diff --git a/src/game/ai/orders.h b/src/game/ai/orders.h
index 215c973..878f5fa 100644
--- a/src/game/ai/orders.h
+++ b/src/game/ai/orders.h
@@ -162,6 +162,45 @@ struct ColonizeOrder {
     int trailing = 0;
 };
 
+// ---------------------------------------------------------------------------------------------
+// A CONTRACT FOR WHOEVER BUILDS THE FLEET GATHER
+//
+// These take `fleetId` as a parameter: this layer emits orders and does not form fleets or mint
+// ids.  The subsystem that does is not written yet, and when it is, it inherits a requirement that
+// was measured rather than guessed, so it is recorded here where it cannot be missed.
+//
+// In the original, two processes running the same turn from the same save with the AI's seeds
+// pinned produce different saves -- 35 leaves of 61,147, and all 35 are one transposition: two
+// fleets exchange contents.  The command blocks differ in exactly three words.  The id counter is
+// deterministic; the ids are minted in the same order every time.  What varies is WHICH GROUP EACH
+// ID LANDS ON, because the pass that assigns them walks a container ordered by
+// `fleet->Location` -- a HEAP POINTER, with the fleet pointer breaking ties.  Under LFH
+// randomisation two allocations of the same size land in either relative order, so the original's
+// whole observed outcome set on that turn is the two orderings of two heap pointers.
+//
+// Three consequences bind us:
+//
+//   1. **Order the groups by a key that is a function of the save, and mint in that order.**
+//      `(systemId, then minimum ship id)` with a total tie-break.  Any deterministic key would do:
+//      the original's order is NOT ascending systemId and NOT ascending minimum ship id in any
+//      process observed, so THERE IS NO ORIGINAL ORDER TO MATCH -- we are choosing one member of
+//      its outcome set, not reproducing a hidden rule.  Say so; do not let a later reader mistake
+//      this for a recovered behaviour.
+//   2. **Impose it where the groups are gathered and the ids are minted, not at assignment.**
+//      The fleets already carry their new ids when the assignment walk first touches them, so
+//      canonicalising at assignment is one level too low and fixes nothing.
+//   3. **A byte-exact replay needs the original's order, and the block already carries it.**
+//      List 10 records `{systemId, fleetId, ships[]}` per group in visit order, so every captured
+//      run states its own order for free.  An optional override that replays a captured order is
+//      what turns "equal modulo a relabelling" into "equal" -- three words per new fleet, the same
+//      cost class as pinning a seed.
+//
+// Evidence: sots-re `findings/control-flow/fleet-visit-order-mechanism.md` (the container and its
+// key, measured across three processes including a branch flip),
+// `findings/control-flow/pinned-seed-rich-turn.md` (the residue),
+// `findings/resolutions/2026-09-09-fleet-id-order-residue.md` (the ruling).
+// ---------------------------------------------------------------------------------------------
+
 // List 8. The route is a counted vector of system ids, so a multi-hop order is longer on the wire
 // than a single-hop one -- but it is still ONE element and therefore ONE counter bump.
 struct FleetMoveOrder {