Inspector can combine research, files, data providers, spreadsheets, presentations and connected tools. The operator must separate three things when reviewing the run:
- Claim: what Inspector says is true.
- Evidence: the source that supports the claim.
- Action: what Inspector proposes or performs because of the claim.
Review claims before recommendations
For every consequential claim, ask:
- Is there a source?
- Does the source support this exact claim?
- Is the source current enough for the decision?
- Is Inspector reporting a fact, an inference or a recommendation?
- Did another domain return contradictory evidence?
Contradictions should remain visible until the source definitions, time periods and populations are aligned. Do not average incompatible figures into a clean number.
Read and write boundaries
Access is not approval. A connected credential may make an action technically possible without making it appropriate.
| Action class | Examples | Safe operator response |
|---|---|---|
| Read | Search, inspect a file, query a data source | Confirm scope and sensitive fields. |
| Prepare | Draft a file, sheet, plan or message | Review the artefact before use. |
| Write | Update a system, create a workflow, change a record | Confirm target, fields and rollback. |
| Publish | Deploy, send, expose or trigger externally | Require explicit final approval. |
For write and publish actions, verify the exact target, expected mutation, owner, rollback and evidence that the action completed.
A defensible synthesis
A strong final answer separates:
- sourced findings,
- unresolved contradictions,
- assumptions and uncertainty,
- recommendations,
- proposed actions requiring approval,
- next checks that would change the decision.
The answer is not complete merely because every sub-agent reported. It is complete when the synthesis addresses the approved decision and the operator can trace the important conclusions back to evidence.