Auditable

Regulated teams are putting AI agents into work that carries supervisory, clinical, financial, or claims risk. A bank uses one to clear trade-surveillance alerts. A health plan uses one to draft an adverse determination. A claims desk uses one to decide whether a file needs human review. Most calls pass through because the evidence looks ordinary.

Then one call is wrong. It is not a jailbreak or an obvious outage. The agent followed policy, called the right tools, cited the file, and returned a result that looked internally consistent. Months later the examiner asks why the action was taken. The vector store has changed, the model has been redeployed twice, and the state that mattered at decision time is no longer live.

auditable is an open-source SDK for agent decisions that need a record. It turns a run into one typed graph: the plan, the tools, the data dependencies, the committed action, and the recovery rail.

Before deploy, the graph lints a declared plan for structural risk: a missing approval edge, evidence collected after the action, a policy gate placed after commit. During the run, it captures the dependency state a decision relied on, replays that decision under live state, and reverses the committed action when it no longer holds. After the run, the same graph ranks the trace and names the load-bearing step.

This is the consolidation layer: a system of record for consequential agent decisions. It replaces a pile of disconnected logs, evals, and guardrails with one representation that engineers, operators, and reviewers can all read.

The evidence is public. In GRADE, across six public agent corpora, the dependency layer predicts which runs fail at ROC-AUC 0.805, where run length carries no signal. The execution layer localizes the faulting step at Top-3 0.614.

PyOD is the credibility base for this work. Yue Zhao created it as an open-source anomaly detection library that production teams could adopt, and its role here is a record: reproducible tools, practical APIs, and evidence quality at scale.

Auditable applies the same engineering posture to agent decisions. The object changes from an anomaly score to a consequential action, but the standard stays the same. The system should be inspectable, reproducible, and clear enough to support a later review.

That work builds on a decade of anomaly detection and trustworthy machine learning, including TrustLLM. The common thread is auditability as an engineering constraint, from model behavior to the decisions an agent commits.