Technical Guide

Evidence that an AI answer was grounded at the time it was given

Sufficient grounding evidence is a record, written at generation time, that links the final response to the specific source versions, retrieved passages, tool results, policy decisions, model and prompt versions, and agent identity that produced it. The record must be integrity-protected and detailed enough to replay or inspect later. It shows what the system was given and allowed to do. It does not prove what the model relied on internally.

Grounding is a record-keeping problem

Evidence that an answer was grounded is a record written at generation time. It links the final response to the specific source versions, retrieved passages, tool results, policy decisions, model and prompt versions, and agent identity that produced it. The question is less about what the model "knew" and more about what the system captured when it answered.

Artifacts to capture at runtime

A sufficient record includes the following artifacts:

  • Specific source versions, including document identifiers and retrieval timestamps
  • Retrieved passages and the queries that returned them
  • Tool inputs and outputs
  • Policy decisions, permissions, and outcomes
  • Model, prompt, and index versions in effect
  • Agent identity
  • Integrity protection for the record itself, such as append-only or verifiable storage

Components of a grounding evidence record

The artifacts above group into five components.

ComponentWhat it contains
Source provenanceDocument identifiers, versions, and retrieval timestamps.
Retrieval and tool recordsQueries, returned passages, tool inputs and outputs.
Governance decisionsAgent identity, permissions, and policy outcomes.
Configuration stateModel, prompt, and index versions in effect.
Integrity protectionAppend-only or verifiable storage for the record.

Grounded versus governed

Two kinds of evidence sit in the same record. Grounding evidence covers what the answer was based on: source provenance, retrieved passages, and tool results. Governance evidence covers what the system was allowed to do: agent identity, permissions, and policy outcomes.

A sufficient record carries both, because it needs to show what the system was given and allowed to do when it produced the answer.

Tamper evidence and point-in-time reproducibility

The record must be integrity-protected and detailed enough to replay or inspect later. Integrity protection, such as append-only or verifiable storage, addresses the first requirement. Configuration state addresses the second: recording the model, prompt, and index versions in effect lets a reviewer inspect the answer against the system as it stood when the answer was given.

What this evidence cannot prove

A runtime record shows what the system was given and allowed to do. It does not prove what the model relied on internally.

Assess whether your runtime records would hold up

Talk through how runtime governance and audit logging can produce grounding and governance evidence for your AI agents.

Talk to an Expert