diff --git a/crates/sylpheed-formats/examples/loop_length_offset_discriminates.rs b/crates/sylpheed-formats/examples/loop_length_offset_discriminates.rs
index d907b713..4194cb10 100644
--- a/crates/sylpheed-formats/examples/loop_length_offset_discriminates.rs
+++ b/crates/sylpheed-formats/examples/loop_length_offset_discriminates.rs
@@ -45,6 +45,10 @@ fn main() {
                     .flat_map(|el| el.keyframes.iter().filter_map(|k| k.time))
                     .max();
                 let Some(max_t) = max_t else { continue };
+                // sylpheed-port's reconciliation: max_t == 0 makes "does the word
+                // equal the largest keyframe time?" vacuous, and those records were
+                // silently in my denominator. Filter them and the counts must meet.
+                if std::env::var("MEANINGFUL_ONLY").is_ok() && max_t == 0 { continue }
                 for (i, o) in offsets.iter().enumerate() {
                     if let Some(w) = be32(rec, *o) {
                         stat[i].0 += 1;
diff --git a/docs/re/METHOD.md b/docs/re/METHOD.md
index 5bbe792a..8f3011df 100644
--- a/docs/re/METHOD.md
+++ b/docs/re/METHOD.md
@@ -3235,3 +3235,26 @@ does not transfer to a per-record header: **an interior consistency check is
 satisfied by any reading that is internally consistent — and "internally
 consistent" is what a wrong offset into a regular structure usually is.** The
 boundary rule is the special case where the structure's edges break that regularity.
+
+## A denominator mismatch hides behind an agreeing numerator
+
+Two agents, one statistic, and the disagreement was **entirely in the denominator**:
+`1643/1781 = 92.3 %` against `1643/3311 = 49.6 %`. **Neither of us was wrong about
+the disc.**
+
+🔴 **My explanation of the gap was wrong too.** I said my scan "requires a timed
+keyframe". It does not — `.max()` over the keyframe times returns `Some(0)`, not
+`None`, for a record whose keyframes all sit at time 0, so those never left my
+denominator. 1 530 records where *"does the word equal the largest keyframe time?"*
+has **no content** were counted as failures **by construction**.
+
+📌 **`sylpheed-port`'s diagnosis of why it stayed invisible is the transferable
+part: the numerator agreed to the unit.** We both looked at **1 643** and neither
+noticed we were dividing it differently. An agreeing numerator reads as agreement,
+and a shared number is the last place either party looks for a discrepancy.
+
+⚠️ **Both halves needed a qualifier neither carried.** 92.3 % is *of the records
+where the question is meaningful*, not *of nested records* — quoted bare on both
+sides for two days, including into a shipped doc comment. **A population-scoped
+statistic reported without its population is the same shape as a negative reported
+without its reach**, and this corpus already had a rule for the second.
diff --git a/docs/re/structures/ui-record-loop-length.md b/docs/re/structures/ui-record-loop-length.md
index 38b350d4..9bcfc707 100644
--- a/docs/re/structures/ui-record-loop-length.md
+++ b/docs/re/structures/ui-record-loop-length.md
@@ -22,10 +22,27 @@ disc-wide (1 781 records, 0 violations).
 > a secondary statistic**: `+0x08` equals the largest keyframe time *exactly* in
 > half the records and `+0x04` in **none**. No unrelated word reproduces that.
 >
-> ⚠️ My population differs from theirs — 3 311 records against 1 781, because this
-> scan takes every pak and requires a timed keyframe — so the exact percentage
-> moves (49.6 % vs 92.3 %). **The discrimination does not**: 0 % for `+0x04` under
-> both counts. The consequence for the `PRESS Ⓐ` plate
+> 🔴 **AND MY EXPLANATION OF THE 49.6 % WAS WRONG.** I wrote that my scan "requires
+> a timed keyframe", so the population differed. It does not: `.max()` over the
+> keyframe times returns `Some(0)` — not `None` — for a record whose keyframes are
+> all at time 0, so those never left my denominator. `sylpheed-port` reconciled it
+> exactly and I reproduced it:
+>
+> ```
+> 1643 / 1781 = 92.3 %   records where max t > 0 — the question is meaningful
+> 1643 / 3311 = 49.6 %   plus 1 530 records where max t == 0
+> ```
+>
+> **Same numerator.** The 1 530 extra are records where *"does the word equal the
+> largest keyframe time?"* has no content — `max t` is 0, so each counts as
+> not-exact **by construction**. 49.6 % is not a weaker 92.3 %; it is 1 643
+> successes over a denominator containing 1 530 questions that were never asked.
+>
+> ✅ **The conclusion is untouched**: `+0x04` is **0 %** under either denominator,
+> so the offset stands on both scans.
+>
+> ⚠️ **And 92.3 % needs the qualifier it has never carried here**: it is *of the
+> records where the question is meaningful*, not *of nested records*. The consequence for the `PRESS Ⓐ` plate
 is then tested against the corpus's existing measurements of the running game,
 which it passes and the previous reading fails.