Stardive Audit Layer

Independent audit, provenance, and governance for AI agents.

Stardive Audit is a vendor-neutral layer that turns observed AI-agent activity into structured evidence for inspection, verification, and human review. It connects agent execution to the records behind it: prompts, tool calls, permissions, file changes, sub-agent activity, and generated artifacts, subject to what each integration exposes.

The platform works alongside existing agent environments, including Claude Code, Codex CLI, and the Stardive demo agent. Agents retain their own interfaces and execution workflows while Stardive provides a common audit contract and review experience.

Problem

An agent's final response gives reviewers limited visibility into how work happened. Teams need to trace outputs to supporting evidence, inspect policy findings, identify missing records, and document how issues were resolved. Different agents expose different evidence, so those limitations must remain visible throughout the review.

Core capabilities

• Activity capture: Collect events through native lifecycle hooks and producer adapters, with an OpenTelemetry adapter for instrumented workloads.
• Execution visibility: Organize captured events into trace graphs linking turns, tool calls, sub-agents, and artifacts.
• Artifact provenance: Connect outputs and file changes to recorded actions, content hashes, and supporting evidence.
• Integrity verification: Use append-only event records, server-side hash chaining, and signed evidence exports to support checks for tampering within the recorded evidence.
• Policy and outcome review: Evaluate traces against versioned rulebooks and check claims against available evidence, keeping automated assessments distinct from human judgments.
• Human governance: Record auditor verdicts, comments, approvals, and corrective action plans (CAPs), including remediation evidence and reviewer verification.
• Evidence-based assurance: Compute assurance from observed evidence and expose gaps, mismatches, unknown outcomes, and capture limitations.
• Consistent review: Present the same canonical trace and audit interpretation across the audit console and embedded Stardive views.

How it works

Agent activity
  → Capture and normalization
  → Redaction and authenticated ingestion
  → Canonical event store and integrity chain
  → Execution graph, provenance, and policy checks
  → Human review and corrective action
  → Evidence export and verification

Intended users

Stardive Audit serves engineering teams investigating agent behavior, auditors and risk teams reviewing evidence and policy findings, and managers tracking remediation and accountability.

Assurance boundary

Each trace carries the assurance its evidence supports. Missing events, incomplete tool outcomes, or absent independent witnesses limit what can be verified. Human approval does not increase capture assurance. Replay is limited to explicitly supported deterministic operations; general agent sessions carry no blanket reproducibility guarantee.

Project objective

Make agent-assisted work inspectable from execution through remediation, giving reviewers a clear record of what was observed, what evidence supports the result, what remains unknown, and who reviewed the outcome.

This description covers product purpose and scope; it does not certify release readiness or a particular trace's assurance level.