diff --git a/crates/sylpheed-formats/examples/voice_len_vs_subs.rs b/crates/sylpheed-formats/examples/voice_len_vs_subs.rs
new file mode 100644
index 0000000..44e19ab
--- /dev/null
+++ b/crates/sylpheed-formats/examples/voice_len_vs_subs.rs
@@ -0,0 +1,78 @@
+//! Is audio actually missing from the resupply banks? Ask the subtitles.
+//!
+//! The `.slb` decode of `VOICE_D_453` yields 0.14 s, which "looks too short" —
+//! but that judgement was an impression. The subtitle track for each movie
+//! carries cue times, so it says independently how long the spoken line runs.
+//! If the last cue lands near the decoded length, nothing is missing; if it lands
+//! far past it, audio really is being lost.
+use sylpheed_formats::{movie_subtitle, slb, PakArchive};
+
+/// XMA1 at 48 kHz stereo, 16-bit → bytes per second of PCM.
+const PCM_BYTES_PER_SEC: f32 = 48000.0 * 2.0 * 2.0;
+
+/// Decode one sub-wave with FFmpeg and return its length in seconds. Measured,
+/// not estimated from a compression ratio — the ratio guess was the first version
+/// of this and it is not good enough to hang a conclusion on.
+fn decoded_secs(riff: &[u8]) -> f32 {
+    use std::io::Write;
+    use std::process::{Command, Stdio};
+    let Ok(mut c) = Command::new("ffmpeg")
+        .args(["-v", "error", "-i", "pipe:0", "-f", "s16le", "pipe:1"])
+        .stdin(Stdio::piped())
+        .stdout(Stdio::piped())
+        .stderr(Stdio::null())
+        .spawn()
+    else {
+        return f32::NAN;
+    };
+    let buf = riff.to_vec();
+    let mut stdin = c.stdin.take().unwrap();
+    std::thread::spawn(move || {
+        let _ = stdin.write_all(&buf);
+    });
+    let out = c.wait_with_output().expect("ffmpeg");
+    out.stdout.len() as f32 / PCM_BYTES_PER_SEC
+}
+
+fn main() {
+    let disc = std::env::var("SYLPHEED_DISC").expect("set SYLPHEED_DISC");
+    let snd = PakArchive::open(format!("{disc}/dat/sound.pak")).expect("sound.pak");
+    let lang = PakArchive::open(format!("{disc}/dat/movie/eng.pak")).expect("eng.pak");
+
+    // movie → the bank the record table binds it to.
+    let bound = [
+        ("hokyu_LS_s02A", 450),
+        ("hokyu_LS_s09A", 451),
+        ("hokyu_DS_s13A", 452),
+        ("hokyu_LS_s02H", 453),
+        ("hokyu_DS_s07H", 454),
+    ];
+    println!(
+        "{:<18} {:>6} {:>10} {:>12} {:>12}",
+        "movie", "bank", "last cue s", "decoded s", "verdict"
+    );
+    for (movie, n) in bound {
+        let cues = movie_subtitle::track_voice_cues(&lang, movie);
+        let last = cues.iter().map(|(_, t)| *t).fold(0.0f32, f32::max);
+        let path = format!("eng\\etc\\VOICE_D_{n}.slb");
+        let Some(entry) = snd.find_by_name(&path) else { continue };
+        let bytes = snd.read(entry).expect("read");
+        // Sum the sub-waves' payloads as the decoder currently sees them.
+        let riffs = slb::to_xma_riffs(&bytes);
+        let secs: f32 = riffs.iter().map(|r| decoded_secs(r)).sum();
+        // One-directional: a subtitle that appears at t seconds cannot sit inside
+        // a clip shorter than t. A cue at 0.0 tells us nothing either way.
+        let verdict = if last == 0.0 {
+            "no cue signal"
+        } else if last > secs {
+            "AUDIO MISSING"
+        } else {
+            "consistent"
+        };
+        println!(
+            "{movie:<18} {:>6} {last:>10.2} {secs:>12.2} {verdict:>12}   cues={}",
+            format!("D_{n}"),
+            cues.len()
+        );
+    }
+}
diff --git a/docs/re/BACKLOG.md b/docs/re/BACKLOG.md
index d6adc7e..07e9763 100644
--- a/docs/re/BACKLOG.md
+++ b/docs/re/BACKLOG.md
@@ -1057,9 +1057,18 @@ premise was wrong.**
   Byte coverage was the wrong success metric. The rule also matches **1524 of
   8021** RIFF-bearing `sound.pak` entries, including `RT*` banks that work today,
   so it risked a wide regression to not-fix five banks.
