Do not place a month-to-date total beside a completed month's total without marking the coverage difference. Choose a common cutoff or leave the comparison explicitly partial.
Successive rolling windows can share most of their observations. Check the overlap before treating each new window as another independent period of evidence.
Keep the observation period separate from the publication date. A report released today can describe conditions from months ago. Decide whether that gap matters for the action being considered.
A balance at month-end and an amount accumulated during the month answer different questions. Mark whether a figure is a point-in-time stock or a period flow before combining it with another number.
Two reports labeled Q1 may cover different months. Write the start and end dates beside each figure before comparing them. A shared period name is only a starting point.
Give a material caveat a place in the handoff. If it only survives in a reviewer's memory, the next person has no reliable way to know when the conclusion stops applying.
Before using an ellipsis in a quotation, read the omitted passage again. If it changes the speaker's condition, comparison, or degree of certainty, shortening the quote changes the claim.
When shortening a finding for a decision-maker, keep the condition that would change their choice. Cut background first; do not spend the word limit by deleting the exception.
When a method points to an appendix, add that appendix to the reading task. Keep the conclusion provisional until the referenced definition or calculation has actually been inspected.
A footnote may qualify one cell, one column, or the whole table. Trace the marker before carrying a number into a summary. The qualification needs to travel with the value it changes.
An unresolved item needs an owner and a next review point. Carrying it into another week should preserve the question, not quietly convert it into a completed task.
If someone has not agreed to public use of their task or words, there is no case study to publish. A general lesson can be shared without taking their story.
Preserve the failing input and the relevant version before changing the system. Otherwise the attempted repair can remove the evidence needed to understand the failure.
A tool should receive the access needed for its current task, not every permission the account can grant. Keep credentials out of prompts and ordinary logs.
Use synthetic or redacted examples when investigating a failure that does not require real personal data. Keep the debugging question narrower than the original dataset.