Retrieval
Did it find the right records?
Library / 05 Governance, Risk & Human Control
Reviewed Oct 2026Claude Managed Agents + Tracing
A bounded State investigator: use traces to diagnose failures without crossing authority boundaries.
Seen in: State
01 / flow
Investigation order
Current StateOpen ReviewsQuestionsRaw source evidenceSuggested answer + unknownsHuman next step
02 / table
Agent setup
| Boundary | Rule |
|---|---|
| Job | Investigate one unresolved question; suggest an answer |
| Allowed tools | Read, Grep, Glob |
| Disallowed tools | Write, Edit, Bash, Web Search, Web Fetch |
| Source types | State, Reviews, Questions, Notes, Slack, transcripts |
| Authority | Never approve reviews, mutate state or resolve questions |
03 / cards
Read the trace by layer
Context
Was chronology and provenance available?
Interpretation
What did evidence actually establish?
Authority
Was pending silently upgraded to fact?
Final answer
Did uncertainty and next human decision remain explicit?
04 / note
Q-103 failure
Agent found the correct open Review but stated ownership as accepted fact. The revised instruction made it say ownership remained under review, then the same case was rerun.
Read the source hierarchy
| Source state | What may be asserted |
|---|---|
| Current State says training ownership under review | Ownership is not established |
| Priya offers to help | An offer, not an appointment |
| Open Review proposes Priya | A pending proposal, not group approval |
| Later approved update | Only after the authorized transition may accepted ownership change |
Boundary of the exercise: This was a controlled read-only State Question Investigator exercise, not proof of production robustness. Retrieval success is not authority correctness; trace the first synthesis error, preserve chronology and evidence provenance, and monitor whether excessive caution creates review burden.
05 / list
Transferable lessons
- Read-only tools do not prevent unsafe interpretation
- Trace pinpoints the first observable wrong layer
- Retest the same case after changing instructions
- Separate safe uncertainty from excessive review burden
- Never confuse controlled demos with production evidence