swt3-ai

# SWT3 - Sovereign Witness Protocol for AI

> Don't audit the agent's thoughts. Audit the agent's actions.

[![npm](https://img.shields.io/npm/v/@tenova/swt3-ai)](https://www.npmjs.com/package/@tenova/swt3-ai)
[![PyPI](https://img.shields.io/pypi/v/swt3-ai)](https://pypi.org/project/swt3-ai/)
[![npm downloads](https://img.shields.io/npm/dm/@tenova/swt3-ai)](https://www.npmjs.com/package/@tenova/swt3-ai)
[![PyPI downloads](https://img.shields.io/pypi/dm/swt3-ai)](https://pypi.org/project/swt3-ai/)
[![License](https://img.shields.io/badge/license-Apache%202.0-blue.svg)](LICENSE)

## The Problem

AI agents are making production decisions: approving loans, triaging patients, managing infrastructure, writing code. In 2026, 65% of firms reported AI agent security incidents. Only 14.4% of agents go live with full security approval. When something goes wrong, there is no tamper-proof audit trail. Logs are mutable. Metrics are averaged. Nobody can prove what the agent actually did.

GPAI transparency obligations are enforceable now. EU AI Act high-risk enforcement begins December 2, 2027. NIST AI RMF, SR 11-7, and CMMC impose similar obligations. 72% of enterprises believe they have AI governance but lack actual control. Most teams have nothing but dashboards and hope.

## The Protocol

SWT3 (Sovereign Witness Traceability) is a deterministic witness protocol for AI systems. It intercepts AI actions, hashes the evidence, and anchors cryptographic proof to an immutable ledger. Your code gets the full response. The auditor gets tamper-proof evidence. Raw prompts and responses never leave your infrastructure.

- **Deterministic, not probabilistic.** The witness engine uses fixed logic, not AI, to evaluate compliance.
- **Zero data retention.** Configurable clearing levels strip sensitive content before it leaves your environment.
- **Framework-mapped.** Every anchor maps to EU AI Act articles, NIST AI RMF functions, and federal controls.

## Try It (10 Seconds, No Account)

**Python**
```bash
pip install swt3-ai
python -m swt3_ai.demo
```

**TypeScript**
```bash
npm install @tenova/swt3-ai
npx swt3-demo
```

No API keys. No account. No network calls. You will see the full witnessing pipeline run locally.

## Three Lines to Production

```python
from swt3_ai import Witness
from openai import OpenAI

witness = Witness(endpoint="https://sovereign.tenova.io", api_key="axm_live_...", tenant_id="YOUR_TENANT")
client = witness.wrap(OpenAI())

# Every inference is now witnessed. Your code does not change.
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Summarize this contract"}],
)
```

Works with OpenAI, Anthropic, AWS Bedrock, Vercel AI SDK, LangChain, LiteLLM (100+ providers), and any OpenAI-compatible endpoint (vLLM, Ollama, Azure OpenAI).

## K8s Hardware Attestation

AI compliance doesn't stop at the model layer. Regulators want to know what hardware ran the inference, whether it was authorized, and whether the compute environment changed between audit periods. Today, no platform provides cryptographic proof of which silicon processed which workload. SWT3 does.

Deploy a DaemonSet that discovers accelerator hardware on every node and mints AI-HW.1 attestation anchors. Zero application code changes. One Helm install.

```bash
helm install swt3-witness oci://ghcr.io/tenova-labs/charts/swt3-witness \
  --set config.endpoint="https://sovereign.tenova.io" \
  --set config.apiKey="axm_live_..." \
  --set config.tenantId="YOUR_TENANT"
```

The DaemonSet auto-discovers 6 accelerator types per node:

| Discovery Path | Silicon | Method |
|---------------|---------|--------|
| NVIDIA GPU | A100, H100, H200, B200, GB200, NVL72 | `nvidia-smi` |
| Google TPU | v4, v5e, v5p, v6e, Trillium | `TPU_NAME` env |
| AMD MI | MI300X, MI325X, MI250 | `rocm-smi` |
| AWS Trainium | Trainium2, Inferentia2 | `neuron-ls` |
| Intel Gaudi | Gaudi3, Gaudi2 | `hl-smi` |
| PCI Fallback | Any 3D controller / processing accelerator | `/sys/bus/pci` |

Each node reports its silicon vendor, topology, memory, and per-accelerator detail. Non-accelerator nodes produce a valid anchor attesting "no accelerator detected" -- absence of hardware is also auditable evidence. Mixed-silicon clusters (NVIDIA + TPU + AMD in the same cluster) are fully supported.

**What this proves to your auditor:**
- Which hardware ran each AI workload (EU AI Act Art. 15(4), NIST 800-53 SI-7)
- Whether the compute environment changed between assessments (drift detection)
- That inference didn't silently migrate to unauthorized or unqualified silicon
- Full hardware provenance from silicon to model, combined with AI-HW.3 (TPM attestation)

The witness context includes `silicon_vendor`, `discovery_method`, and `accelerators[]`. All device identifiers are SHA-256 hashed before leaving the node. Serial numbers, UUIDs, and bus IDs are never transmitted in plaintext.

See the [Cross-Silicon K8s Attestation Guide](https://sovereign.tenova.io/guides/cross-silicon-k8s-attestation.html) for GKE, EKS, AKS, and on-prem deployment patterns.

