How a plan is scored
The violations panel says what is wrong. The score says how close the plan is to right, in one figure, so two plans of the same guest list can be compared and a change can be judged.
Every rule on KB-2 contributes, hard and soft alike.
The arithmetic
Each rule declares how many chances the guest list gave it, and how many of them it missed.
sub-score = 1 − missed ÷ opportunities
plan score = weighted mean of every rule's sub-score, on 0–100
Opportunities are what the rule could have got right, given the guest list. Conflicts gets one per recorded conflict pair. Partners adjacent gets one per couple, minus the pairs KB-2 exempts. Capacity gets one per table. A rule the guest list gave no chances — no conflict pairs recorded, say — has no sub-score and is left out of the mean rather than counted as perfect.
Opportunities are counted over the guest list, not over how much of the plan is seated. Without that, a plan leaving most people unseated scores higher than one seating everybody with a few things wrong, and an incomplete plan outscoring a complete one makes the figure useless for the comparison it exists to serve. The cost is that an empty plan scores nought rather than being unscored. That is deliberate.
Rounding is plain and there is no cap, so 100 means the mean rounds to 100. It has never meant the plan is clean.
Weight
Weight is declared on each rule. A hard rule defaults to 3 and a soft rule to 1, because missing a thing that stops the plan going out is not the same as missing a thing the couple may not mind.
A rule may set its own and override the default, so changing how much one thing matters is a line in that rule's file rather than an edit to anything shared. Nothing else has to know.
The score is not permission
This is the part to get right, because the figure is easy to read as a verdict.
A plan publishes when its hard violation count is zero, and for no other reason. The score has no say in it. A plan can score well and be unpublishable, and a plan can be publishable and score poorly because the couple's preferences have been widely missed.
So the score is never shown on its own. Wherever it appears, whether the plan can go out appears with it, in words. A figure alone invites somebody to read 96% as finished and send it.
What changed, and why hard rules are in it now
The score began over the soft rules only, on the reasoning that hard violations decide publishability and should not move a number about fit.
That left a gap. A plan can be unpublishable in one small way or in twenty, and the old score read identically for both — the only thing distinguishing them was a list somebody had to count. "How far from publishable am I" is a real question and the score could not answer it.
Counting every rule answers it, and the rule above is what stops the answer being mistaken for permission.
Open
The hard default of 3 against a soft 1 is a starting ratio, not a measured one. Say if a hard miss should weigh more heavily than that, or if any single rule should sit outside the default.
Whether a plan carrying a hard violation should be barred from rounding up to 100 is also open. It is currently not: rounding stays plain, and the publishable state beside it is what carries the truth.