Skip to main content
Reference Specification — Draft v1.0·Published 2026-05-15

The AI Forensics Audit Trail Specification

An open reference standard for tamper-evident, cryptographically-signed audit trails of autonomous AI agents. Profiles OCSF, OpenTelemetry GenAI, MITRE ATLAS, SPIFFE, and NIST AI RMF as a single coherent schema for forensic reconstruction — verifiable offline by auditors with no vendor dependency.

The Spec at a Glance

Four pillars, one coherent schema

The spec organizes forensic primitives into four pillars. Each pillar maps to a specific question an auditor will ask after an incident.

PillarThe auditor's questionSpec capability
IdentityWho is this agent?Per-agent API keys, lifecycle management, scoped credentials.
PolicyWhat is it allowed to do?Fail-closed gateway, deny-by-default policy evaluation.
ComplianceCan we prove rules were followed?DSSE-signed session attestations, HMAC-verifiable audit logs, automated compliance assessments.
ForensicsWhat happened, provably?Hash-chained logs, incident replay, export, offline cryptographic verification.

§9.1 Capability Matrix

Per-provider logprob exposure

Token-level log-probabilities are the strongest forensic signal — they let a verifier replay sampling decisions deterministically. Not every provider exposes them. This matrix tracks the state of public APIs as of 2026-05.

ProviderLogprobsDetailForensic sufficiency
OpenAIYes`logprobs` parameter returns top-N log-probabilities per token. Up to 5 alternatives via `top_logprobs`.Yes
AnthropicNo (as of 2026-05)No public logprob exposure on the Messages API. Forensic attestations rely on response provenance (model_id, system fingerprint, request_id) instead.Partial
Google GeminiYes`logprobs` and `responseLogprobs` available on the GenerateContent API. Up to 5 alternatives.Yes
AWS BedrockVaries by modelClaude on Bedrock: no logprobs (parity with Anthropic direct). Other models: provider-dependent.Partial
Self-hosted (vLLM, TGI, llama.cpp)YesFull logprob distribution exposed. Strongest forensic signal — verifier can replay token-by-token sampling decisions.Full

See the full spec §9.1 for the verifier algorithm + fallback rules when logprobs are unavailable.

§9 Open Questions

What v1.0 doesn't solve yet

A specification that hides its uncertainty is not a standard — it's a marketing document. These are the questions v1.0 leaves open, with v1.1 and v1.2 commitments where applicable.

§9.1

Per-provider logprob exposure

Anthropic and Bedrock-Claude do not expose token logprobs. Forensic attestations rely on response provenance (model_id, system fingerprint) — but this is weaker than logprob-backed replay. We track upstream changes here.

§9.2

Multimodal context capture

Image, audio, and video inputs are out of scope for v1.0. v1.1 plan in progress — likely capture content-hash + MIME + size rather than raw bytes, with optional vendor-specific perceptual hashes.

§9.3

Cross-organization federation

Agent identity assertions across trust domains depends on IETF AIP draft maturity. v1.0 assumes single-org trust domain; federation profile deferred to v1.2.

§9.4

Vendor-hosted agent attestation

Agents running entirely inside vendor-hosted runtimes (Perplexity hosted agents, OpenAI Custom GPTs, Anthropic Claude Projects, Manus, Devin) cannot be intercepted by the gateway-proxy pattern — calls originate inside vendor infrastructure with no operator-controlled injection point. v1.0 supports operator-assertion only (declarative registration, manual artifact logging). v1.1 will introduce Level 0 declarative conformance + a browser-side intercept profile; v1.2 tracks vendor-API attestation hooks. The most strategically important unanswered piece of v1.0.

Help us land v1.1

The spec is open and seeking community review. File issues on GitHub, comment on the OCSF discussion, or join the AI Identity design partner cohort to shape v1.1 against real production deployments.

OCSF discussion issue: filing pending · Spec license: CC-BY-4.0 (content) + Apache-2.0 (reference impl)