Technical Crosswalk.
1. The Evidence Gap
Your customers deploy OpenShell to contain agent behavior. The sandbox works. It blocks unauthorized network calls, restricts file access, enforces tool permissions. None of that is in question.
The question is: when an auditor asks "prove that Agent-47 was sandboxed on Tuesday at 14:00 UTC," where is the evidence?
Enforcement logs exist, but they sit on local disk. They have no cryptographic integrity. They cannot be verified by a third party. An auditor must trust the operator -- and trust is not evidence.
| Aspect | Enforcement Only | Enforcement + Evidence |
|---|---|---|
| What exists after an event | Log entries on local disk | Cryptographic witness anchor with SHA-256 fingerprint |
| Integrity guarantee | None -- logs can be edited or deleted | HMAC-signed, fingerprinted, independently verifiable |
| Third-party verification | Not possible without system access | Any party can verify with the published formula |
| Compliance reporting | Manual screenshots and narrative | Auditor-ready exports mapped to 80 regulatory frameworks |
| Trust model | Auditor trusts the operator | Evidence is independently attestable |
Key distinction: SWT3 does not enforce anything. It does not intercept, block, or modify agent behavior. It records what your tool already enforces. This is the distinction that matters to auditors, and the distinction that sets your integration apart from pure enforcement tools.
2. What is SWT3?
SWT3 (Sovereign Witness Traceability) is an open protocol for recording cryptographic evidence of AI system behavior. Think of it as TLS for accountability -- infrastructure that any tool can plug into, not a vendor product that competes with yours.
When your tool makes a decision -- allow a network call, deny file access, block a guardrail violation -- SWT3 records a witness anchor: a signed, timestamped, fingerprinted attestation of what happened. The anchor is immutable and independently verifiable. The formula that produces the fingerprint is published and locked across 10 SDK languages.
SWT3 speaks OCSF natively. The OpenShell adapter consumes OCSF v1.7.0 events and automatically maps them to the correct accountability procedure. No schema translation, no custom integration code, no runtime dependency.
Key Terms
- Witness Anchor: A cryptographic attestation record -- signed, fingerprinted, mapped to a specific accountability procedure. Not a log entry. Not a metric. An independently verifiable proof that something happened.
- Procedure: A specific accountability obligation. For example, AI-SHELL.1 records that a sandboxed runtime enforced its containment policy during a specific observation window. There are 279 procedures covering everything from inference witnessing to agent delegation.
- Clearing Level: A privacy tier that controls what metadata survives in the anchor. L0 (analytics) keeps everything. L3 (classified) strips all identifying information while preserving the verdict. Your customers choose the level that matches their data sensitivity.
3. Three-Layer Architecture
Every compliance question has three layers. Your tool handles Layer 1. SWT3 handles Layer 2. Auditors and regulators consume Layer 3.
OpenShell
Your sandbox enforces containment policies. It blocks unauthorized network calls, restricts file access, manages tool permissions. This is the action.
SWT3
SWT3 records each enforcement decision as a witness anchor. SHA-256 fingerprinted, HMAC-signed, clearing-level-aware. This is the proof.
Auditor
Auditors query the evidence layer to verify compliance posture. Framework-mapped to EU AI Act, NIST AI RMF, CMMC, and 77 more. This is the verdict.
Your customers need all three layers. You already deliver Layer 1. SWT3 adds Layer 2 with zero changes to your enforcement logic. The three-layer diagram is the clearest way to explain the integration to a customer's compliance team.
4. How It Works
SWT3 stays out of the critical path. It reads events after they occur. Zero runtime dependencies on your sandbox. Zero performance impact on sandbox operations. If SWT3 goes down, your sandbox keeps running exactly as before.
OCSF Event Mapping
The OpenShell adapter maps OCSF v1.7.0 event classes to SWT3 procedures automatically:
| OCSF Event Class | Disposition | SWT3 Procedure | What It Records |
|---|---|---|---|
network_activity | allow / deny | AI-INF.1 / AI-SEC.1 | Inference call permitted or blocked |
process_activity | launch / tool_call | AI-ID.1 / AI-TOOL.1 | Agent lifecycle and tool invocation |
file_activity | allow / deny | AI-DATA.1 / AI-SEC.1 | Data access permitted or blocked |
configuration_change | update | AI-MDL.2 | Policy mutation |
security_finding | guardrail_allow / deny | AI-GRD.1 | Guardrail passed or blocked |
Integration
The adapter is a consumer, not a proxy. It reads your OCSF events from a log file, a stream, or a queue. Your sandbox never knows SWT3 exists.
from swt3_ai.adapters.openshell import OpenShellWitness
osw = OpenShellWitness()
osw.process_event(ocsf_event) # That's it.
