diff --git a/phase-ends/CURRENT_PHASE.md b/phase-ends/CURRENT_PHASE.md
index 055bef8..3b224e9 100644
--- a/phase-ends/CURRENT_PHASE.md
+++ b/phase-ends/CURRENT_PHASE.md
@@ -488,3 +488,25 @@ check a row's final instructions for this pattern BEFORE attempting it and
 record as class-rare-epilogue; the class stays excluded-adjacent (not yet
 excluded from the worklist, since a source spelling or a future post-pass
 may reach it, but recorded).
+
+## Cycle 4 — RARE-EPILOGUE CLASS: consequence corrected (worker B falsified the skip)
+
+The 47-row census stands (definitive pattern count). But worker B falsified
+the coordinator's "skip the class as a block" guidance with a REGISTERED
+match: 0x80021FF4 carries the rare-epilogue signature (restore in the jr
+slot) and cc1 produced it with DEFAULT settings. The class is not
+collectively unmatchable. Worker B's three-way split (verified):
+
+(a) cc1 fills the slot itself → MATCHES normally (proof 0x80021FF4)
+(b) cc1 leaves the slot empty, maspsx's unconditional nop blocks the fill →
+    --no-maspsx fixes the epilogue but breaks move->addu (the harness
+    mutual exclusion reported earlier, NOT the class)
+(c) cc1 emits the nop → genuinely stuck (0x800FBDC0)
+
+REVISED guidance: attempt class rows normally; if the ONLY residual is the
+epilogue, try --no-maspsx; if that trades the epilogue for an or-encoded
+move, record as the harness limitation. Pre-screen only rows in the class
+that ALSO contain a move pseudo — those block on the mutual exclusion.
+LESSON: the coordinator over-generalized a class consequence from a census;
+worker B's empirical falsification is the correct checks-and-balances that
+makes the team structure work.