A pattern catalog for developers and engineering leads. No code here, just the menu of what is possible and what each pattern proves.
SWT3 is an industry-agnostic cryptographic witness protocol. These patterns work with any AI provider, any framework, any deployment model. The SDK is published on PyPI and npm under Apache 2.0. Nothing here is vendor-specific. Every pattern produces tamper-evident evidence that satisfies multiple regulatory frameworks simultaneously.
SWT3 works as a transparent overlay on your existing AI client. It supports OpenAI, Anthropic, AWS Bedrock, LiteLLM (100+ providers), Ollama, vLLM, Google ADK, CrewAI, Microsoft Foundry, and any custom client you build.
The SDK does not modify your AI responses. It observes, hashes, and generates tamper-evident evidence alongside your existing workflow. Your application logic stays unchanged. The witness layer runs in parallel, producing cryptographic anchors that auditors and regulators can independently verify.
Whether you run a single-model chatbot or a multi-agent orchestration pipeline, the same patterns apply. Instrumentation is additive. You can start with one pattern and expand coverage as your compliance requirements grow.
Transparent proxy around any AI client. Wrapping your client adds a witness layer that records every inference without changing the response. Each call generates a tamper-evident anchor containing the model identifier, timestamp, and evidence factors.
Witness document retrieval alongside inference. Records which documents were retrieved, their content hashes, and relevance scores. Proves that the AI system's responses were grounded in specific source material.
Record every tool call an agent makes. When your agent calls external tools such as APIs, databases, or file systems, each call is witnessed with the tool name, input hash, and result hash.
Bind a persistent cryptographic identity to an agent's actions. Every anchor minted by this agent carries its identity fingerprint, creating an unbroken chain of attribution across sessions and deployments.
Pre-inference authorization check. Before every inference, the SDK verifies that the requesting identity has permission to use this model with this data at this clearing level. The authorization decision is witnessed as part of the anchor.
Recall a previous attestation. If a model is recalled, data is contaminated, or a policy violation is discovered after the fact, the SDK can mint a revocation anchor that cryptographically links to the original. Seven reason codes are supported, from model recall to regulatory order.
Record model version, weights hash, and adapter stack. Proves which exact model artifact was running at inference time, preventing disputes about what was deployed when. Covers the full model lifecycle from training to serving.
Run the SWT3 doctor in your deployment pipeline. Validates SDK configuration, checks connectivity, and reports coverage gaps before code ships. Catches missing instrumentation before your auditor does.
Each pattern you adopt closes a gap in your auditor's assessment. When you add client wrapping, your auditor sees AI-INF.1 checked off in their printable checklist. When you add RAG provenance, AI-RAG.1 and AI-RAG.2 appear as verified. The assessment mapping shows exactly which regulatory requirements each procedure satisfies across 16 frameworks, including the EU AI Act, NIST AI RMF, CMMC, and SR 11-7.
Your auditor follows a framework-specific walkthrough guide that maps these procedures to their regulatory obligations. The evidence you generate through these patterns flows directly into the auditor's verification workflow. No manual evidence collection, no screenshot folders, no spreadsheets.
Start with the patterns that address your most pressing compliance requirements. Every additional pattern you adopt expands your evidence coverage without disrupting patterns already in place.