Replace the EP-family wall-clock figure with the measured backprop ratiomain

The 0.4x figure compared relaxation-iteration counts across methods that count
different things: the closest EP-family baseline uses asynchronous local updates
where one iteration moves information roughly one layer, while ours transmits
through the whole stack per round. Wall-clock against backprop on matched
hardware is the currency that survives that difference, and it measures 2.7x.

1 files changed, 2 insertions, 2 deletions

diff --git a/index.html b/index.html
index 24e2872..e3d0477 100644
--- a/index.html
+++ b/index.html
@@ -1249,8 +1249,8 @@
           <span>matched quality at 72M parameters</span>
         </div>
         <div class="proof">
-          <strong>0.4×</strong>
-          <span>training wall-clock vs the closest EP-family baseline</span>
+          <strong>2.7×</strong>
+          <span>training wall-clock vs backprop, measured on matched hardware</span>
         </div>
       </div>
     </section>