diff --git a/fsst/src/test/java/io/github/dfa1/vortex/fsst/PaperFigure2Test.java b/fsst/src/test/java/io/github/dfa1/vortex/fsst/PaperFigure2Test.java
new file mode 100644
index 00000000..ab90789c
--- /dev/null
+++ b/fsst/src/test/java/io/github/dfa1/vortex/fsst/PaperFigure2Test.java
@@ -0,0 +1,83 @@
+package io.github.dfa1.vortex.fsst;
+
+import org.junit.jupiter.api.Test;
+
+import java.nio.charset.StandardCharsets;
+import java.util.List;
+
+import static org.assertj.core.api.Assertions.assertThat;
+
+/// Golden test reproducing the worked example in Figure 2 of the FSST paper (Boncz, Neumann, Leis,
+/// "FSST: Fast Random Access String Compression", PVLDB 13(11), 2020, p. 2653, Figure 2: "Four
+/// iterations of symbol table construction algorithm on the corpus 'tumcwitumvldb' with a maximum
+/// symbol length of 3 and a maximum symbol table size of 5").
+///
+/// After the paper's fourth iteration the symbol table is `{tum, cwi, vld, b}` and the figure's
+/// "Compressed" row at the bottom shows `tum | cwi | tum | vld | b`, len = 5 (down from the 13-byte
+/// uncompressed corpus). This test hand-builds exactly that symbol table and asserts the greedy
+/// longest-match compressor reproduces the figure's compressed code sequence and length, then that
+/// the result round-trips back to the original 13 bytes.
+///
+/// The symbol table is hand-constructed on purpose. Using [CompressorBuilder#train(byte[][])] would
+/// re-derive a table through this codebase's own training heuristics — a circular check. A golden
+/// test needs an external, human-verifiable fixed point, so the symbols and their codes come
+/// straight off the paper's figure, not off our trainer.
+///
+/// Symbol codes equal the symbol's index in the gain-descending list [Compressor#of(List)] receives.
+/// Figure 2's iteration-4 table lists the symbols in gain order `tum` (gain 6), `cwi` (gain 3),
+/// `vld` (gain 3), `b` (gain 1), so the codes assigned here are `tum`=0, `cwi`=1, `vld`=2, `b`=3.
+class PaperFigure2Test {
+
+    /// The paper's Figure 2 corpus, exactly as printed. ASCII, so US-ASCII and UTF-8 are equivalent.
+    private static final byte[] CORPUS = "tumcwitumvldb".getBytes(StandardCharsets.US_ASCII);
+
+    /// Codes for the figure's compressed row `tum | cwi | tum | vld | b`, using the code assignment
+    /// documented on the class: tum=0, cwi=1, vld=2, b=3.
+    private static final byte[] EXPECTED_CODES = {0, 1, 0, 2, 3};
+
+    @Test
+    void compress_paperFigure2Corpus_producesFigure2CodeSequenceAndLength() {
+        // Given — the exact iteration-4 symbol table from Figure 2, in gain-descending order so
+        // each symbol's code matches its column in the figure (tum=0, cwi=1, vld=2, b=3).
+        Compressor sut = Compressor.of(List.of(
+                symbol("tum"),
+                symbol("cwi"),
+                symbol("vld"),
+                symbol("b")));
+        byte[] out = new byte[CORPUS.length * 2];
+
+        // When — greedy longest-match compression of the 13-byte corpus.
+        long compressedLength = sut.compress(CORPUS, 0, CORPUS.length, out, 0);
+
+        // Then — the figure's "Compressed ... len = 5" row: five codes [tum, cwi, tum, vld, b], no
+        // escapes, exactly the sequence the paper draws.
+        assertThat(compressedLength).isEqualTo(5L);
+        assertThat(java.util.Arrays.copyOf(out, (int) compressedLength)).isEqualTo(EXPECTED_CODES);
+    }
+
+    @Test
+    void compress_thenDecompress_paperFigure2Corpus_roundTripsToOriginal() {
+        // Given — same hand-built Figure 2 table.
+        Compressor sut = Compressor.of(List.of(
+                symbol("tum"),
+                symbol("cwi"),
+                symbol("vld"),
+                symbol("b")));
+        byte[] compressed = new byte[CORPUS.length * 2];
+
+        // When — compress, then decompress through the same table.
+        long compressedLength = sut.compress(CORPUS, 0, CORPUS.length, compressed, 0);
+        byte[] decompressed = new byte[CORPUS.length];
+        long decompressedLength = sut.toDecompressor()
+                .decompress(compressed, 0, (int) compressedLength, decompressed, 0);
+
+        // Then — the primary correctness gate: the exact original 13 bytes come back.
+        assertThat(decompressedLength).isEqualTo(CORPUS.length);
+        assertThat(decompressed).isEqualTo(CORPUS);
+    }
+
+    private static Symbol symbol(String text) {
+        byte[] bytes = text.getBytes(StandardCharsets.US_ASCII);
+        return Symbol.of(bytes, 0, bytes.length);
+    }
+}
diff --git a/writer/src/test/java/io/github/dfa1/vortex/writer/encode/FsstEncodingEncoderTest.java b/writer/src/test/java/io/github/dfa1/vortex/writer/encode/FsstEncodingEncoderTest.java
index 55edd5a2..c57b0785 100644
--- a/writer/src/test/java/io/github/dfa1/vortex/writer/encode/FsstEncodingEncoderTest.java
+++ b/writer/src/test/java/io/github/dfa1/vortex/writer/encode/FsstEncodingEncoderTest.java
@@ -196,9 +196,13 @@ void encode_repeatedText_learnsMultiByteSymbols_compressesBelowRaw() {
             byte maxSymLen = maxSymbolLength(result);
 
             // Then — real compression (< half raw) proves multi-byte symbols are in play, and the
-            // table contains symbols longer than the old fixed 2-byte limit.
+            // table learns symbols well past a bigram. The paper-faithful trainer grows symbols to
+            // the 8-byte cap on this highly repetitive input (ten of eleven learned symbols are
+            // length 8, one is length 3, with the fixed 0x5EED encoder seed), so a `> 4` floor is a
+            // stronger regression guard than the old `> 2` with ample headroom — a training change
+            // that quietly stopped growing symbols past a trigram would now be caught.
             assertThat(compressedBytes).isLessThan(rawBytes / 2);
-            assertThat(maxSymLen).isGreaterThan((byte) 2);
+            assertThat(maxSymLen).isGreaterThan((byte) 4);
         }
 
         @Test