diff --git a/POLISH-PLAN.md b/POLISH-PLAN.md
index e4ed711..0e22036 100644
--- a/POLISH-PLAN.md
+++ b/POLISH-PLAN.md
@@ -181,21 +181,49 @@ oder nicht."
 Note path 2 has a partial gate already: `G.shameGame` withholds the tin if *they* took a
 game since the last delivery. That is not the same as winning the duel.
 
-**Build:** make winning the duel the ONLY source of weapons.
-- Delete the `tinDue` scheduling in `splatPoint()` and the `deliverTin()` call in
-  `startChangeover()`. Keep the "CHANGEOVER" flash. `G.tinDue` / `TUNE.ballChangeFirst` /
-  `ballChangeEvery` / `G.shameGame` all become dead — remove them rather than leaving
-  vestigial state.
-- Keep `resolveInterlude('P') → deliverTin()`.
+**The design decision (Oliver, 2026-07-27):** duels must come round on their own after a
+while, "damit es immer die chance auf waffen gibt". So the schedule is not the problem — it
+is attached to the wrong event. **Move it from delivering tins to calling duels.** The
+cadence stays, the reward becomes something you have to earn, and the umpire-restores-order
+framing is unchanged.
 
-**Consequence to check:** duels only trigger when the court is clear and a racket zombie is
-the last one standing (`gameplay.js:~760`). If that turns out to be rare, weapons become
-near-unobtainable. **Measure before shipping:** run a 5-minute simulated session and count
-duels. If it's under ~1 per two games, tell Oliver rather than silently re-adding a second
-source — the design question (how often should duels happen?) is his call.
+**Why the duel is rare today:** the trigger at `gameplay.js:~760` requires
+`zombies.length === 1 && zombies[0].variant === 'r' && !zombies[0].dueled` — the court has to
+be completely empty apart from one never-yet-dueled racket zombie. Racket zombies only appear
+when you win a game (`splatPoint`) or survive a horde (`endHorde`), and fresh shamblers spawn
+again the moment a clear finishes. So it fires by luck.
 
-**Acceptance:** simulate a full game where the player loses the duel; `G.bombs` etc. must be
-unchanged. Win one; exactly one tin arrives.
+**Build:**
+- Rename the schedule to what it now schedules: `TUNE.ballChangeFirst`/`ballChangeEvery` →
+  `duelFirst`/`duelEvery`, `G.tinDue` → `G.duelDue`. Keep the same arithmetic in
+  `splatPoint()`. Suggested start: first duel after game 2, then every 3.
+- Delete the `deliverTin()` call in `startChangeover()` (keep the "CHANGEOVER" flash).
+  `G.shameGame` — which withheld a tin if they took a game — has nothing left to gate;
+  remove it rather than leaving vestigial state.
+- Keep `resolveInterlude('P') → deliverTin()` as the only weapon source.
+- **Fire the scheduled duel off the existing clear beat.** `proceedAfterClear()`
+  (`gameplay.js:290`) already runs at the one moment the court is guaranteed empty — the
+  player just cleared the wave by hand, and no new shamblers have spawned yet. That is the
+  natural slot. There: if `G.duelDue`, consume it, ensure a racket zombie exists
+  (`spawnRacketZombie()` if none, and clear `dueled` on the existing one so a veteran can be
+  challenged again), and start the interlude instead of the changeover/game beat. The
+  changeover follows after the duel resolves.
+- Leave the existing opportunistic trigger in place. The schedule is a floor, not a ceiling —
+  a lucky clear can still produce a bonus duel, which is pure upside for weapon supply.
+
+**Why this slot and not a timer:** clearing the wave is already mandatory before play
+advances (`G.clearing`, a settled feel decision Oliver was emphatic about), so the schedule
+can never stall. A wall-clock timer would need its own "wait for a safe moment" logic and
+could fire mid-swarm.
+
+**Acceptance:**
+- Simulate ~10 games. A duel must occur at game 2, 5, 8 … regardless of how the waves fall.
+- Lose a duel → `G.bombs`/`triples`/`flames`/`guns` unchanged, and the next duel still
+  arrives on schedule.
+- Win one → exactly one tin, and the item-8 modal names it.
+- No path reaches `deliverTin()` except `resolveInterlude('P')`
+  (`grep -n "deliverTin" *.js` should show the definition, that one call, and the `window.T`
+  test export — nothing else).
 
 ---