Accuracy Saturates Before Reality-Channel Use Stabilizes: Interventional Tests of Mechanism Non-Stationarity in Associative-Recall Transformers

Task accuracy is the standard test of whether a sequence model uses memory correctly. On a controlled associative-recall task with an explicit present/valid channel, we show it is insufficient: at matched, often perfect, accuracy, trained seeds occupy a continuous reality-channel-use axis that is bimodal, splitting into a mode that uses the reality channel (genuine) and a reality-channel-independent bypass that reaches the same answers without the tag or channel. We separate the two by intervention rather than observation — turning the reality tag off, swapping it, and ablating the channel at answer time — and two positive probes for a static bypass cue both return null, so the bypass has no local signature in the tested factors and is defined by its interventional profile. A data-design ablation that closes the non-tag route to absent yields zero bypass across bands; because the redesign leaves no such route by construction, we read this as design validation plus a trainability result: with the route closed, every seed learns reality-channel use at no accuracy cost, without an anti-bypass penalty. The split is not two settled strategies: under continued training at already-saturated accuracy the same seeds drift, bypass seeds first acquiring reality-channel use then eroding back by 25000 while the deepest genuine seeds resist crossing yet all weaken, so accuracy converges long before reality-channel use, which does not converge within the training we ran. A passive activity monitor is blind to the bypass, whereas the same necessity test, re-applied at a deployed checkpoint, reads out current axis position, a checkpoint-conditional state readout the passive axis cannot provide; greater embedding width shifts the snapshot toward the tag-dependent end on seven of eight bands (one reverses) without removing the route. Deterministic reruns are byte-identical across invocations and two physical nodes. We release the substrate, diagnostics, and per-seed evidence as a testbed for memory-pathway diagnostics.