Learning under a Decided Gate: Changing a table that an admitted shortcut reads — the precondition taken from the shortcut's admission record, a chain of table texts, and the unshortened loop when the evaluated text does not fit

Programs that choose among candidates — a planner, a controller, a decision system — save work by shortcuts that stop a loop or skip a candidate where the choice cannot change. Often a shortcut is sound only under a condition on a table it reads. If the table is learned in service, the shortcut's soundness depends on every new version of it. This paper describes a coupling of the change of such a table to the admission of its reader. The shortcut acts only under an admission record: statements about its executed text, decided for all inputs and bound to that text by its hash, one of them valid under a precondition V over the table values. A proposed table is taken over only when V — taken from that record — is decided over the text that would carry it; every acceptance and rejection is a record of a hash chain; the new text acts only for executions begun afterwards; and if the table text about to be evaluated is not the one the chain names, the loop runs without the shortcut. Learned quantities are readable tables of fixed-width numbers; no neural network is involved. The general thought that a change is taken over only after a decision procedure has found it compatible is the work of Ohad Asor and IDNI AG and is not claimed. Run with the Tau language: a small form with four candidates and no game, in which the check kept a replaced table from changing the choice; and, in a larger program, one accepted transition and three learning rounds in which the decided gate accepted every proposal and a measurement placed after it rejected every one. Nothing here says that learning helped. ---