-  ▶️ Next cheap probe: vary the synthesised `fmt` (channels / streams / sample
-  rate) instead of assuming the container. Then a human has to listen; audio
-  judgement cannot be done here. See
+  ✅ **(same day) Audio really IS missing — proven by the subtitle cue times**,
+  not by the "sounds too short" impression the docs recorded. A subtitle that
+  appears at *t* seconds cannot sit inside a clip shorter than *t*, and three of
+  five banks fail that: `D_450` cue 4.00 s vs 1.41 s decoded, `D_451` 3.70 vs
+  1.81, `D_453` **4.70 vs 0.07**. The other two have their only cue at 0.0 s and
+  give no signal. Artifact `examples/voice_len_vs_subs.rs`, FFmpeg-measured.
+  ❔ **The fmt-variation probe was INCONCLUSIVE** — 36 combinations over
+  `VOICE_D_453`'s leading region all produced 0 PCM bytes, *including* ones
+  equivalent to the crate's working `synth_xma1_fmt`, so the probe tested my
+  hand-built `fmt` chunk rather than the hypothesis. Not evidence the region is
+  non-XMA. ▶️ Retry building the chunk with the crate's own helper and varying
+  its parameters. See
   [`voice-bank-leading-region.md`](voice-bank-leading-region.md).
 * ❌ **(2026-08-25) My own boot-nav diagnosis, MEASURED AND WITHDRAWN.** I said
   the run died because `skip_intro.sh` gates the title test at `rmse <= 1500`
diff --git a/docs/re/voice-bank-leading-region.md b/docs/re/voice-bank-leading-region.md
index 7b76a2c..66efe2f 100644
--- a/docs/re/voice-bank-leading-region.md
+++ b/docs/re/voice-bank-leading-region.md
@@ -120,6 +120,43 @@ it.** Two measurements killed it:
 The refuted attempt is recorded in `slb.rs` beside the code, so the next person
 does not re-derive the arithmetic and re-make the same change.
 
+## ✅ Audio really is missing — proven by the subtitles, not by impression
+
+The corpus's original wording was that 0.14 s is "far too short for the spoken
+line". That is a judgement, and judgements about audio cannot be made in this
+container. The subtitle tracks settle it instead: each carries **cue start
+times**, and a subtitle that appears at *t* seconds cannot sit inside a clip
+shorter than *t*.
+
+Decoded with FFmpeg (measured, not estimated from a compression ratio — the
+first version of this used an 8:1 guess and that is not good enough to hang a
+conclusion on), artifact `examples/voice_len_vs_subs.rs`:
+
+| movie | bank | last cue | decoded audio | verdict |
+|---|---|---|---|---|
+| `hokyu_LS_s02A` | `D_450` | 4.00 s | 1.41 s | **audio missing** |
+| `hokyu_LS_s09A` | `D_451` | 3.70 s | 1.81 s | **audio missing** |
+| `hokyu_LS_s02H` | `D_453` | 4.70 s | **0.07 s** | **audio missing** |
+| `hokyu_DS_s13A` | `D_452` | 0.00 s | 1.21 s | no signal |
+| `hokyu_DS_s07H` | `D_454` | 0.00 s | 0.21 s | no signal |
+
+Three of the five are decisive; the other two have their only cue at 0.0 s, which
+tells us nothing in either direction. So **something is genuinely missing from
+these banks** — independent of anything above, and now measured rather than felt.
+
+## ❔ Varying the XMA format — inconclusive, and for a boring reason
+
+The recorded next step was to vary the synthesised `fmt` (channels, streams,
+sample rate) rather than assume the container. I tried 36 combinations over
+`VOICE_D_453`'s 22-packet leading region and **every one produced 0 PCM bytes** —
+including combinations that should be equivalent to the crate's own
+`synth_xma1_fmt(2, 2, 48000)`, which does at least parse (it yields 1792 bytes).
+
+That means the probe tested **my hand-built `fmt` chunk**, not the hypothesis. It
+is not evidence that the region is non-XMA. The next attempt should build the
+chunk with the crate's own helper and vary its parameters, rather than
+hand-rolling the WAVEFORMATEX.
+
 ## What this does not settle
 
 * ❔ **What the leading region holds.** Its *size* is now exact (1392 + n·2048)