Anthropic Claude Compliance Witnessing

v0.7.9 RELEASED – Anthropic adapter with streaming witnessing

1. The Evidence Gap

Anthropic provides substantial safety infrastructure for Claude models. The Responsible Scaling Policy (RSP) classifies models by capability level and applies corresponding safety commitments. Model cards document known limitations. Usage policies define acceptable use. This is the provider's responsibility, and Anthropic does it well.

But none of this addresses the deployer's evidence obligation. When your application calls Claude, you need proof that the specific inference was governed -- not that the model was generally safe when Anthropic evaluated it months ago.

Anthropic Provides	You Need	Gap
RSP capability evaluations at training time	Per-inference evidence at runtime	No runtime evidence
Model cards with known limitations	Proof that limitations were managed in your deployment	No deployment-specific attestation
Usage policies and acceptable use terms	Evidence that your usage complied with those terms per call	No per-call compliance record
API response with model version string	Immutable record tying model version to specific input/output	API logs are mutable
Safety training (Constitutional AI, RLHF)	Evidence that safety training held for your specific use case	No use-case-specific evidence
System prompt support (separate parameter)	Proof that the intended system prompt was active at inference time	No system prompt attestation

This is not a criticism of Anthropic. It is a statement about the boundary between provider responsibility and deployer responsibility. Anthropic ensures the model is safe to deploy. You must ensure your deployment is governed. SWT3 creates the evidence that bridges this gap.

2. Witnessing Claude Inference

The SWT3 Anthropic adapter wraps the Anthropic client. Every call to messages.create() is witnessed automatically. The response is untouched. Your application code does not change.

Python

from anthropic import Anthropic
from swt3_ai import Witness
from swt3_ai.adapters.anthropic import wrap_anthropic

witness = Witness(
    endpoint="https://your-instance.example.com/api/v1/witness",
    api_key="axm_live_...",
    tenant_id="YOUR_TENANT",
)

client = wrap_anthropic(Anthropic(), witness)
# Or: client = witness.wrap(Anthropic())  # auto-detected

response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    system="You are a compliance analyst.",
    messages=[{"role": "user", "content": "Review this policy document."}],
)
# response is untouched -- witnessing happens in the background

witness.flush()

TypeScript

import Anthropic from '@anthropic-ai/sdk';
import { Witness, wrapAnthropic } from '@tenova/swt3-ai';

const witness = new Witness({
  endpoint: 'https://your-instance.example.com/api/v1/witness',
  apiKey: 'axm_live_...',
  tenantId: 'YOUR_TENANT',
});

const client = wrapAnthropic(new Anthropic(), witness);
// Or: const client = witness.wrap(new Anthropic());  // auto-detected

const response = await client.messages.create({
  model: 'claude-sonnet-4-20250514',
  max_tokens: 1024,
  system: 'You are a compliance analyst.',
  messages: [{ role: 'user', content: 'Review this policy document.' }],
});
// response is untouched -- witnessing happens in the background

await witness.flush();

3. Streaming Witnessing

Claude's streaming API sends response tokens incrementally, which is critical for user-facing applications. The TypeScript adapter handles both streaming patterns -- create({ stream: true }) and the .stream() convenience method -- and assembles the full response for witnessing after the stream completes.

TypeScript (stream: true)

import Anthropic from '@anthropic-ai/sdk';
import { Witness, wrapAnthropic } from '@tenova/swt3-ai';

const witness = new Witness({
  endpoint: 'https://your-instance.example.com/api/v1/witness',
  apiKey: 'axm_live_...',
  tenantId: 'YOUR_TENANT',
});

const client = wrapAnthropic(new Anthropic(), witness);

const stream = await client.messages.create({
  model: 'claude-sonnet-4-20250514',
  max_tokens: 1024,
  system: 'You are a compliance analyst.',
  messages: [{ role: 'user', content: 'Summarize the key risks.' }],
  stream: true,
});

for await (const event of stream) {
  if (event.type === 'content_block_delta' && event.delta.type === 'text_delta') {
    process.stdout.write(event.delta.text);
  }
}

// Anchor minted after stream completes -- full response hash covers all chunks
await witness.flush();

TypeScript (.stream() convenience)

const stream = client.messages.stream({
  model: 'claude-sonnet-4-20250514',
  max_tokens: 1024,
  system: 'You are a compliance analyst.',
  messages: [{ role: 'user', content: 'Summarize the key risks.' }],
});

stream.on('text', (text) => process.stdout.write(text));

const finalMessage = await stream.finalMessage();
// Anchor already queued -- covers the complete assembled response

await witness.flush();

4. System Prompt Drift Detection

Anthropic's API accepts the system prompt as a separate system parameter, distinct from the messages array. This is an important design choice -- the system prompt sets the behavioral boundary for the model and should be treated as a governance artifact, not just another message.

The SWT3 adapter extracts the system parameter and hashes it independently as system_prompt_hash. This creates a continuous record of which system prompt was active at each inference. If your system prompt changes between deployments -- intentionally or accidentally -- the hash changes, and the ledger shows exactly when.

How drift appears in the ledger

Timestamp	Model	system_prompt_hash	Signal
2026-10-18 09:00	claude-sonnet-4-20250514	a3f8c2...	Stable
2026-10-18 14:30	claude-sonnet-4-20250514	a3f8c2...	Stable
2026-10-19 02:15	claude-sonnet-4-20250514	7e1b09...	Drift detected

The system prompt hash changed overnight. Either a deployment pushed a new prompt, or configuration management failed. Either way, the evidence exists.

