custody

The mechanism is that the money never sits anywhere in between. A client signs an authorization naming an amount, an asset and a recipient; the facilitator submits that signature. The funds move from the payer's address to the address the payer signed for, and no account belonging to the service is a stop along the way.

What the property buys is narrow and worth stating exactly. Because the recipient is inside the signed message, a facilitator cannot send the payment somewhere else, and because it holds no balance there is nothing of the payer's for it to lose or freeze. Those are the risks non-custody removes.

What it leaves open is everything that is not the money. A facilitator sees the terms and the signature; it can decline to submit, submit late, or fail to submit at all, and under the protocol's own error vocabulary a settlement can be reported as pending when its confirmation could not be established. A signed authorization also stays a bearer instrument until it is used or lapses, which is what replay protection exists to bound.

The word is used differently a step away, and the difference matters when comparing rails. On the card side an agent wallet does hold the credential and issues a scoped grant against it, so the holding is the point rather than the thing being avoided.