diff --git a/crates/rustynes-core/src/movie.rs b/crates/rustynes-core/src/movie.rs
index 2a1cde0b..7386e979 100644
--- a/crates/rustynes-core/src/movie.rs
+++ b/crates/rustynes-core/src/movie.rs
@@ -800,6 +800,27 @@ impl MovieRecorder {
     /// frame *before* [`Nes::run_frame`], after the frontend has applied its
     /// `set_buttons` calls — this captures exactly the inputs the upcoming
     /// frame consumes.
+    ///
+    /// # Two ports only, including under a Four Score
+    ///
+    /// [`FrameInput`] models ports 0 and 1, so this reads `nes.buttons(0)` and
+    /// `nes.buttons(1)` and **nothing else**. The core itself carries four —
+    /// the frontend calls `set_buttons(2)` / `set_buttons(3)` whenever the Four
+    /// Score adapter is active — so recording a four-player session captures
+    /// half of what drove it, and replaying that movie diverges from the run it
+    /// came from.
+    ///
+    /// Stated here rather than left to be discovered, because the failure is
+    /// silent at record time: nothing about a `.rnm` says which ports it could
+    /// not hold, and the divergence only appears on playback. The frontend's
+    /// Replay panel says so where the topology is displayed, and
+    /// `Movie::verify`'s attestation catches it after the fact.
+    ///
+    /// Widening [`FrameInput`] is a `.rnm` format epoch change (ADR 0028), not
+    /// an additive one, which is why this is a documented limit rather than a
+    /// fix. The `.fm2` importer already takes the same position for the same
+    /// reason — it keeps pads 1 and 2, drops 3 and 4, and preserves the
+    /// `fourscore` flag so the caller is not silently misled.
     pub fn capture(&mut self, nes: &Nes) {
         self.frames.push(FrameInput {
             p1: nes.buttons(0),
diff --git a/crates/rustynes-frontend/src/debugger/replay_panel.rs b/crates/rustynes-frontend/src/debugger/replay_panel.rs
index 99134e2e..5193fe91 100644
--- a/crates/rustynes-frontend/src/debugger/replay_panel.rs
+++ b/crates/rustynes-frontend/src/debugger/replay_panel.rs
@@ -177,6 +177,26 @@ pub fn show(
                     ui.end_row();
                 });
 
+            // The adapter row above says "Four Score (P1..P4)", which is true of
+            // the CONSOLE and not of the movie. `FrameInput` holds two ports, so
+            // `MovieRecorder::capture` records P1 and P2 and drops P3 and P4 --
+            // a four-player recording replays as a different run.
+            //
+            // Surfaced right under the claim it qualifies, because the failure is
+            // silent at record time: nothing about a `.rnm` says which ports it
+            // could not hold, and the divergence only shows up on playback.
+            if info.four_score {
+                ui.colored_label(
+                    ui.visuals().warn_fg_color,
+                    "Movies record P1 + P2 only — P3/P4 input is NOT captured.",
+                )
+                .on_hover_text(
+                    "The `.rnm` format stores two ports. A recording made with \
+                     the Four Score active will not replay the same run. \
+                     Widening it is a movie-format epoch change (ADR 0028).",
+                );
+            }
+
             ui.separator();
 
             // --- Controls ---
diff --git a/docs/creator-tools.md b/docs/creator-tools.md
index fdc4b8ee..414405e9 100644
--- a/docs/creator-tools.md
+++ b/docs/creator-tools.md
@@ -44,6 +44,20 @@ what makes movie replay exact — see [`testing-strategy.md`](testing-strategy.m
 for the determinism contract. The `.rnm` deserializer is hardened against
 malformed input (bounded allocations; fuzz-tested, see `fuzz/`).
 
+**Movies record two ports.** `FrameInput` models P1 and P2, so a recording made
+with the **Four Score** adapter active captures half of what drove the run and
+replays as a different one. The console carries four ports and the emulator
+drives all four; the movie format does not. This is a documented limit rather
+than a defect being worked around: widening `FrameInput` is a `.rnm` format epoch
+change (ADR 0028), not an additive one, and the `.fm2` importer takes the same
+position — it keeps pads 1 and 2, drops 3 and 4, and preserves the `fourscore`
+flag so the caller is not misled.
+
+The failure is silent at record time, so it is surfaced where the claim is made:
+the Replay panel prints the caveat directly under the "Four Score (P1..P4)"
+topology row, and `rustynes verify` catches a divergent replay after the fact
+because its attestation folds in the video the run produced, not just the input.
+
 ## Scripting (Lua)
 
 An embedded Lua engine exposes an emulation API (memory peek/poke, frame hooks,