Story 02 · AI Assurance Investigator
Investigate whether evidence influenced AI agent’s decision
Test one approved evidence change in an isolated replay to measure whether a risk score materially changed the agent’s outcome.
The situation
A claims agent recommends reviewing a claim after consulting a risk score. An investigator needs to know whether that score influenced the outcome, rather than merely appearing in the execution record.
The question
When one permitted evidence input changes, does the measured outcome change under the supported controlled-replay contract?
The governance workflow
From question to governed outcome
- 01
Select completed, auditable evidence
Choose a completed execution and a tool result whose runtime provides the durable evidence references and isolated replay capability required for an audit.
- 02
Define and approve one intervention
Create a versioned intervention policy that narrowly permits a replace, nullify, or bounded perturbation of the selected evidence.
- 03
Verify eligibility before work starts
BHANUJ - AIGP (AI Governance Platform) checks the source evidence, policy version, runtime capability, and safe replay preconditions before it queues a controlled replay.
- 04
Run the isolated controlled replay
The runtime produces a new counterfactual execution; the historical source execution and its evidence remain immutable.
- 05
Measure and classify the difference
Review original evidence, intervention lineage, replay outcome, measured evidence influence, and the supported causal classification.
In Studio
The screens behind the story
The audit begins with a completed execution and an auditable tool result.
The policy permits a bounded evidence change before any controlled replay is queued.
The result connects the original evidence, the controlled replay, and the measured outcome difference.
Verify the outcome yourself
Run one governed intervention and inspect the measured difference
The causal-audit tutorial uses the local fraud scenario to select evidence, activate a bounded policy, check eligibility, and inspect the new audit record.
Expected result
Studio shows the original evidence, the approved policy version, a distinct controlled replay, its counterfactual outcome, and a causal classification only when the required replay and lineage contract is complete.
The outcome
The investigation produces bounded evidence about outcome sensitivity to one governed intervention, rather than a claim about hidden reasoning or universal causation.
OSS capabilities used