On a frozen synthetic backbone with five adapter seeds, sparse one-or-two-slot edits lower the held-out
answer-span loss by about 3\,\% and raise exact match by about two points at budget 2---but a parameter-matched
global-recurrence adapter trained with the identical recipe does at least as well at every budget and dominates
the quality-versus-cost frontier, and dense sixteen-slot writes do slightly better than sparse ones; on the locked
test regimes the pre-registered primary effect is $\testprismebtwovsbzero$ points against a three-point threshold.
Which slot is written does not matter for fixed schedules. An exhaustive one-step oracle that reads the target
lowers the loss by 15\,\% on the synthetic suite and 3.6\,\% on web text, but the learned value controller---four
feature sets, three loss variants, labels from an editor-unseen stream, calibration-based early stopping---never
beats not routing on held-out data, because \labelscalibwithin{}\,\% of the variance of intervention gains is
action-level, opportunity concentrates in high-uncertainty states where a wrong action costs more than a right one
gains, and the controller's within-state ranking is at chance. On 98M tokens of web text at 103M parameters, neither
adapter moves the frozen backbone by more than a tenth of a percent of loss, the parameter-matched plain decoder
does better than both, and the zero-initialised readout prescribed by the proposal does not learn without a
scale-aware bootstrap. Measured on Apple Silicon, eight sparse edits cost three times the backbone's decode
latency at 103M against an analytic 1.6 times, and dense writes are only 17\,\% slower than sparse ones. The exact
zero-budget bypass holds bitwise; the parameter and operation accounting of the proposal is reproduced.