diff --git a/internal/engine/rng.go b/internal/engine/rng.go
new file mode 100644
index 0000000..009148c
--- /dev/null
+++ b/internal/engine/rng.go
@@ -0,0 +1,16 @@
+// Package engine implements the agent-based spread model: build a social
+// network, run an independent cascade over it, and measure the effect of
+// education strategies. It is pure: no I/O, no web dependencies, and all
+// randomness flows from seeds in the config, so identical inputs always
+// produce identical runs.
+package engine
+
+import "math/rand/v2"
+
+// newRand returns a deterministic random stream for the given seed. The
+// engine never touches global random state; every source of randomness
+// (graph build, edge thresholds, education sampling) gets its own seeded
+// stream so each can be varied independently.
+func newRand(seed uint64) *rand.Rand {
+	return rand.New(rand.NewPCG(seed, 0))
+}
diff --git a/internal/engine/rng_test.go b/internal/engine/rng_test.go
new file mode 100644
index 0000000..0ba8de4
--- /dev/null
+++ b/internal/engine/rng_test.go
@@ -0,0 +1,26 @@
+package engine
+
+import "testing"
+
+// Determinism is a locked project decision: the same seed must always yield
+// the same stream, and different seeds must not collide. With fixed seeds
+// these assertions are exact, not probabilistic.
+
+func TestNewRandSameSeedSameStream(t *testing.T) {
+	a, b := newRand(17), newRand(17)
+	for i := range 100 {
+		if got, want := a.Float64(), b.Float64(); got != want {
+			t.Fatalf("draw %d: streams diverged: %v != %v", i, got, want)
+		}
+	}
+}
+
+func TestNewRandDifferentSeedsDiffer(t *testing.T) {
+	a, b := newRand(17), newRand(18)
+	for range 100 {
+		if a.Float64() != b.Float64() {
+			return // diverged, as expected
+		}
+	}
+	t.Fatal("seeds 17 and 18 produced 100 identical draws")
+}