Trainable Linear Bottlenecks Add Forgetting Under a Shared Head: Five Pre-Registered Experiments on Permuted-MNIST

On Permuted-MNIST with a shared head, a trainable linear bottleneck between two hiddenlayers adds forgetting of +0.11 [0.09,0.13] at width k = 8 and +0.15 [0.13,0.16] at k = 64relative to the same bottleneck held frozen, over 50 seeds in five identically configured sets(descriptive aggregate; set means 0.085–0.132 at k = 8, heterogeneous), and +0.085 and +0.155in a pre-registered replication, with plasticity not reduced (higher by at most 0.004). Withfrozen per-task readers selected at test by the task identity (a confound), the excess is absentat k = 64 and absent or reversed at k = 8 (E3: −0.033 [−0.053,−0.013]). Frozen-bottleneckcontrols with a threefold upstream step and a norm-matched gain reach the excess at k = 8 withthe bottleneck output amplified 1.79-fold over the frozen one (trainable: 1.46; rule undecided,not equivalence) and 0.58 of it at k = 64, at 1.25 against 1.50 (pre-registered failure). A largereffective step is not excluded; the residual is not attributed. A lemma: with the bottleneckfrozen, the layer feeding it writes only in the bottleneck’s row space. The work started fromstructural opacity, a learner reading its activations only through a compressed linear projection.Because structural opacity is the same in every pair of conditions we compare, it could not betested as a cause; the four hypotheses derived from it (stability, rotation, readers, movement)were tested and not supported, the archive remained inconclusive, and the fifth experiment isthe replication above. Of the motivating hypotheses, two are clean pre-registered negatives, onepartial and one by construction; every other failed rule is reported. All results come from onesmall architecture at one scale. Code, configurations, pre-registrations and FROZEN result files are in the attached archive (see README.md). This record continues doi:10.5281/zenodo.21613205.