Clean and compare without rewriting the evidence

Cleaning can involve unit conversion, quality filtering, duplicate handling and alignment. Each action should answer a defined problem. Preserve the original input and record transformations so another person can distinguish what was observed from what the analysis changed.

Do not remove a value merely because it weakens an expected result. A documented quality rule can justify exclusion; a surprising valid observation may instead be the most interesting part of the investigation. Count excluded records and retain their reasons.

In interval A, suppose target signal is 100 and reference signal — open definition is 50 in compatible units. Their ratio is 2. In interval B, signals are 80 and 40, so the ratio remains 2 even though both measured signals fell by 20%. This comparison does not show a relative target dimming.

Now consider a separate interval C with target 90 and reference 50. The ratio is 1.8, a 10% decrease from the baseline ratio of 2. These invented numbers demonstrate the calculation. They do not establish the physical cause or represent a real telescope detection.

A reference changes the interpretation

A target reading should be matched to an appropriate reference observation rather than whichever row happens to share its position in a file. Missing rows can shift indexing and create false pairs. Matching identifiers or timestamps makes the intended relationship explicit.

Try a justified alternative quality threshold and report whether the main result changes materially. Sensitivity to a reasonable analysis choice can reveal that the conclusion is fragile. This does not mean choosing the threshold that gives the preferred answer; it means documenting how the result depends on a decision that deserves scrutiny.