Can someone else reconstruct the decision?
We propose comparing chat transcripts, ordinary application logs and structured evidence bundles across synthetic workflows. Independent reviewers will answer predefined questions about sources, permissions, approvals and outcomes. We will measure reconstruction accuracy, missing links, review time and the amount of sensitive data retained. Evidence quality alone does not establish regulatory compliance.
Scope and question
A useful record should let someone who was not present reconstruct what happened: which information was used, which rules applied, who approved the action and what changed. That becomes harder when documents, policies or model versions move on.
The proposed study builds on synthetic workflows from the agent-authority simulator. It compares transcript-only records, application event logs and evidence bundles that connect sources, configuration, policy decisions, approvals and final state.
Research approach
Freeze the case and ground truth.
Use synthetic record updates, document classifications and draft recommendations. Record the intended action and outcome before the reviewer sees the artifacts.
Introduce realistic missing context.
Include a changed source, missing event, stale approval, replay and corrected decision. Make uncertainty visible when provider internals or version identifiers are unavailable.
Ask someone else to reconstruct it.
Give a reviewer the evidence without the answer key and ask six predefined questions. Compare the record formats against the same cases.
What we will measure
- Primary: correctly answered reconstruction questions.
- Missing links, false attribution and time required for review.
- Storage requirements and the amount of sensitive information retained.
What this can and cannot tell us
The study concerns evidence quality, not a legal opinion or a guarantee that a regulator will accept a record.
A hash alone does not establish authenticity against someone able to rewrite both the record and its hash. Integrity claims need a defined threat model.
Private model chain-of-thought is not a substitute for externally verifiable evidence. This proposal has no completed reviewer results.
Next steps
Define the six questions a process owner must answer after an incident. Once the simulator exists, create a minimal evidence schema, replay fixtures and a sample reviewer report.
Discuss the research