From cc71f17b7b311a2619cadb03b6d94698888fb48a Mon Sep 17 00:00:00 2001 From: Meta Vens Date: Tue, 22 Sep 2026 21:43:52 +0200 Subject: [PATCH] Say what the cell-mass test proves, and what it does not Soul: a constant erf passes it. It is one necessary property, not correctness. Co-Authored-By: Claude Opus 5 --- packages/cultmath/tests/CultMath.Tests/MathTests.cs | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/packages/cultmath/tests/CultMath.Tests/MathTests.cs b/packages/cultmath/tests/CultMath.Tests/MathTests.cs index d2e18b1e..b47a389e 100644 --- a/packages/cultmath/tests/CultMath.Tests/MathTests.cs +++ b/packages/cultmath/tests/CultMath.Tests/MathTests.cs @@ -494,9 +494,11 @@ public sealed class MathTests } } - // The consumer-level property that actually matters: fire control turns adjacent erf - // values into a cell's probability mass, so a backward step in erf would show up as a - // negative mass. Soul checked these geometries clean: 16/64/256/1024/4096 cells over + // One property the consumer needs, not the whole of correctness: fire control turns + // adjacent erf values into a cell's probability mass, so a backward step in erf would + // show up as a negative mass. An erf that returned a constant would pass this and is + // caught by ErfMatchesReferenceValues instead. + // Soul checked these geometries clean: 16/64/256/1024/4096 cells over // +-4 sigma, plus a tight 1 sigma spread and a very tight 0.25 sigma spread. [Theory] [InlineData(16, 4f)]