diff --git a/README.md b/README.md
index 1653e61..49fbce0 100644
--- a/README.md
+++ b/README.md
@@ -242,7 +242,7 @@ let current = bestBid.load(ordering: .acquiring)
 let nan = FixedPointDecimal.nan
 nan.isNaN                    // true
 nan == nan                   // true (sentinel semantics)
-(nan + someValue).isNaN      // true (propagates)
+nan + someValue              // traps (NaN is signalling)
 nan.description              // "nan"
 ```
 
@@ -301,7 +301,7 @@ bash Fuzz/run.sh debug run
 
 The fuzzer validates invariants across all operations:
 
-- **Arithmetic**: commutativity, NaN propagation, no silent NaN sentinel creation
+- **Arithmetic**: commutativity, NaN trapping, no silent NaN sentinel creation
 - **Comparisons**: strict total order (exactly one of `<`, `==`, `>`)
 - **Conversions**: String, Double, Decimal, Codable round-trips
 - **Rounding**: scale-8 identity, no overflow
diff --git a/Sources/FixedPointDecimal/FixedPointDecimal.docc/GettingStarted.md b/Sources/FixedPointDecimal/FixedPointDecimal.docc/GettingStarted.md
index 69570f0..32cd4a9 100644
--- a/Sources/FixedPointDecimal/FixedPointDecimal.docc/GettingStarted.md
+++ b/Sources/FixedPointDecimal/FixedPointDecimal.docc/GettingStarted.md
@@ -97,13 +97,13 @@ let i = Int(somePrice)                    // 123
 
 ## NaN Support
 
-A sentinel-based NaN value propagates through all operations:
+NaN is signalling — any arithmetic operation on a NaN value traps immediately:
 
 ```swift
 let missing = FixedPointDecimal.nan
 missing.isNaN                              // true
-(missing + someValue).isNaN                // true (propagates)
 missing == .nan                            // true (sentinel semantics)
+missing + someValue                        // traps (precondition failure)
 ```
 
 ## Codable
diff --git a/Sources/FixedPointDecimal/FixedPointDecimal.docc/TypeDesign.md b/Sources/FixedPointDecimal/FixedPointDecimal.docc/TypeDesign.md
index 0050467..65935af 100644
--- a/Sources/FixedPointDecimal/FixedPointDecimal.docc/TypeDesign.md
+++ b/Sources/FixedPointDecimal/FixedPointDecimal.docc/TypeDesign.md
@@ -31,7 +31,7 @@ This value was chosen because:
 NaN uses **sentinel semantics**, not IEEE 754:
 - `NaN == NaN` returns `true` (required for `Hashable` and `Comparable` correctness)
 - NaN compares less than all non-NaN values (provides a strict total order for sorting)
-- Arithmetic with NaN propagates NaN (any operation involving NaN returns NaN)
+- Arithmetic with NaN traps (NaN is signalling — any operation involving NaN is a precondition failure)
 - `init(_ decimal: Decimal)` maps `Decimal.nan` to `.nan`; `Decimal(.nan)` returns `Decimal.nan`
 - `isFinite` returns `false` for NaN, `true` for all other values
 - `sign` returns `.plus` for NaN (NaN is excluded from the negative check)
diff --git a/Sources/FixedPointDecimal/FixedPointDecimal.swift b/Sources/FixedPointDecimal/FixedPointDecimal.swift
index 21db149..3acf2cf 100644
--- a/Sources/FixedPointDecimal/FixedPointDecimal.swift
+++ b/Sources/FixedPointDecimal/FixedPointDecimal.swift
@@ -393,7 +393,7 @@ public struct FixedPointDecimal: Sendable, BitwiseCopyable {
     /// Follows the swift-numerics `ElementaryFunctions` naming convention.
     ///
     /// Negative exponents compute the reciprocal: `pow(x, -n) == 1 / pow(x, n)`.
-    /// Returns `.nan` if the base is NaN, or if the result overflows Int64 storage.
+    /// - Precondition: The base must not be NaN. Traps on overflow.
     ///
     /// ```swift
     /// FixedPointDecimal.pow(10, 3)     // 1000