diff --git a/plan/candidates.md b/plan/candidates.md
index c928c19..d778ca7 100644
--- a/plan/candidates.md
+++ b/plan/candidates.md
@@ -2649,3 +2649,45 @@ enemy settles, one of ours moves so the stand set gains and loses stands, and
 that same step lands inside a third enemy's envelope so its positions shift
 under it - then asserts the fixture produced all three before comparing against
 a full re-sweep of the same state.
+
+### Across a round boundary it carries nothing, and the reason is a product
+
+With rebuilds two to four cheap, the *first* rebuild of each round is what is
+left: about three quarters of a round's movement planning. `Memos::take` drops
+every entry from an older round, so that rebuild is always a full sweep. The
+obvious next question is whether it has to be.
+
+It does. Measured the same way, before anything was built - 11 boundary
+crossings over four rounds of a pinned 4v4:
+
+| | within a round | across a boundary |
+|---|---|---|
+| stands shared | 93-100% | 0-11% (one outlier at 80%) |
+| enemy positions shared | 93-99% | 0-18% |
+| **pairs carried (ceiling)** | **91-99%** | **0-3%, median 0%** |
+
+**A pair needs both halves, and the shares multiply.** Within a round one unit
+moves between rebuilds, so `0.97 x 0.97` is about 94% and the carry pays.
+Between round `N`'s last rebuild and round `N + 1`'s first, *every* unit has
+moved: our own reachable set is drawn from a new hex, and a `Reach` pins MP
+spent, so even a hex we can still reach is usually a different stand. About
+`0.06 x 0.04` is a fifth of one percent. The one crossing that kept 80% of its
+stands still carried 3%, because all four enemies had moved regardless.
+
+No relaxation of the key recovers this, and the instrument checked rather than
+assumed it: reading "the same machine" three ways - the whole wire record, the
+record without position and turn state, and that without heat as well - reached
+**0%** in every case, because the geometry is already gone before state is
+asked about.
+
+**What this also confirms.** `carry::Carried` pins the *whole* wire record of
+our unit and of every enemy, which includes `locations` and their armour. The
+hazard that armour changes in the firing phase is therefore already guarded, and
+the instrument shows the guard firing exactly when it should: **0 of 4** enemies
+kept their whole record across any boundary, and `mover` equality failed at all
+11 crossings because our own unit had moved. Within a round neither of those
+ever fires, which is why the strictness costs nothing there.
+
+So the first rebuild of a round is not reachable this way. If it is worth
+attacking, it is worth attacking as a sweep - the volley memo, or the size of
+`L x M` - and not as a carry.