What the suggestions on the payment screens are worth

Some payment screens show a suggested order for a queue, an estimate of
how long a bank usually takes, a warning that an account may run short, or
a shortlist of statement lines that might match a batch.

This page says what those are, and what they are not allowed to do.

They advise, and they never decide

No calculation in the payment system may approve something, satisfy a
check, stand in for a bank signature, submit a payment, declare that
somebody has been paid, relax the rule that different people do different
jobs, close a question, change any evidence, or refuse to pay somebody.

That is not a promise about intentions. The commands that actually move a
payment take no score of any kind as an input, so there is no route
through which one could reach a decision, and a test fails the build if
one ever appears.

What each one is for

• Queue order. Which question somebody should pick up first, given how old it is, how much money is involved, and whether a control was breached. It changes the order and nothing else: a row at the bottom is worked later, not less.
• Expected time on a rail. How long payments of this kind have actually been taking. It tells somebody a good moment to ring the bank. It never moves a payment out of unknown.
• Liquidity warning. Whether the account looks like it will cover what is scheduled to leave it. It goes to Treasury as a warning. It is not permission to delay somebody who has been approved and budgeted for.
• Match suggestions. A shortlist of statement lines that might be the debit for a batch. A person decides; a score links nothing on its own.
• Duplicate warning. A flag that a payment resembles one already made. A high score raises a question for somebody to look at. A low score excuses nothing: the real protection is a database rule that refuses the same payment twice regardless.

They are allowed to say they do not know

A matcher that cannot separate two candidates says so instead of putting
one at the top of a list. An estimate built from four payments is not
published at all.

That is deliberate. A confident-looking answer built from too little
evidence is worse than no answer, because somebody acts on it.

Every suggestion is on file

Each one is recorded with what it was calculated from, which version of
the calculation produced it, and what the person decided afterwards.

That last part matters most. A calculation whose suggestions are always
followed is either very good or is being rubber-stamped, and the only way
to tell them apart is to count how often somebody looked and decided
otherwise.

New ones are shown to nobody first

A calculation runs quietly for a while before anybody sees it, so it can
be judged against real history rather than against how sensible it sounds.
Anything that has not been through that is marked, and it does not appear
on a working screen.