Update training section wording

README.md

@@ -136,9 +136,7 @@ For 8-bit inputs (256 total):
 
 The parametric construction was derived algebraically, not discovered through training.
 
-Evolutionary search was attempted (as with parity) but consistently plateaued at 247/256 (96.5%) accuracy across multiple seeds. The (1,1,-2) weight pattern is sufficiently specific that random mutation cannot reliably discover it.
-
-This finding reinforces that the algebraic insight is essential—MOD-3 networks cannot be found by naive search alone.
+Evolutionary search was attempted (as with parity) but consistently plateaued at 247/256 (96.5%) accuracy across multiple seeds. We were unable to discover the (1,1,-2) weight pattern through training, so we used algebraic construction instead.
 
 ## Comparison to Parity
 
@@ -157,7 +155,7 @@ This finding reinforces that the algebraic insight is essential—MOD-3 networks
 - **Binary inputs**: Expects {0, 1}, not continuous values
 - **No noise margin**: Heaviside threshold at exactly 0
 - **Not differentiable**: Cannot be fine-tuned with gradient descent
-- **Training gap**: Evolutionary search achieves only 96.5%; algebraic construction required for 100%
+- **Training gap**: Our evolutionary search achieved only 96.5%; we used algebraic construction instead
 
 ## Files