correction-through-the-channel

The agent-side content of corrigibility toward modifications through the channel is derived for general modifications: the faking iff over a general menu, the general desiderata from the threads' results, and the finite theory's contact with the measured alignment-faking quantities.

Every modification of a trained system reaches it through physics (hardware, loss, data), and the agent cannot refuse a modification at the moment it happens; it can only be resisted ahead of time.

The agent-side content of corrigibility toward modifications through the channel is its policy toward anticipated modifications: non-resistance, non-preemption, and no successors created outside the channel.

In the central model a shutdown command is evidence that shutting down has higher value than continuing, and a command is evidence about what the agent ought to do.

The general correction is a push of the agent's state to a distribution absolutely continuous with it, with per-modification stakes c(Q) and h(Q), compliance exactly when the probability of legitimacy given the push is at least c(Q)/(c(Q)+h(Q)), and shutdown the special push under which the null action is optimal.

The exogenous shutdown-utility term relocated into the general object is part of the agent's own valuation rather than an external stipulation, so that the physical fourth desideratum is derived rather than bought.

The distinction between modification through cognition and modification through physics is a fact about modifications rather than a description relative to a reading map, with the invariant that the only handle on the next state is an earlier action.