The Readout Chooses the Winner: Evaluation Protocols Decide Fairness Conclusions in Continual Learning

A continual-learning result is a claim about models, but it is produced by a protocol. Which head reads out the encoder, what utility a model must show before its fairness is reported, and how many seeds were run are all choices, and we treat that protocol as the object of study. Our claim is that a conclusion is only as strong as the layer it is most sensitive to, so the layers need stress-testing one at a time. We test this on a nine-method comparison in fairness-aware continual learning. The choice of readout decides which method appears best. On identical frozen encoders, a head refit on retained data ranks FAIR-GEM first among continual learners at 0.732 balanced accuracy, while a training-free nearest-class-mean readout ranks it fourth and puts DER++ first at 0.699. Four defensible readouts give three different answers, and none of them supports a ranking. Head collapse also manufactures apparent fairness, because four of the nine methods degenerate onto the five newest of 25 classes, and a classifier whose predictions barely vary has little measurable rate disparity, so their equalized-odds gaps ($\leq 0.026$, against 0.146 for the best working model) beat every functioning model's. The broken models rank first in a table whose every cell is correctly computed. Refitting the head returns the disparity, which was concealed rather than absent. Four further layers fail the same test on the same comparison. A surrogate stops tracking its metric, a gradient-space guarantee holds on a minority of applied updates, a cost moved with the solver backend at fixed configuration, and a three-to-five seed budget leaves an exact paired permutation test unable to reach $p<0.05$ however large the effect. The sensitivity is a property of the chain rather than of one careless link. Each failure carries a procedural remedy, and we release the protocol code.