diff --git a/reikhelm-core/src/lib.rs b/reikhelm-core/src/lib.rs
index a4dfaae..4e636ef 100644
--- a/reikhelm-core/src/lib.rs
+++ b/reikhelm-core/src/lib.rs
@@ -5,3 +5,4 @@
 
 pub mod geometry;
 pub mod grid;
+pub mod rng;
diff --git a/reikhelm-core/src/rng.rs b/reikhelm-core/src/rng.rs
new file mode 100644
index 0000000..8924811
--- /dev/null
+++ b/reikhelm-core/src/rng.rs
@@ -0,0 +1,347 @@
+//! Deterministic, reproducible RNG (spec §4.4, §7).
+//!
+//! [`Rng`] wraps [`rand_chacha::ChaCha8Rng`] — a portable, byte-stable stream
+//! cipher RNG. The same seed yields the same sequence on every machine and
+//! every run; we never touch `thread_rng`/`rand::random()`.
+//!
+//! The keystone feature is [`Rng::fork`]: it derives an *independent* child
+//! stream from the parent's **basis seed** plus a label. Forking does not draw
+//! from or otherwise mutate the parent (`fork(&self)`), so a child stream is a
+//! pure function of `(basis_seed, label)` — independent of how many values were
+//! pulled from the parent first. The [`Pipeline`](crate) leans on this to give
+//! each pass its own sub-stream keyed by `"<pass name>#<occurrence>"`, which is
+//! what makes inserting or reordering a pass non-destructive to existing seeds.
+
+use rand_chacha::ChaCha8Rng;
+// `rand_core` is not a direct dependency; reach its traits through the
+// re-export on `rand_chacha` so the two stay version-locked. `Rng` carries
+// `next_u64`; `SeedableRng` carries `seed_from_u64`.
+use rand_chacha::rand_core::{Rng as _, SeedableRng};
+
+/// A deterministic random number generator backed by ChaCha8.
+///
+/// Construct one with [`Rng::from_seed`], draw values with [`range`](Rng::range),
+/// [`chance`](Rng::chance), [`choose`](Rng::choose) and [`shuffle`](Rng::shuffle),
+/// and derive isolated child streams with [`fork`](Rng::fork).
+#[derive(Clone, Debug)]
+pub struct Rng {
+    /// The stable basis used for all forks: the seed this `Rng` was built from.
+    /// Forking mixes *this* (never the live ChaCha state) with the label, so a
+    /// fork is independent of how many draws the parent has taken.
+    basis_seed: u64,
+    /// The live ChaCha8 stream that actually produces random words.
+    inner: ChaCha8Rng,
+}
+
+impl Rng {
+    /// Create an `Rng` from a 64-bit seed, storing the seed as the fork basis.
+    ///
+    /// The seed is expanded into ChaCha8's 256-bit key by ChaCha8's own
+    /// portable `seed_from_u64`, so two `Rng`s built from the same `u64`
+    /// produce identical streams everywhere.
+    pub fn from_seed(seed: u64) -> Self {
+        Self {
+            basis_seed: seed,
+            inner: ChaCha8Rng::seed_from_u64(seed),
+        }
+    }
+
+    /// Derive an independent child `Rng` from a stable basis (this `Rng`'s
+    /// originating seed) combined with `label`.
+    ///
+    /// This **does not** advance or otherwise mutate `self` (note `&self`), and
+    /// the result depends only on `(self.basis_seed, label)` — never on how many
+    /// values have already been drawn from `self`. Distinct labels (including
+    /// `"name#0"` vs `"name#1"`) yield distinct streams.
+    ///
+    /// The label→seed mix is the explicit, fixed integer routine in
+    /// [`mix_seed`] (splitmix64 + FNV-1a), chosen because it is stable across
+    /// runs and crate versions. We deliberately avoid
+    /// `std::collections::hash_map::DefaultHasher`, whose output is
+    /// unspecified and version-unstable (spec §7).
