Back to guides

Guide · Safety

AI Decision Traces and Citations for Audit

"The model said so" is not an audit trail. APRF Explainability expects reconstructable traces—citations, tool calls, and policy decisions—tied to versions you can roll back.

Ungrounded answers fail Explainability—cite sources, log policy hits, and keep reconstructable traces. Primary control: Explainability

Explainability that ops can use

Production explainability means a reviewer can reconstruct context, tools, and policies behind an answer—not reverse-engineer every weight. An internal assistant claimed a discount "per policy"; finance only had the final string. After citation objects and a decision trace (retrieval → tools → policy → answer), disputes closed in minutes and bad citations became eval cases.

Trace schema that survives an audit

Log `request_id`, `model_id`, `prompt_version`, source/segment IDs, tool calls, and policy/refusal outcomes. Require citations for grounded claims—or mark ungrounded explicitly. Record which safety/AuthZ rule fired without dumping secrets. Retain by risk class with access control. Show citations and a trace summary in the support UI, not only chat text. Fail evals when citations are missing or fabricated.

APRF Explainability is evidence, not theater

Pair with Data Governance for retention and Incident Readiness when a bad answer becomes a SEV. Traces are how you tell retrieval error from invention.

Assess against APRF Core

Run the Core Profile quiz — gated pass/fail blockers for AI production readiness, not a vanity score.

Frequently asked questions

Is this the same as model interpretability research?
No. APRF Explainability here is auditability of the serving path—context, tools, and policy—not neuron-level explanation.
Must every answer cite sources?
For RAG and policy-bound domains, yes or mark ungrounded. Creative or chit-chat flows can use a lighter schema.
How does this help incidents?
Traces tell you whether to roll back prompt, corpus, model, or tool allowlist—see Incident Readiness playbooks.