How to Witness Kimi K3 on DigitalOcean Inference Engine with Cryptographic Compliance Evidence

1. The Serverless Evidence Gap

When you deploy a model on your own infrastructure, you have access to server logs, request traces, GPU metrics, and container state. You have too much data and need to distill it into compliance evidence.

Serverless inference is the opposite problem. DigitalOcean manages the infrastructure. You send a request, get a response, and the serving environment is ephemeral. There are no server logs to subpoena, no container state to inspect, no GPU metrics to export. The infrastructure that ran your inference does not belong to you.

Under EU AI Act Article 12, deployers must maintain logs that demonstrate the AI system operated as intended. Under Article 26, deployers of GPAI-based systems bear transparency obligations independent of the model provider. DigitalOcean provides an excellent inference platform, but platform uptime is not compliance evidence.

The solution is straightforward: wrap your inference calls with a witness that captures the provenance of every request at the application layer, where you have full control regardless of who owns the GPU.

2. Model Routing and the Provenance Problem

DigitalOcean Inference Engine supports 70+ models across serverless, dedicated, and routed deployments. The platform includes inference routing that can match requests to optimal models based on cost, latency, and capability.

This creates a specific governance challenge: when a router selects the model, the developer may not know which model actually served a given request. The model you requested and the model that responded may differ. For compliance purposes, the model that actually ran is the one that matters.

witness.wrap() solves this by capturing the model field from the API response, not the request. If a router sends your request to deepseek-v4-flash instead of kimi-k3, the witness anchor records what actually happened.

3. 3-Line Integration

DigitalOcean Inference uses OpenAI-compatible endpoints. Point the standard OpenAI SDK at DO's base URL and wrap with SWT3.

Python

import os
from openai import OpenAI
from swt3_ai import Witness

client = OpenAI(
    base_url="https://inference.do-ai.run/v1",
    api_key=os.environ["DO_API_KEY"]
)
witness = Witness(agent_id="do-inference-prod")

response = witness.wrap(client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": prompt}]
))

TypeScript

import OpenAI from 'openai';
import { Witness } from '@tenova/swt3-ai';

const client = new OpenAI({
  baseURL: 'https://inference.do-ai.run/v1',
  apiKey: process.env.DO_API_KEY
});
const witness = new Witness({ agentId: 'do-inference-prod' });

const response = await witness.wrap(client.chat.completions.create({
  model: 'kimi-k3',
  messages: [{ role: 'user', content: prompt }]
}));

The witness adds sub-millisecond overhead. No proxy, no sidecar, no infrastructure changes. The base_url stays pointed at DigitalOcean. The evidence stays with you.

4. What Gets Witnessed

At Clearing Level 2 and above, prompt and response content are excluded from the witness anchor. The factors above are metadata only. No PII passes through the witness pipeline.

5. Procedure Coverage

6. Quick Reference for Examiners

7. References

• DigitalOcean Inference Engine Documentation
• Kimi K3 Self-Hosted Compliance Witnessing (the companion guide for on-premises deployments)
• Self-Hosted Quickstart (air-gapped and private endpoint deployments)
• NVIDIA NeMo Guardrails Witnessing (guardrail evidence for open-weights models)
• SDK Documentation (9 languages, 21 adapters)
• Machine-Readable Crosswalks (JSON)