+    pub fn fork(&self, label: &str) -> Rng {
+        Rng::from_seed(mix_seed(self.basis_seed, label))
+    }
+
+    /// Draw a `u64` of raw randomness from the live stream.
+    fn next_u64(&mut self) -> u64 {
+        self.inner.next_u64()
+    }
+
+    /// Return an integer in the **half-open** range `[lo, hi)` — `lo` inclusive,
+    /// `hi` exclusive (so `range(0, 3)` yields one of `0, 1, 2`).
+    ///
+    /// If `hi <= lo` the range is empty; we return `lo` rather than panicking
+    /// (no panics on valid input — spec §8). Uniformity is achieved with
+    /// Lemire-style rejection sampling over the `u64` span, so the result is
+    /// unbiased across the whole range.
+    pub fn range(&mut self, lo: i32, hi: i32) -> i32 {
+        if hi <= lo {
+            return lo;
+        }
+        // Width fits in u64: (hi - lo) where hi > lo and both are i32.
+        let span = (hi as i64 - lo as i64) as u64;
+        lo + self.below(span) as i32
+    }
+
+    /// Uniform integer in `[0, span)` for `span > 0`, via rejection sampling to
+    /// avoid modulo bias. `span` is assumed non-zero by callers.
+    fn below(&mut self, span: u64) -> u64 {
+        // Largest multiple of `span` that fits in u64; anything at or above the
+        // threshold would skew the distribution, so we reject and redraw.
+        let threshold = u64::MAX - (u64::MAX % span);
+        loop {
+            let x = self.next_u64();
+            if x < threshold {
+                return x % span;
+            }
+        }
+    }
+
+    /// Return `true` with probability `p`.
+    ///
+    /// `p` is clamped to `[0.0, 1.0]`, so out-of-range probabilities are
+    /// saturated rather than panicking. `p <= 0.0` is always `false`,
+    /// `p >= 1.0` is always `true`.
+    pub fn chance(&mut self, p: f64) -> bool {
+        if p <= 0.0 {
+            return false;
+        }
+        if p >= 1.0 {
+            return true;
+        }
+        // Map the next u64 into a uniform f64 in [0, 1) and compare.
+        // 53 bits gives the full mantissa precision of an f64.
+        let bits = self.next_u64() >> 11; // top 53 bits
+        let unit = bits as f64 / (1u64 << 53) as f64;
+        unit < p
+    }
+
+    /// Return a reference to a uniformly chosen element of `items`, or `None`
+    /// if the slice is empty.
+    pub fn choose<'a, T>(&mut self, items: &'a [T]) -> Option<&'a T> {
+        if items.is_empty() {
+            return None;
+        }
+        let idx = self.below(items.len() as u64) as usize;
+        items.get(idx)
+    }
+
+    /// Shuffle `items` in place into a uniformly random permutation.
+    ///
+    /// Uses an in-place Fisher–Yates shuffle built on our own [`range`] so the
+    /// permutation is fully determined by the stream and identical on every
+    /// machine. A slice of length 0 or 1 is left unchanged.
+    pub fn shuffle<T>(&mut self, items: &mut [T]) {
+        let len = items.len();
+        if len < 2 {
+            return;
+        }
+        // Walk from the end; swap each element with a random earlier-or-equal
+        // index. Iterate down to 1 because index 0 has nothing left to swap.
+        for i in (1..len).rev() {
+            let j = self.below((i + 1) as u64) as usize;
+            items.swap(i, j);
+        }
+    }
+}
+
+/// Mix a 64-bit basis seed with a label into a fresh 64-bit seed for [`Rng::fork`].
+///
+/// This is the load-bearing determinism primitive (spec §7), so the algorithm
+/// is spelled out explicitly and uses only fixed integer arithmetic — no
+/// `DefaultHasher`, no version-unstable hashing.
+///
+/// Recipe:
+/// 1. Start from a splitmix64 scramble of the basis seed (so the basis is well
+///    diffused before any label bytes are folded in).
+/// 2. Fold each label byte in with the FNV-1a step (`xor` then multiply by the
+///    64-bit FNV prime), which makes the result order-sensitive across bytes —
+///    so `"name#0"` and `"name#1"` diverge.
