diff --git a/crates/sim/src/tuning.rs b/crates/sim/src/tuning.rs
index f403ce4..62a1b08 100644
--- a/crates/sim/src/tuning.rs
+++ b/crates/sim/src/tuning.rs
@@ -149,7 +149,7 @@ pub const FRESH_PICKS: Picks = Picks {
     level: BotLevel::Basic,
     tyres: Tyres::On,
     laps: DEFAULT_LAPS,
-    // Until rain has been played (ADR 0103).
+    // Until rooms have played Rain (ADR 0104).
     weather: Weather::Dry,
 };
 /// Seconds the karts are held on the grid before Go: the start lights, one
diff --git a/docs/adr/0104-rain-keeps-its-starting-grip-and-dry-stays-the-default.md b/docs/adr/0104-rain-keeps-its-starting-grip-and-dry-stays-the-default.md
new file mode 100644
index 0000000..bfa7e38
--- /dev/null
+++ b/docs/adr/0104-rain-keeps-its-starting-grip-and-dry-stays-the-default.md
@@ -0,0 +1,73 @@
+# ADR 0104: Rain keeps its starting grip and Dry stays the default
+
+## Status
+
+Accepted (2026-11-09). Records a bot probe in place of the playtest and retune
+ADR 0103 left for after the build; nobody has played rain yet. Changes no
+factor in `crates/sim/src/tuning.rs`.
+
+## Context
+
+ADR 0103 set one row of rain factors as a starting tune and left four
+questions for a playtest: whether worn tyres in the wet read as rain or as a
+broken kart on a keyboard, whose steering is all or nothing; how rain plays on
+lap 3 of 3, where a stop cannot pay, and on the oval, whose 3-lap race ends
+before the track has had its 20 s to wet; whether an Ace bot's dry lap costs
+need a wet row; and whether Rain can become the default pick.
+
+A throwaway probe raced eight Ace bots on Tyres On over 40 seeds, dry against
+rain on the same seeds, on every course at 3 and 5 laps, split by the lap the
+seed picked: 18 to 22 races a row at 3 laps, 12 to 14 at 5. Beside it, a lone
+Ace bot held at each life of `BOT_TYRE_LOSS_S_PER_LAP` raced Bell Deer,
+Goldstone and Monarch without their pit lanes or item boxes, on seeds whose
+rain began on lap 2, and timed laps 4 and 5, which run fully wet. It is too
+small to tune from, and it is the evidence below.
+
+- **What a stop wins is about what the dry table says.** Against a dry lap at
+  the same life, a wet lap costs about 2 to 3 s from 100% down to 30% and under
+  1 s at 20% and 10%. Measured from new tyres in the same weather, going from
+  30% to new wins 10.8 s a lap wet against 9.6 s dry; the table holds 10.2, and
+  its measuring test allows 1 s either way. At 0% the wet laps were quicker than
+  the dry ones on the mean, too noisy across courses to read further; ADR 0103
+  expected a worn-out kart to slide less in the wet.
+- **Lap 3 of 3 changes the bots' stops by nothing.** Stops made from the rain's
+  lap on are 0.16 a kart wet against 0.15 dry on Bell Deer, 0.11 against 0.12 on
+  Goldstone and 0.26 against 0.26 on Monarch. The winner is 2 to 3.5 s slower.
+  How a human's last lap feels is not something bots show.
+- **The oval barely gets wet.** Rain on lap 3 of 3 leaves the track 72% wet at
+  the winner's flag and costs the winner about 0.6 s; on lap 2 it is fully wet
+  and costs about 0.2 s.
+- **Five laps add a few stops and some capped karts.** Rain adds 0.1 to 0.5
+  stops a kart over five laps, most on Monarch. Goldstone sends more karts past
+  the end cap in rain, of about 100 karts a row: 18 against 16 with rain on lap
+  2, 19 against 8 on lap 3 and 13 against 6 on lap 4. The other courses move by
+  noise either way, and the oval's five laps overrun the cap in the dry as much
+  as in rain. Karts on grass are 0 or 1 in every row.
+
+Neither the keyboard nor the gamepad playtest was held. The maintainer chose to
+record the probe now and leave play for later.
+
+## Decision
+
+- **Rain keeps ADR 0103's starting factors**: steering rate and acceleration x
+  0.92, braking x 0.80, wash-out + 0.05 rad to at most 0.18 rad.
+- **An Ace bot's lap costs stay dry.** The stop a bot weighs wins within the
+  table's own tolerance wet or dry, since no stop takes the rain away, and its
+  stops barely change. There is no wet row.
+- **Lap 3 of 3 and the oval stay as they are.** A last lap with no stop worth
+  making, and an oval that only partly gets wet, are what a seeded lap and 20 s
+  of wetting give, and neither broke a bot race.
+- **`RACE_END_CAP_S` stays**, the choice ADR 0103 left to the maintainer.
+  Goldstone's five laps in rain are recorded here and left alone.
+- **Dry stays the default pick** until rooms have played Rain.
+
+## Consequences
+
+- Nothing in the sim, the protocol or the replay fixtures changes.
+- Whether worn tyres in the wet read as rain or as a broken kart on a keyboard,
+  and how lap 3 of 3 in the rain feels to a human, are still open. A playtest
+  that answers them gets its own ADR.
+- Rain becoming the default is a later decision, with play in rooms as its
+  evidence.
+- If Goldstone's five-lap rain races start ending at the cap in play, the
+  maintainer decides between the row and `RACE_END_CAP_S`, as ADR 0103 left it.
diff --git a/tools/repo-checks/test/adr.test.ts b/tools/repo-checks/test/adr.test.ts
index e968697..2beb018 100644
--- a/tools/repo-checks/test/adr.test.ts
+++ b/tools/repo-checks/test/adr.test.ts
@@ -108,6 +108,7 @@ const DECISION_SLUGS = [
   "a-ramp-is-a-slope-across-part-of-the-road",
   "a-hop-on-the-lip-is-a-trick",
   "rain-changes-grip-on-a-lap-the-seed-picks",
+  "rain-keeps-its-starting-grip-and-dry-stays-the-default",
 ];
 
 const expectedFiles = DECISION_SLUGS.map(