5. Anthropic RSP Alignment

Anthropic's Responsible Scaling Policy (RSP) defines AI Safety Levels (ASL) and commits to specific safeguards at each level. SWT3 witnessing does not replace RSP -- it creates the deployer-side evidence that RSP's provider-side commitments are being honored in production.

RSP Commitment	What Anthropic Does	What SWT3 Witnesses	Procedure
Capability evaluation	Tests models before release for dangerous capabilities	Records which model version was used at each inference	AI-INF.1
Deployment safeguards	Applies usage restrictions based on ASL classification	Witnesses whether guardrails were active and whether they intervened	AI-GRD.1
Model version tracking	Publishes model version strings (e.g., claude-sonnet-4-20250514)	Anchors the exact version string to each inference decision	AI-INF.1
Safety training (Constitutional AI)	Aligns model behavior via RLHF and Constitutional AI	Detects behavioral drift via response hash patterns over time	AI-DRIFT.1
Incident response	Commits to model recall if dangerous capabilities are discovered	Provides evidence chain for revocation anchors (AI-REV.1)	AI-REV.1
Third-party evaluation	Supports external red-teaming and auditing	Gives auditors independently verifiable evidence of production behavior	AI-AUDIT.1

RSP is the provider's safety commitment. SWT3 is the deployer's operational evidence. Together, they close the accountability loop: Anthropic proves the model is safe to deploy, and SWT3 proves your deployment is governed. For the full crosswalk mapping RSP commitments to SWT3 procedures, see the Anthropic RSP Crosswalk guide.

6. Regulatory Evidence Mapping

Every Claude inference anchor maps to specific regulatory obligations. This is not theoretical -- the anchor is the artifact that satisfies the requirement.

Obligation	Regulation	SWT3 Evidence	Procedure
Deployer must log AI system operation	EU AI Act Art. 26	Per-inference anchor with model ID, prompt hash, response hash, timestamp	AI-INF.1
Automatic logging capabilities	EU AI Act Art. 12	Continuous witnessing -- every call logged with cryptographic fingerprint	AI-INF.1
Map AI risks in deployment context	NIST AI RMF MAP 1.5	Model version, system prompt hash, and usage patterns per deployment	AI-INF.1, AI-INF.2
Measure AI system behavior over time	NIST AI RMF MEASURE 2.2	Drift detection via hash comparison across assessment windows	AI-DRIFT.1
AI management system -- operational planning	ISO 42001 A.6.2.6	Continuous evidence of AI system operation under management controls	AI-INF.1, AI-INF.2
Model risk management -- ongoing monitoring	SR 11-7 / SR 26-2	Per-decision evidence trail with model version pinning and behavioral fingerprints	AI-INF.1
Human oversight measures	EU AI Act Art. 14	Guardrail intervention evidence, system prompt governance records	AI-GRD.1, AI-INF.2

7. Quick Start

Install

# Python
pip install swt3-ai anthropic

# TypeScript
npm install @tenova/swt3-ai @anthropic-ai/sdk

Three Lines to Compliance Evidence

# Python -- three lines added to any Claude integration
from anthropic import Anthropic
from swt3_ai import Witness
from swt3_ai.adapters.anthropic import wrap_anthropic

witness = Witness(endpoint="https://your-instance.example.com/api/v1/witness", api_key="axm_live_...", tenant_id="YOUR_TENANT")
client = wrap_anthropic(Anthropic(), witness)

# Everything below is your existing code, unchanged
response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    system="You are a compliance analyst.",
    messages=[{"role": "user", "content": "What are the key risks in this contract?"}],
)
print(response.content[0].text)
witness.flush()

// TypeScript -- three lines added to any Claude integration
import Anthropic from '@anthropic-ai/sdk';
import { Witness, wrapAnthropic } from '@tenova/swt3-ai';

const witness = new Witness({ endpoint: 'https://your-instance.example.com/api/v1/witness', apiKey: 'axm_live_...', tenantId: 'YOUR_TENANT' });
const client = wrapAnthropic(new Anthropic(), witness);

// Everything below is your existing code, unchanged
const response = await client.messages.create({
  model: 'claude-sonnet-4-20250514',
  max_tokens: 1024,
  system: 'You are a compliance analyst.',
  messages: [{ role: 'user', content: 'What are the key risks in this contract?' }],
});
console.log(response.content[0].text);
await witness.flush();

Three lines added. Zero changes to your application logic. Every Claude inference now has a cryptographic witness anchor that proves what model was used, what was asked, and what was answered -- without storing the content itself.

8. References

• Anthropic RSP Crosswalk -- full mapping of RSP commitments to SWT3 procedures
• Anthropic Zero Trust Crosswalk -- zero trust architecture alignment for Claude deployments
• Anthropic Messages API Documentation
• Anthropic Responsible Scaling Policy
• SWT3 SDK Documentation
• swt3-ai on PyPI
• @tenova/swt3-ai on npm

Other Witnessing Guides

• OpenAI GPT Compliance Witnessing
• Google Vertex AI and Gemini Compliance Witnessing
• LangChain and LangGraph Compliance Witnessing
• DeepSeek GPAI Compliance Witnessing
• Groq Compliance Witnessing