Learned Structure Shows Up Earlier in the Task's Own Basis: A Synchronisation Precursor in Grokking
arXiv preprintHolds activations and the metric fixed and varies only the readout basis: structure appears first in the task's own discrete-log basis in 60/60 transformers on modular multiplication across three primes, and ≥16/20 per prime under untrained-calibrated thresholds. Extends to six finite groups via irreps — first on all six, 4/6 under stricter thresholds. v3 corrects v1's causal claims and a metric artefact, and reports the negatives: memorisation precedes grokking in 179/179 runs, and synchronisation predicts it only weakly (ρ = 0.18).