# Or tail a log file in real time:
import asyncio
asyncio.run(osw.watch_log("/var/log/openshell/sandbox-001.jsonl"))
import { OpenShellWitness } from "@tenova/swt3-ai/adapters/openshell";
const osw = new OpenShellWitness(witness);
osw.processEvent(ocsfEvent); // That's it.
// Or consume from an async stream:
await osw.watchStream(eventIterator);
No OCSF? No problem. If your sandbox does not emit OCSF events, use the one-liner for periodic attestation:
witness.witness_runtime_containment(
runtime_type="openshell",
policy_hash=policy_sha256,
violation_count=0,
observation_window_ms=300000
)
5. What Your Customers Get
Every sandbox decision becomes a verifiable anchor
No more manual log screenshots. No more narrative explanations of what your sandbox did. Compliance teams export framework-mapped reports in one click. Auditors verify evidence independently without contacting any vendor.
80 frameworks out of the box
SWT3 maps your enforcement decisions to EU AI Act Article 9 (risk management), NIST AI RMF MAP 1.1 (risk identification), CMMC AC-2 (access control), NERC CIP-007 (system security management), and 76 more. Your containment enforcement automatically generates evidence for all of them. No custom compliance code required.
Passive witness -- never touches the enforcement path
SWT3 is a passive observer. It never intercepts, blocks, or modifies agent behavior. If SWT3 goes down, your sandbox keeps running. If your sandbox changes APIs, the OCSF contract stays stable. The evidence layer and the enforcement layer fail independently.
Verifiable without contacting any vendor
Anchor fingerprints are SHA-256 hashes computed with a published, locked formula. The formula is identical across 10 SDK languages. Any party can verify independently. Customers can self-host the protocol. Auditors can verify evidence offline. No phone-home, no license check, no vendor dependency.
6. Also Works With
SWT3's runtime containment procedure (AI-SHELL.1) is duck-typed. OpenShell is the headline integration, but the same procedure works with any sandbox that emits OCSF events or provides summary statistics.
| Runtime | Integration Path | Status |
|---|---|---|
| NVIDIA OpenShell | Full OCSF adapter (7 event classes) | Production |
| gVisor (Google) | AI-SHELL.1 periodic attestation | Production |
| Kata Containers (CNCF) | AI-SHELL.1 periodic attestation | Production |
| Firecracker (AWS) | AI-SHELL.1 periodic attestation | Production |
| WASM runtimes | AI-SHELL.1 periodic attestation | Production |
| Custom | runtime_type="custom" | Production |
If your product uses a different sandbox runtime, the same evidence layer applies. The procedure does not prescribe the enforcement mechanism -- it records what the mechanism did.
7. Next Steps
For Integration Engineers
Install the SDK and point it at your OCSF event stream. The adapter handles procedure mapping automatically.
- Python SDK --
pip install swt3-ai - TypeScript SDK --
npm install @tenova/swt3-ai - Technical Crosswalk -- full OCSF mapping, clearing levels, code examples
For Partner Managers
Share this page with your customer's compliance team. The three-layer diagram (Section 3) is the fastest way to explain the integration value. For customers who need the deep technical detail, send the Technical Crosswalk.
For Sales Engineers
Use the before/after comparison (Section 1) in customer conversations. The question "prove that your sandbox was enforced last Tuesday at 14:00 UTC" is the one that opens the door. The answer is: enforcement + evidence.
- UCT Procedure Registry -- browse all 279 procedures
- Anchor Verification -- live verification demo
- Protocol Specification -- v2.0.0
Questions or partnership inquiries: partners@tenovaai.com
8. FAQ
Is SWT3 a security tool?
No. SWT3 is an independent witness protocol. It records evidence of what your security tools enforce. It never blocks, intercepts, or modifies agent behavior. Your tools are the enforcement layer. SWT3 is the evidence layer.
Does SWT3 add latency to sandbox operations?
No. SWT3 reads events after they occur. It stays out of the critical path. The adapter is a consumer, not a proxy. Zero performance impact on sandbox operations.
What if my sandbox does not emit OCSF events?
Use witness_runtime_containment() to mint periodic attestation anchors from summary statistics (policy hash, violation count, observation window). No OCSF required. The procedure is duck-typed and works with any runtime that can provide those three values.
Do customers need a TeNova account?
A free Open-tier account is available for testing and evaluation. Production deployments use Pro or Enclave tiers depending on clearing level requirements. The SDK works in local mode (no account) for development and CI/CD.
Can SWT3 evidence be verified without contacting TeNova?
Yes. Anchor fingerprints are SHA-256 hashes computed with a published, locked formula. The formula is identical across 10 SDK languages. Any party can verify independently using the SDK, the public verification page, or their own implementation of the formula. No phone-home required.