+/// 3. Run one final splitmix64 finalizer to avalanche the combined value.
+fn mix_seed(basis: u64, label: &str) -> u64 {
+    // The 64-bit FNV-1a offset basis and prime (fixed, well-known constants).
+    const FNV_PRIME: u64 = 0x0000_0100_0000_01B3;
+
+    let mut state = splitmix64(basis);
+    for &byte in label.as_bytes() {
+        state ^= byte as u64;
+        state = state.wrapping_mul(FNV_PRIME);
+    }
+    splitmix64(state)
+}
+
+/// One round of the splitmix64 finalizer: a fixed, portable integer avalanche.
+fn splitmix64(seed: u64) -> u64 {
+    // Constants from the reference splitmix64 (Steele, Lea & Flood 2014).
+    let mut z = seed.wrapping_add(0x9E37_79B9_7F4A_7C15);
+    z = (z ^ (z >> 30)).wrapping_mul(0xBF58_476D_1CE4_E5B9);
+    z = (z ^ (z >> 27)).wrapping_mul(0x94D0_49BB_1331_11EB);
+    z ^ (z >> 31)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// Same seed → identical sequence of `range` draws.
+    #[test]
+    fn same_seed_reproduces_sequence() {
+        let mut a = Rng::from_seed(0xABCD_1234);
+        let mut b = Rng::from_seed(0xABCD_1234);
+        let seq_a: Vec<i32> = (0..64).map(|_| a.range(0, 1000)).collect();
+        let seq_b: Vec<i32> = (0..64).map(|_| b.range(0, 1000)).collect();
+        assert_eq!(seq_a, seq_b);
+    }
+
+    /// Different seeds should (overwhelmingly) produce different sequences.
+    #[test]
+    fn distinct_seeds_diverge() {
+        let mut a = Rng::from_seed(1);
+        let mut b = Rng::from_seed(2);
+        let seq_a: Vec<i32> = (0..64).map(|_| a.range(0, 1_000_000)).collect();
+        let seq_b: Vec<i32> = (0..64).map(|_| b.range(0, 1_000_000)).collect();
+        assert_ne!(seq_a, seq_b);
+    }
+
+    /// `fork` is order-independent: the child stream does not depend on how many
+    /// values were drawn from the parent beforehand, and forking does not mutate
+    /// the parent.
+    #[test]
+    fn fork_is_order_independent() {
+        let parent = Rng::from_seed(99);
+
+        // Fork immediately.
+        let mut early = parent.fork("a");
+
+        // Now draw a bunch from a *clone* of the parent, then fork.
+        let mut drained = parent.clone();
+        for _ in 0..50 {
+            let _ = drained.range(0, 100);
+        }
+        let mut late = drained.fork("a");
+
+        let early_seq: Vec<i32> = (0..32).map(|_| early.range(0, 10_000)).collect();
+        let late_seq: Vec<i32> = (0..32).map(|_| late.range(0, 10_000)).collect();
+        assert_eq!(early_seq, late_seq, "fork must depend only on basis seed + label");
+    }
+
+    /// `fork(&self)` does not advance the parent's own stream.
+    #[test]
+    fn fork_does_not_mutate_parent() {
+        let mut parent = Rng::from_seed(7);
+        let before: Vec<i32> = {
+            let mut probe = parent.clone();
+            (0..16).map(|_| probe.range(0, 1000)).collect()
+        };
+        // Forking should leave `parent`'s stream untouched.
+        let _child = parent.fork("whatever");
+        let after: Vec<i32> = (0..16).map(|_| parent.range(0, 1000)).collect();
+        assert_eq!(before, after);
+    }
+
+    /// Distinct labels produce distinct streams.
