Traceability

Traceability means a person can follow an AI system's path from a real-world event to a model output and, when relevant, to a human or automated action. It answers: what happened, where did the information come from, which version made the decision, and who touched it?

Traceability is not the same as explaining every parameter inside a model. It is an operational record. A traceable loan denial, content moderation decision, or medical triage recommendation can be reconstructed without relying on memory or guesses.

Replay a Loan Decision

A denied applicant disputes the decision. Step through the pipeline that produced it, and decide which stages your system actually captures. A trace only helps if a reviewer can rebuild what happened.

Stage detail

Input snapshot

The loan application as submitted: income, requested amount, and applicant-entered fields.

Reviewer's checklist

Input snapshot(needed to replay)
Preprocessing(needed to replay)
Model version(needed to replay)
Retrieved context(needed to replay)
Policy threshold(needed to replay)
Model output(needed to replay)
Human override
Final action(needed to replay)

A reviewer could replay this decision.

Traceability is not about logging everything — it's about preserving the specific links a reviewer needs to reconstruct a decision without guessing.

A useful trace usually includes the input snapshot, preprocessing steps, model or prompt version, retrieved context, thresholds, policy rules, output, override, and final action. The goal is to preserve enough context to inspect the decision later without storing unnecessary sensitive data.

Good traces are designed before deployment. If teams wait until a dispute or incident, the missing link is often exactly the one they need.

Trace the System Boundary

Many AI failures happen outside the model call: a stale feature, a missing document, a changed threshold, or an unrecorded human override. Trace the full workflow, not only the model response.

Checkpoint

A team stores the final model score but not the input version or policy threshold. What is the traceability gap?