Recursive Channel Cost in Budgeted Arbitration: Reliability-Blind Arbitration Cannot Be Repaired for Free

A budgeted agent given several channels over one observation stream must choose what to read, and the channels hand it claims that compete. We add the arbiter to the access-structure formalism and measure what arbitration does. (1) A blind arbiter attains a presentation-weighted average of its channels' reliabilities rather than the best of them, and we state the identity with a third term that vanishes exactly when the arbiter is blind, so blindness becomes a hypothesis a transcript settles: on this grid, settled for (C = +0.016 over 366 arbitrated decisions). Telling the arbiter which channel to follow, by a one-line directive keyed on the query class, is the first manipulation in this line of work to act at the arbitration point and survive its own control; the directive helps where its classifier is right every time and hurts where it is right about half, so knowing which channel to trust is itself another unreliable channel over the same stream. (2) Appending a weaker shard of the same index (same form, weaker content) subtracts capability in proportion to its weakness, the two blocks arbitrated symmetrically; conditioning on the loop's exit, the subtraction lives in committed accuracy and not in rounds, so appending is arbitration, not budget. (3) The compressive channel is followed where it holds the queried quantity (90.2 one of 114$ injected decisions follows a planted wrong value) and holds nothing where the answers are attributes; a registered probe that renders its facts as citations leaves a silent channel unfollowed, so presentation cannot rescue a channel with nothing to say. Grounding is a function of the access loop's exit, not of the architecture: separated from commitment, the apparent benefit of removing a channel reverses. A table lists every claim with its interval and verdict, and every control is read against the spread of a difference known to be nothing.