agentrust-trace 0.2.0

TRACE v0.1 — hardware-attested governance records for AI agents

TRACE: Trust Runtime Attestation and Compliance Evidence

Developer Preview. Launching at Confidential Computing Summit, June 23 2026. May have breaking changes before v1.0.

Specification  |  Schema  |  Examples  |  Registry  |  Reference Impl

An open specification for hardware-attested AI agent governance records. TRACE defines the format, anchoring protocol, and verification rules for cryptographically provable evidence that an AI agent ran under a specific policy, in a verified hardware environment, on classified data, invoking identified tools, bound into a single signed artifact rooted in silicon attestation.

What a TRACE Trust Record Is

{
  "eat_profile": "tag:agentrust.io,2026:trace-v0.1",
  "iat": 1750676142,
  "subject": "spiffe://trust.example.org/agent/payments-processor/prod",  // or "did:mesh:spiffe://..."
  "model": {
    "provider": "anthropic",
    "model_id": "claude-sonnet-4-6",
    "version": "20251001",
    "weights_digest": "sha256:a3f8d2c1..."
  },
  "runtime": {
    "platform": "amd-sev-snp",
    "measurement": "sha384:c9e4b1d2e3f4...",
    "rim_uri": "https://kdsintf.amd.com/vcek/v1/..."
  },
  "policy": {
    "bundle_hash": "sha256:b2c3d4e5...",
    "enforcement_mode": "enforce",
    "version": "1.2.0"
  },
  "data_class": "confidential",
  "tool_transcript": {
    "hash": "sha256:d4e5f6a7...",
    "call_count": 3
  },
  "build_provenance": {
    "slsa_level": 2,
    "builder": "https://github.com/slsa-framework/slsa-github-generator",
    "digest": "sha256:e5f6a7b8..."
  },
  "appraisal": {
    "status": "affirming",
    "verifier": "https://trust-authority.example.org",
    "policy_ref": "https://trust-authority.example.org/policy/agent-v1"
  },
  "transparency": "https://registry.agentrust.io/claim/trace-2026-06-23T09:15:42Z-f2a8d1",
  "cnf": {
    "jwk": {"kty": "EC", "crv": "P-256", "x": "MEkwEw...", "y": "..."}
  }
}

The record is a single EAT envelope (RFC 9711). Each field is independently verifiable. No callback to the issuer is required.

Supported platforms

Platform	v0.1	v0.2 (planned)	Notes
intel-tdx	Yes	Yes
amd-sev-snp	Yes	Yes
nvidia-h100	Yes	Yes
gpu-cc	No	Planned	Generic GPU confidential compute
opaque	Yes	Yes	Explicit opt-in; contact maintainers

Specification

• spec/trace-v0.1.md -- full specification
• schema/trace-claim.json -- JSON Schema
• examples/ -- example Trust Records for Intel TDX, AMD SEV-SNP, and NVIDIA H100

Standards composition

TRACE profiles existing standards rather than replacing them:

Primitive	Role in TRACE
RATS / EAT (RFC 9711)	Wire envelope and claim model
SLSA Provenance v1.0	Build-time provenance (build_provenance)
SPIFFE SVID / DID URI	Workload identity (subject)
SCITT	Append-only transparency anchoring (transparency)
EAR (draft-ietf-rats-ar4si)	Verifier appraisal output (appraisal)
MCP / A2A	Agent tool-call transcript surface (tool_transcript)
AIBOM (SPDX 3.0 / CycloneDX 1.7)	Model component inventory (model)

Reference implementation

agentrust-io/cmcp -- Confidential MCP Runtime. Hardware-attested policy enforcement at the MCP tool-call boundary on Intel TDX, AMD SEV-SNP, and NVIDIA H100/Blackwell.

Registry

A public append-only Merkle registry of TRACE Trust Record anchors: agentrust-io/trace-registry.

Install

pip install agentrust-trace

from agentrust_trace import TrustRecord, sign_record, load_signing_key

key = load_signing_key()  # reads TRACE_PRIVATE_KEY_PEM or generates ephemeral
record = sign_record(payload, key)
TrustRecord.model_validate(record)  # validate before writing

Status

v0.2. subject now accepts DID URIs (did:) in addition to SPIFFE SVIDs. slsa_level: 0 supported for software-only deployments. Published at Confidential Computing Summit, San Francisco, June 23 2026. Targeting submission to the Agentic AI Foundation (AAIF) under the Linux Foundation.

Community

Questions, feedback, integration help: Discord.

License

Creative Commons Attribution 4.0 International (CC BY 4.0)