diff --git a/docs/notes/ideas.md b/docs/notes/ideas.md
index 89eec39..6618289 100644
--- a/docs/notes/ideas.md
+++ b/docs/notes/ideas.md
@@ -63,3 +63,15 @@ Adding subscriptions to `@causaloop/core` would:
 2. Align with Elm's architecture more faithfully
 3. Make replay/restore safe by default — the framework's core value proposition
 4. Position Causaloop as solving a problem that even Elm sidesteps rather than solves
+
+### Conditional subscription pausing
+
+The `subscriptions` function receives the current model, which enables conditionally starting or stopping subscriptions based on state. No consumer currently exercises this — causal-factory's `subscriptions` function ignores the model parameter entirely (always returns the same animationFrame subscription).
+
+Potential use cases:
+
+- Return an empty array when `model.isPaused === true` to pause the game loop
+- Start a countdown timer subscription only during a specific game phase
+- Switch from animationFrame to a slower timer when entity count exceeds a threshold
+
+The dispatcher already handles subscription diffing (`diffSubscriptions`) — adding/removing subscriptions between commits is fully supported. This just needs a real consumer to exercise it.
diff --git a/docs/notes/immutable-bulk-operations.md b/docs/notes/immutable-bulk-operations.md
new file mode 100644
index 0000000..8589209
--- /dev/null
+++ b/docs/notes/immutable-bulk-operations.md
@@ -0,0 +1,80 @@
+# Immutable Bulk Operations: The O(n²) Spread Trap
+
+## Problem
+
+Dispatching N individual messages that each modify the same N-sized collection is always O(n²) with immutable state — regardless of data structure.
+
+## Discovery
+
+In [causal-factory](https://github.com/bitkojine/causal-factory), the "Event Storm" button resets all bots to idle. The naive approach dispatched one `reset_bot` message per bot:
+
+```
+window.triggerMarketCrash = () => {
+    for (const bot of Object.values(latestSnapshot.bots)) {
+        dispatcher.dispatch({ kind: 'reset_bot', botId: bot.id });
+    }
+};
+```
+
+### Array-based bots (`Bot[]`)
+
+Each `reset_bot` did `.map()` over the entire array to find and replace one bot:
+
+```
+bots: model.bots.map(b => b.id === msg.botId ? { ...b, state: idle } : b)
+```
+
+With 30k bots: 30k messages × 30k iterations = **~900M operations**. Browser froze.
+
+### Record-based bots (`Record<string, Bot>`)
+
+Converted to Record for O(1) lookup:
+
+```
+bots: { ...model.bots, [msg.botId]: { ...model.bots[msg.botId], state: idle } }
+```
+
+Lookup is now O(1), but `...model.bots` still **copies all 30k entries** to create the new immutable object. With 30k messages: 30k × 30k copies = **~900M copy operations**. Browser still froze.
+
+### Why this is fundamental
+
+Each message in the dispatcher queue produces a new model. The next message operates on that new model. There is no way to "batch" mutations across messages without breaking the sequential guarantee of MVU.
+
+```
+Message 1: spread 30k entries → new model
+Message 2: spread 30k entries from new model → new model
+...
+Message 30k: spread 30k entries → new model
+```
+
+This is inherent to immutable state management. The bottleneck isn't the lookup — it's the spread.
+
+## Solution: Bulk messages
+
+The correct pattern is a single message that handles the entire operation in one pass:
+
+```
+case 'market_crash': {
+    const resetBots: Record<string, Bot> = {};
+    for (const id in model.bots) {
+        resetBots[id] = { ...model.bots[id], state: { kind: 'idle' } };
+    }
+    return { model: { ...model, bots: resetBots }, effects: [] };
+}
+```
+
+One spread of the collection, one pass over all entities = O(n). With 41,700 bots this runs smoothly at ~45 FPS.
+
+## Key takeaway
+
+**Individual messages are for individual operations.** When you need to touch every entity, use a bulk message. The dispatcher's sequential guarantee is a strength for correctness, but it means each message pays the full cost of immutable state creation. Design messages accordingly.
+
+## When individual entity messages ARE appropriate
+
+The `reset_bot` message type still exists and works well for targeted operations — resetting a single bot in response to a user click, for example. The O(1) Record lookup makes this fast. The problem only arises when dispatching thousands of them in a tight loop.
+
+## Related
+
+- `Record<string, Bot>` is still the better data structure (consistent with `machines`, O(1) individual access)
+- Persistent data structures (HAMTs) could theoretically solve this with O(log n) updates, but add complexity
+- Elm has the same constraint — bulk operations are the standard pattern there too