How to Distinguish Actual Results, Inferences, and Next-Step Suggestions?
Check answers with evidence to avoid treating plans as already completed.
First Find Evidence for Each Conclusion
When reading an answer, you can categorize the content into three types: facts or actual outputs from tools you provided, explanations based on those facts, and suggestions that have not yet been executed. For example, a command example is just a next-step method; only after the corresponding execution output appears can you discuss what it returned. This kind of check helps you quickly identify overly confident statements.
Have the Answer Clearly Mark the Status
You can request, "Please separately list observed facts, inferred causes, and items pending verification, and provide evidence for each." For important results, further ask for corresponding logs, files, or source locations. If a check fails, keep the reason for failure and do not omit it from the summary as a success.
- Select the conclusion that most impacts the decision.
- Identify the original input or tool result supporting it.
- For judgments lacking support, require an explanation of uncertainty.
Narrow Next Steps to Verifiable Actions
If multiple possible causes exist, ask for the minimal check that can distinguish among them, rather than making large-scale changes simultaneously. After completing the check, bring the actual results back to the same task for continued judgment. Tool output does not necessarily mean the entire goal is completed; you also need to see whether the output answers the question and whether the current environment and test conditions are applicable. When contradictory materials appear, first preserve the differences before explaining possible reasons.