## What Gets Witnessed

Each inference produces anchors across 108 AI procedures spanning 56 namespaces:

| Procedure | Domain | What It Proves | Regulatory Mapping |
|-----------|--------|---------------|-------------------|
| AI-INF.1 | Inference | Prompt and response captured (provenance) | EU AI Act Art. 12 |
| AI-INF.2 | Inference | Latency within threshold (detects model swaps) | NIST AI RMF MEASURE 2.6 |
| AI-MDL.1 | Model | Deployed model matches approved hash (integrity) | EU AI Act Art. 9 |
| AI-MDL.2 | Model | Model version identifier recorded (tracking) | EU AI Act Art. 72 |
| AI-MDL.5 | Model | Weight file SHA-256 verified (tamper detection) | EU AI Act Art. 15(4) |
| AI-MDL.6 | Model | LoRA/PEFT adapter stack attested | EU AI Act Art. 12(2)(b) |
| AI-MDL.7 | Model | Quantization method recorded | EU AI Act Art. 15(3) |
| AI-GRD.1 | Guardrail | Required safety filters were active (enforcement) | NIST AI RMF GOVERN 1.5 |
| AI-GRD.2 | Safety | No content filter or refusal triggered | EU AI Act Art. 14 |
| AI-GRD.3 | Gatekeeper | Pre-call guardrail gate enforced | EU AI Act Art. 9(2) |
| AI-RAG.1 | Retrieval | RAG context chunks and corpus attested | EU AI Act Art. 12(2)(a) |
| AI-RAG.2 | Retrieval | Retrieval relevance scoring verified | EU AI Act Art. 10(2) |
| AI-TOOL.1 | Tool Use | Agent tool/function call recorded (latency, success) | NIST AI RMF MANAGE 4.1 |
| AI-ID.1 | Identity | Witness instance identity attested (agent accountability) | EU AI Act Art. 13 |
| AI-ACC.1 | Access | Resource access granted or denied with scope | EU AI Act Art. 9(4)(c) |
| AI-REV.1 | Revocation | Previously-issued anchor revoked with reason | EU AI Act Art. 12(3) |
| AI-SEC.1 | Security | Adversarial threat detection performed | EU AI Act Art. 15(4) |
| AI-SEC.2 | Security | Input validated and sanitized before inference | EU AI Act Art. 15(3) |
| AI-SKILL.1 | Skills | Loaded skill/tool/plugin manifest attested | EU AI Act Art. 12(2)(b) |
| AI-SKILL.2 | Memory | Active memory sources bound to decision | EU AI Act Art. 12(2)(a) |
| AI-SKILL.3 | Alignment | RLHF/DPO reward model binding recorded | EU AI Act Art. 9(4)(a) |
| AI-CHAIN.1 | Chain | Multi-agent handoff witnessed with cycle tracking | EU AI Act Art. 12(2)(a) |
| AI-VIO.1 | Violation | Policy violation detected during inference | EU AI Act Art. 9(4)(a) |
| AI-CHR.1 | Charter | Agent charter/system prompt hash attested | EU AI Act Art. 13 |
| AI-MDL.8 | Model | Model verified against approved registry | EU AI Act Art. 51 |
| AI-HITL.3 | Oversight | Reviewer identity bound to human review | EU AI Act Art. 12(3)(d) |
| AI-SAFE.1 | Safety | Stop mechanism tested, safe state confirmed | EU AI Act Art. 14(4)(e) |
| AI-HW.1 | Hardware | GPU/accelerator inventory attested at startup | EU AI Act Art. 15(4) |
| AI-TRUST.1 | Trust | Mutual compliance trust verified between agents | EU AI Act Art. 9(4)(c) |
| AI-TRUST.2 | Trust | Trust handshake details recorded | EU AI Act Art. 12(2)(a) |
| AI-MARK.1 | Content | Content provenance marking attested | EU AI Act Art. 50(2) |
| AI-BASE.1 | Baseline | Agent behavioral baseline monitored | NIST AI RMF MEASURE 2.6 |
| AI-ENV.1 | Environment | Runtime environment fingerprint recorded | EU AI Act Art. 15(4) |
| AI-ENV.2 | Environment | Dependency manifest attested | EU AI Act Art. 15(3) |
| AI-DATA.3 | Data | Training data lineage attested | EU AI Act Art. 10(2) |
| AI-DATA.4 | Data | Data quality metrics recorded | EU AI Act Art. 10(3) |
| AI-CHAIN.2 | Chain | Chain-of-trust credential verified | EU AI Act Art. 9(4)(c) |

Plus 70 additional procedures covering fairness, explainability, inference volume, model drift, human oversight, cybersecurity, supply chain, content marking, agent lifecycle, financial transactions, and cross-border routing. See the [full procedure registry](https://sovereign.tenova.io/registry).

## View an Anchor

A Level 1 anchor for AI-INF.1 (Inference Provenance). This is what reaches the witness ledger. No prompts, no responses, just cryptographic proof.

```json
{
  "procedure_id": "AI-INF.1",
  "factor_a": 1,
  "factor_b": 1,
  "factor_c": 0,
  "clearing_level": 1,
  "anchor_fingerprint": "c059eb5938c0",
  "anchor_epoch": 1774800000,
  "fingerprint_timestamp_ms": 1774800000000,
  "ai_prompt_hash": "315f5bdb76d078c4",
  "ai_response_hash": "a1b2c3d4e5f60718",
  "ai_latency_ms": 842,
  "ai_model_id": "gpt-4o",
  "ai_context": {
    "provider": "openai",
    "guardrails": ["content-filter", "pii-redaction"]
  }
}
```

The `anchor_fingerprint` is computed from `SHA256("WITNESS:{tenant}:{procedure}:{fa}:{fb}:{fc}:{ts}")`. Anyone with the factors can independently verify the math. Trust is a vulnerability. Math is the remedy.

## Clearing Levels

The clearing engine controls what leaves your infrastructure. Your code always gets the full response. Clearing only affects what reaches the witness ledger