+    #[test]
+    fn distinct_labels_diverge() {
+        let parent = Rng::from_seed(42);
+        let mut a = parent.fork("a");
+        let mut b = parent.fork("b");
+        let seq_a: Vec<i32> = (0..32).map(|_| a.range(0, 1_000_000)).collect();
+        let seq_b: Vec<i32> = (0..32).map(|_| b.range(0, 1_000_000)).collect();
+        assert_ne!(seq_a, seq_b);
+    }
+
+    /// The pipeline's occurrence-keying scheme works: `"name#0"` and `"name#1"`
+    /// must yield distinct streams.
+    #[test]
+    fn occurrence_keys_diverge() {
+        let parent = Rng::from_seed(2024);
+        let mut zero = parent.fork("name#0");
+        let mut one = parent.fork("name#1");
+        let seq_zero: Vec<i32> = (0..32).map(|_| zero.range(0, 1_000_000)).collect();
+        let seq_one: Vec<i32> = (0..32).map(|_| one.range(0, 1_000_000)).collect();
+        assert_ne!(seq_zero, seq_one);
+    }
+
+    /// `choose` on an empty slice returns `None`; on a non-empty slice it
+    /// returns an in-bounds element.
+    #[test]
+    fn choose_empty_and_nonempty() {
+        let mut rng = Rng::from_seed(5);
+        let empty: [i32; 0] = [];
+        assert_eq!(rng.choose(&empty), None);
+
+        let items = [10, 20, 30, 40];
+        for _ in 0..64 {
+            let picked = rng.choose(&items).copied();
+            assert!(picked.map(|v| items.contains(&v)).unwrap_or(false));
+        }
+    }
+
+    /// `shuffle` is a permutation: it preserves the multiset of elements.
+    #[test]
+    fn shuffle_is_a_permutation() {
+        let mut rng = Rng::from_seed(123);
+        let mut items: Vec<i32> = (0..50).collect();
+        let original = items.clone();
+        rng.shuffle(&mut items);
+
+        // Same multiset.
+        let mut sorted = items.clone();
+        sorted.sort_unstable();
+        assert_eq!(sorted, original);
+
+        // And the same shuffle reproduces under the same seed.
+        let mut rng2 = Rng::from_seed(123);
+        let mut items2: Vec<i32> = (0..50).collect();
+        rng2.shuffle(&mut items2);
+        assert_eq!(items, items2);
+    }
+
+    /// Tiny / empty slices are left untouched by `shuffle`.
+    #[test]
+    fn shuffle_trivial_slices() {
+        let mut rng = Rng::from_seed(1);
+        let mut empty: Vec<i32> = vec![];
+        rng.shuffle(&mut empty);
+        assert!(empty.is_empty());
+
+        let mut single = vec![42];
+        rng.shuffle(&mut single);
+        assert_eq!(single, vec![42]);
+    }
+
+    /// `range` honors the half-open convention and the degenerate `hi <= lo`
+    /// contract (returns `lo`, never panics).
+    #[test]
+    fn range_bounds_are_half_open() {
+        let mut rng = Rng::from_seed(77);
+        for _ in 0..256 {
+            let v = rng.range(3, 7);
+            assert!((3..7).contains(&v), "expected [3,7), got {v}");
+        }
+        // Degenerate ranges return `lo`.
+        assert_eq!(rng.range(5, 5), 5);
+        assert_eq!(rng.range(5, 2), 5);
+        // Negative bounds work too.
+        for _ in 0..256 {
+            let v = rng.range(-5, -1);
+            assert!((-5..-1).contains(&v), "expected [-5,-1), got {v}");
+        }
+    }
+
+    /// `chance` saturates outside `[0, 1]` and is reproducible.
+    #[test]
+    fn chance_saturates_and_reproduces() {
+        let mut rng = Rng::from_seed(9);
+        assert!(!rng.chance(0.0));
+        assert!(!rng.chance(-1.0));
+        assert!(rng.chance(1.0));
+        assert!(rng.chance(2.0));
+
+        // p = 0.5 over many draws should land somewhere in the middle.
+        let mut a = Rng::from_seed(2);
+        let hits = (0..10_000).filter(|_| a.chance(0.5)).count();
+        assert!((4000..6000).contains(&hits), "0.5 chance produced {hits}/10000");
+    }
+}