When an AI agent takes a wrong action in production, this engine reconstructs what happened, attributes it to a specific version change, measures the blast radius, and exports an audit-ready incident report.
Problem
As AI systems move from advisory roles into production where they take real actions, a new class of operational risk emerges. When these systems fail, the failure is often semantic rather than technical: an AI support agent refunds the wrong account, an agent exposes another user's data, an automated decision violates policy — all while every API returns 200 OK and infrastructure monitoring stays green. Teams then struggle to reconstruct what happened, attribute it to a specific change across prompts, models, tools and policies, and prove control to auditors, regulators, or customers. Existing tooling was not designed to treat AI behaviour itself as a first-class operational incident.
Overview
A forensic engine that treats AI behaviour as a first-class operational concern. Given an append-only record of what an AI system did, it reconstructs the failing run into a timeline and execution graph, correlates the failure with the exact version that changed since the last known-good run, computes the semantic blast radius across affected runs and users, and assembles an audit-ready incident report. It is the open-source core — reconstruction, root-cause, blast-radius and reporting — of a broader AI incident response platform, shipped as a zero-dependency Python library and CLI.
How It Works (Approach)
Decision-relevant telemetry for each run — the rendered prompt, model and parameters, tool calls, policy checks, and the action taken — is captured as typed, append-only events tied to a version bundle (the exact prompt, model, policy and agent-config hashes in force). The reconstruction engine orders these into a timeline and execution graph and identifies the failure point, including semantic failures that return no error. A version diff compares the failing run's bundle against the most recent known-good run to surface precisely what changed. A blast-radius pass finds every run that executed under the faulty version, the distinct users affected, the exposure window, and the downstream actions needing remediation. Findings are rendered as Markdown or JSON for legal and compliance export.
Impact / Value
Engineers move from speculation ('we think it was the prompt') to evidence-based root cause in minutes, with full execution visibility and far less manual documentation. Engineering leadership gains demonstrable control over autonomous AI systems, and compliance teams get consistent, audit-ready incident records aligned with emerging requirements such as the EU AI Act, SOC 2 and ISO 27001. The result shifts organisations from reactive damage control toward the operational confidence needed to deploy AI safely rather than disabling it to reduce risk.
Key Features
- Reconstruction of a failing run into an ordered timeline and execution graph
- Detection of semantic failures that return no API error
- Root-cause version diff against the last known-good run (prompt, model, policy, agent config)
- Semantic blast-radius analysis: affected runs, users, exposure window, remediation list
- Audit-ready incident reports exportable as Markdown or JSON
- Zero-dependency Python library plus a CLI; faithful to a documented data model