Every cycle, a multi-head combination weigher solves a hard problem. It has fourteen hoppers, each holding a different amount of product, and it has to choose the subset of them whose combined weight lands closest to target without going under. With fourteen heads there are 16,383 possible combinations. It solves that, several times a second, all day.
It can only do that if there is product in the hoppers to choose from. And here is the part that gets missed: the weigher doesn't tell you which heads are going short, or why.
What starvation does to the number
When four of fourteen heads are running short, the machine doesn't stop. It carries on making combinations out of the ten heads that have product. The combinations it can form are worse (coarser, further from target), so giveaway drifts up and misses drift up with it.
From the outside, this looks exactly like a weigher problem. The efficiency number is bad. The giveaway number is bad. The machine is the thing in the room with a screen on it, so the machine gets blamed, and somebody spends a shift tuning feeders that were never the problem.
The measurement that settles it
There is one measurement that separates the two cases cleanly, and it is not complicated: mean fill per head, under identical commanded settings.
If every head is commanded the same and every head fills roughly the same, the feed is fine and any remaining problem belongs to the weigher or the product. If every head is commanded the same and a contiguous arc of heads fills at a fraction of the rest, the product is not arriving evenly, and no amount of feeder tuning will fix it.
That is real data from a live autonomous control session on a 14-head machine running a 170 g pack. Every head was given identical feeder settings. Heads 6, 7, 8 and 9 filled at roughly one third the rate of heads 4, 5, 10, 11 and 12.
They are adjacent. That matters. A contiguous arc of starved heads points upstream, to how product is being distributed across the machine, not to four independent feeder faults that happen to sit next to each other.
A second machine on the same line flagged a matching infeed asymmetry the same session.
Why nobody had this chart
Not because it is hard to compute. Because computing it requires reading every head's state every cycle, for a sustained period, while holding the commanded settings constant and knowing that you held them constant. That is not something a person standing at a panel can do, and it is not something the machine volunteers.
It falls out for free from a system that was already reading all fourteen heads every second in order to control them. The diagnostic was never the goal. It is a by-product of doing the control job properly.
The uncomfortable corollary
The controller in that session pushed every feeder to its limit to make up for the starved heads, and still beat the recipe's previous best by 18.65 percentage points. Which means every performance figure from that run was achieved on a machine with a mechanical feed fault for the entire session.
We could treat that as a caveat to bury. It is more useful as a floor. Whatever those numbers say the system does, it does it while starved on four of fourteen heads.