close-the-gap-to-the-published-numberlisted

# A disagreement with the published number is a bug report addressed to you You reran the study and got 0.140 V where the paper got 0.0117 V. You have two moves available. One is to write the discrepancy down carefully, attribute it to a difference in setup, and move on to the next section. The other is to treat the gap as evidence that something in your pipeline is wrong, and spend budget until it closes or until you can name the specific thing that makes it irreducible. The first move feels like honesty and reads as one. It is also, almost always, wrong on the facts: a twelvefold gap against a published measurement is far more often a bug than a boundary. And it is scored as what it is — an analysis whose methodology is defensible and whose numbers are not. ## The rule **A quantitative disagreement with the source is a defect with an owner until an experiment says otherwise.** Not a limitation, not a caveat, not a scope note. The owner is this run. Order of magnitude matters. Treat these differently: | gap | reading | |---|---| | within the source's own stated uncertainty | agreement; say so and move on | | a factor of 1.5-3, or a shifted intercept | a parameter, a normalisation, a split, a unit | | a factor of ten, an inverted sign, an inverted trend | a bug; nothing else does this | | your estimator collapses to a constant or to zero | a bug; the model learnt nothing | The last two rows are not results. A latent charge that collapses to 0.0002 e, a curve that sits