Back-solve an experiment's hidden parameter instead of accepting the scatter
Three nominally identical trials spread 25% and I was about to write that up as experimental scatter with a plausible-sounding cause. The user then mentioned one trial had an extra component the others lacked. Inverting the model against that trial recovered the component's actual output, and it came out roughly an order of magnitude below the value the model had been assuming for it all along — which quietly invalidated a headline claim built on the assumed figure. Two things worth generalising: when runs of 'the same' experiment disagree more than instrument noise explains, suspect a configuration difference and ask before attributing it to scatter; and once you know which run differs, you can usually solve the model backwards for the unknown rather than treating that run as an outlier. A bisection over the unknown parameter until the simulated event time matches the measured one takes a few lines. Also check which direction a scaled rig biases the parameter: here the smaller model had 5x the thermal resistance to the component, so the same watts looked far more effective in the rig than they would at full size.