OpenAI

OpenAI Audit Logging & EU AI Act Evidence

Article 12 evidence at the model boundary

Wrap Chat Completions, Responses, and tool-enabled calls so every inference produces tamper-evident ai_act events — without copying prompts into a SIEM.

Where Sigigo sits

Sigigo wraps your existing OpenAI client at the inference boundary. When your app calls the API, the wrapper emits canonical evidence events to Sigigo's ingest pipeline — model metadata, timing, tool calls, and cryptographic commitments over sensitive payloads.

What you can prove

Auditors and regulators can verify a complete inference chain:

  • Which model and version ran
  • That inputs and outputs existed at decision time (commitment mode)
  • Tool calls and downstream actions tied to the same session
  • Human overrides when wired through workflow hooks

Private Commitment Mode (EU default)

Prompts and completions stay in your vault. Sigigo witnesses signed commitments only — unlinkable without your held nonce and payload. Disclosure stays under your control for audit or legal review.

Integration shape

Thin wrapper over the TypeScript OpenAI SDK — shared Sigigo client, session context, and idempotency keys so stacked layers do not duplicate events.

  • wrapOpenAI(client, sigigo) — drop-in client proxy
  • EU compliance preset: commitment mode + public anchor off
  • Same event schema whether you emit manually or via wrapper

Evidence events

Canonical event types emitted through this integration path:

  • ai_act.inference_started / inference_completed
  • ai_act.input_received (Private Commitment Mode)
  • ai_act.output_generated
  • ai_act.action_executed (tool calls)

Related

Integration availability may vary by tenant, region, and deployment model. This page describes product architecture — not a guarantee of feature availability. See compliance disclaimers and contact us for early access.