diff --git a/pharmsol-dsl/src/analyze.rs b/pharmsol-dsl/src/analyze.rs
index 034708d7..2750efd5 100644
--- a/pharmsol-dsl/src/analyze.rs
+++ b/pharmsol-dsl/src/analyze.rs
@@ -916,7 +916,11 @@ impl<'a> Analyzer<'a> {
             let id = self.insert_global_symbol(
                 &ident.text,
                 kind,
-                PendingSymbolType::Scalar(None),
+                // Derived values and outputs are stored in f64 buffers, so they
+                // are always real-valued. Letting a whole-number initialiser
+                // narrow one to Int writes an f64 and reads it back as an
+                // integer, which silently returns the bit pattern.
+                PendingSymbolType::Scalar(Some(ValueType::Real)),
                 ident.span,
             )?;
             match kind {
diff --git a/tests/dsl_constant_derived.rs b/tests/dsl_constant_derived.rs
new file mode 100644
index 00000000..295ea377
--- /dev/null
+++ b/tests/dsl_constant_derived.rs
@@ -0,0 +1,128 @@
+//! A derived assignment whose value folds to a compile-time constant must still
+//! be materialised, because other statements read it by slot.
+
+#![cfg(feature = "dsl")]
+
+use pharmsol::dsl::compile_module_source_to_runtime;
+use pharmsol::{prelude::*, Parameters};
+
+fn prediction(
+    model_source: &str,
+    params: &[(&str, f64)],
+) -> Result<f64, Box<dyn std::error::Error>> {
+    let model = compile_module_source_to_runtime(model_source, Some("const_derived"), |_, _| {})?;
+    let support_point = Parameters::with_model(&model, params.iter().copied())?;
+    let subject = Subject::builder("const_derived")
+        .missing_observation(0.0, "cp")
+        .build();
+    let predictions = model
+        .estimate_predictions(&subject, &support_point)?
+        .into_subject()
+        .ok_or("expected subject predictions")?;
+    Ok(predictions.predictions()[0].prediction())
+}
+
+const CONSTANT_DERIVED: &str = r#"
+name = const_derived
+kind = ode
+
+params = p1, p2
+states = central, sink
+outputs = cp
+
+z1 = 1.0
+
+init(central) = 2.0
+init(sink) = 0.0
+
+dx(central) = 0.0
+dx(sink) = 0.0
+
+out(cp) = z1
+"#;
+
+const FRACTIONAL_DERIVED: &str = r#"
+name = const_derived
+kind = ode
+
+params = p1, p2
+states = central, sink
+outputs = cp
+
+z1 = 1.5
+
+init(central) = 2.0
+init(sink) = 0.0
+
+dx(central) = 0.0
+dx(sink) = 0.0
+
+out(cp) = z1
+"#;
+
+const PARAMETER_DERIVED: &str = r#"
+name = const_derived
+kind = ode
+
+params = p1, p2
+states = central, sink
+outputs = cp
+
+z1 = p1
+
+init(central) = 2.0
+init(sink) = 0.0
+
+dx(central) = 0.0
+dx(sink) = 0.0
+
+out(cp) = z1
+"#;
+
+const MIN_MAX_CONSTANT_BRANCHES: &str = r#"
+name = const_derived
+kind = ode
+
+params = p1, p2
+states = central, sink
+outputs = cp
+
+z1 = if (p1 > 0.0) 1.0 else 2.0
+
+init(central) = 2.0
+init(sink) = 0.0
+
+dx(central) = 0.0
+dx(sink) = 0.0
+
+out(cp) = z1
+"#;
+
+#[test]
+fn derived_value_from_a_parameter_is_read_correctly() -> Result<(), Box<dyn std::error::Error>> {
+    let value = prediction(PARAMETER_DERIVED, &[("p1", 4.25), ("p2", 1.0)])?;
+    assert!((value - 4.25).abs() < 1e-9, "got {value}");
+    Ok(())
+}
+
+#[test]
+fn derived_value_from_a_constant_is_read_correctly() -> Result<(), Box<dyn std::error::Error>> {
+    let value = prediction(CONSTANT_DERIVED, &[("p1", 4.25), ("p2", 1.0)])?;
+    assert!((value - 1.0).abs() < 1e-9, "got {value}");
+    Ok(())
+}
+
+#[test]
+fn derived_value_from_a_fractional_constant_is_read_correctly(
+) -> Result<(), Box<dyn std::error::Error>> {
+    let value = prediction(FRACTIONAL_DERIVED, &[("p1", 4.25), ("p2", 1.0)])?;
+    assert!((value - 1.5).abs() < 1e-9, "got {value}");
+    Ok(())
+}
+
+#[test]
+fn constant_branch_conditional_is_read_correctly() -> Result<(), Box<dyn std::error::Error>> {
+    let value = prediction(MIN_MAX_CONSTANT_BRANCHES, &[("p1", 1.0), ("p2", 1.0)])?;
+    assert!((value - 1.0).abs() < 1e-9, "got {value}");
+    Ok(())
+}