diff --git a/crates/gc-extract/tests/disc.rs b/crates/gc-extract/tests/disc.rs
index 9f10abe..fa05946 100644
--- a/crates/gc-extract/tests/disc.rs
+++ b/crates/gc-extract/tests/disc.rs
@@ -2920,6 +2920,25 @@ fn chunk_records_start_16_byte_aligned() {
 
     assert!(chunks >= 1500, "only {chunks} chunks parsed");
     assert!(records >= 34000, "only {records} chunk records");
+    // Every chunk has a uniform record length -- length varies between chunks
+    // of a type, never within one -- and on most chunks the leading byte is a
+    // per-chunk phase counter rather than data.
+    let mut uniform = 0usize;
+    let mut ruled = 0usize;
+    for n in 0..27 {
+        let core = std::fs::read(data.join(format!("level{n}/core_data.bin"))).expect("core_data");
+        for c in parse_all_chunks(&core).expect("chunks") {
+            if c.has_uniform_record_length() {
+                uniform += 1;
+            }
+            if c.bucket_rule().is_some() {
+                ruled += 1;
+            }
+        }
+    }
+    assert_eq!(uniform, 1509, "uniform-record-length count moved off 1509");
+    assert_eq!(ruled, 1262, "bucket-rule count moved off 1262");
+
     // Not universal. Three chunk types break even the 8-byte rule, all of them
     // the variable-length or 8-stepped ones. A parallel probe puts 1,142
     // records across seven types outside a *16*-byte rule, which is how the
@@ -2942,8 +2961,9 @@ fn chunk_records_start_16_byte_aligned() {
     );
 
     eprintln!(
-        "chunks: {chunks} chunks, {records} records; leading byte a multiple of 8 on {aligned}, \
-         {bad} exceptions across {} types (8-byte step and one variable-length type)",
+        "chunks: {chunks} chunks, {records} records; uniform record length on {uniform}; \
+         leading byte a multiple of 8 on {aligned}, {bad} exceptions across {} types; \
+         leading byte fully explained by a per-chunk phase counter on {ruled}",
         unaligned.len()
     );
 }
diff --git a/crates/gc-formats/src/chunks.rs b/crates/gc-formats/src/chunks.rs
index 1009427..8030c17 100644
--- a/crates/gc-formats/src/chunks.rs
+++ b/crates/gc-formats/src/chunks.rs
@@ -122,6 +122,65 @@ pub struct ResolvedChunk<'a> {
     pub records: Vec<&'a [u8]>,
 }
 
+impl<'a> ResolvedChunk<'a> {
+    /// The per-chunk phase and step that explain every record's leading byte, or
+    /// `None` when no `(step, phase)` in that form fits.
+    ///
+    /// **Measured: 1,262 of 1,509 chunks** across all 27 levels are fully explained
+    /// by `record[i].leading == step * ((i + phase) mod (256 / step))`, with
+    /// `step` taking 8 on 51 chunks, 16 on 554, 32 on 520 and 64 on 137. The other
+    /// 247 chunks, across 11 chunk types, are not explained by it.
+    ///
+    /// Where this returns `Some`, the leading byte carries **no per-record
+    /// information**: it is a deterministic function of the record's index and the
+    /// chunk's phase, so it is a position marker rather than data. The 247 chunks
+    /// where it returns `None` are the ones where that has *not* been shown, and
+    /// nothing should skip their leading byte on the strength of this.
+    pub fn bucket_rule(&self) -> Option<(u8, usize)> {
+        use std::collections::BTreeSet;
+        if self.records.is_empty() {
+            return None;
+        }
+        let n = self.records.len();
+        for step in [8u16, 16, 32, 64] {
+            let m = (256 / step) as usize;
+            for phase in 0..m {
+                let mut ok = true;
+                for (i, r) in self.records.iter().enumerate() {
+                    let Some(&b) = r.first() else {
+                        ok = false;
+                        break;
+                    };
+                    if b != ((((i + phase) % m) * step as usize % 256) as u8) {
+                        ok = false;
+                        break;
+                    }
+                }
+                if ok {
+                    return Some((step as u8, phase));
+                }
+            }
+        }
+        let _ = n;
+        let _: Option<BTreeSet<u8>> = None;
+        None
+    }
+
+    /// True when every record in this chunk has the same length.
+    ///
+    /// **Measured: true for 1,509 of 1,509 chunks** across all 27 levels. Record
+    /// length varies *between* chunks of the same type -- one type runs 76 to 104
+    /// bytes -- but never within one. An earlier claim that some type had
+    /// variable-length records was pooling across chunks; it is withdrawn.
+    pub fn has_uniform_record_length(&self) -> bool {
+        let mut it = self.records.iter().filter(|r| !r.is_empty());
+        let Some(first) = it.next().map(|r| r.len()) else {
+            return true;
+        };
+        it.all(|r| r.len() == first)
+    }
+}
+
 /// True if `offset` looks like the start of a chunk.
 ///
 /// The signature is the header's last byte being `0xff`, the 10-byte gap
diff --git a/docs/formats/core-data-regions.md b/docs/formats/core-data-regions.md
index cb04bf9..8b5690a 100644
--- a/docs/formats/core-data-regions.md
+++ b/docs/formats/core-data-regions.md
@@ -1316,3 +1316,39 @@ Worth stating plainly: this is the second time this session a correct number
 carried a wrong label, after the tie four-value retraction. Both times the test
 was the thing that caught it, which is the argument for asserting the number in
 a test at all rather than only in prose.
+
+## The chunk record's leading byte is a per-chunk phase counter
+
+Follow-up to the alignment result, and it settles what the leading byte is.
+
+Within one chunk the leading byte steps by a constant and wraps:
+
+    offsets: 1420 1748 2072 2392 2716 3044 3372 3700 4028 4352 ...
+    byte0:      0  16   32   48   64   80   96  112  128  144 ...
+
+i.e. `byte0 == step * ((index + phase) mod (256 / step))`. Sweeping every
+`(step, phase)` over all 27 levels:
+
+    fully explained: 1,262 of 1,509 chunks
+    by step: 8 -> 51 chunks, 16 -> 554, 32 -> 520, 64 -> 137
+    unexplained:  247 chunks across 11 chunk types
+
+**So on 1,262 chunks the leading byte carries no per-record information.** It is
+a deterministic function of the record's index and the chunk's phase. That makes
+it a position marker, not data -- which is why the earlier "16 discrete values
+with uniform counts" pattern showed up.
+
+`ResolvedChunk::bucket_rule` returns the `(step, phase)` when the whole chunk
+fits and `None` otherwise, so the 247 unexplained chunks are distinguishable
+from the 1,262 explained ones and nothing skips a leading byte on the strength of
+a rule that does not apply.
+
+### Every chunk has a uniform record length
+
+**1,509 of 1,509 chunks.** Length varies between chunks of the same type -- one
+type runs 76 to 104 bytes -- but never within one chunk.
+
+This **withdraws the previous commit's "variable-length type"**. That came from
+taking `minlen`/`maxlen` over all of a type's records at once, pooling across
+chunks; rule 3a's pooled claim, third time this session. Within any one chunk the
+length is fixed.