diff --git a/docs/math-validation-plan.md b/docs/math-validation-plan.md
index ceb143c..77fb546 100644
--- a/docs/math-validation-plan.md
+++ b/docs/math-validation-plan.md
@@ -42,9 +42,10 @@ Goal: ensure Echo’s math module produces identical results across environments
 ---
 
 ## Tooling
-- Use Vitest with snapshot-style comparisons for key operations.
+- Rust harness (in `rmg-core/tests/math_validation.rs`) validates scalar/vector/matrix/quaternion + PRNG behaviour against JSON fixtures.
+- Provide deterministic reference values generated offline (e.g., via high-precision Python or Rust) stored in fixtures.
+- Next step: mirror the fixtures in Vitest with snapshot-style comparisons for the TypeScript layer.
 - For cross-environment checks, add Playwright-driven tests that run the same suite in headless Chromium/WebKit (call into math module via bundled script).
-- Provide deterministic reference values generated offline (e.g., via high-precision Python script) stored in fixtures.
 - Fixed-point tests compare against integer expectations.
 
 ---
@@ -57,8 +58,9 @@ Goal: ensure Echo’s math module produces identical results across environments
 ---
 
 ## Tasks
-- [ ] Generate reference fixtures (JSON) for vector/matrix/quaternion cases.
-- [ ] Implement Vitest suites covering scalar, vector, matrix, quaternion, transform, PRNG.
+- [x] Generate reference fixtures (JSON) for scalar/vector/matrix/quaternion/PRNG cases.
+- [x] Implement Rust-based validation suite (`cargo test -p rmg-core --test math_validation`).
+- [ ] Mirror fixtures in Vitest to cover the TypeScript bindings (float32 mode).
 - [ ] Integrate Playwright smoke tests for browser verification.
 - [ ] Add CI job running math tests across environments.
 - [ ] Document any environment-specific deviations in decision log.
diff --git a/docs/spec-deterministic-math.md b/docs/spec-deterministic-math.md
index 5dcde68..0a2a361 100644
--- a/docs/spec-deterministic-math.md
+++ b/docs/spec-deterministic-math.md
@@ -16,23 +16,26 @@ Echo’s math module underpins every deterministic system: physics proxies, anim
 ## Numeric Modes
 
 ### Float32 Mode (default)
+
 - All operations clamp to IEEE 754 float32 using `Math.fround`.
 - Inputs converted to float32 before computation; outputs stored in float32 buffers (`Float32Array`).
 - Stable across JS engines as long as `Math.fround` available (polyfill for older runtimes).
 
 ### Fixed-Point Mode (opt-in)
+
 - 32.32 fixed-point representation using BigInt internally, surfaced as wrapper `Fixed` type.
 - Configured via engine options (`mathMode: "float32" | "fixed32"`).
 - Useful for deterministic networking or hardware without stable float operations.
 - Bridges through helper functions: `fixed.fromFloat`, `fixed.toFloat`, `fixed.mul`, `fixed.div`.
 
-Mode chosen at engine init; math module provides factory returning mode-specific implementations.
+Mode chosen at engine init; math module provides factory returning mode-specific implementations. The Rust runtime already exposes the float32 primitives in `rmg_core::math`, so FFI/WASM adapters can reuse a single source of truth while TypeScript bindings converge on the same fixtures.
 
 ---
 
 ## Core Types
 
 ### Vec2 / Vec3 / Vec4
+
 ```ts
 interface Vec2 {
   readonly x: number;
@@ -45,22 +48,30 @@ type VecLike = Float32Array | number[];
 - Methods: `create`, `clone`, `set`, `add`, `sub`, `scale`, `dot`, `length`, `normalize`, `lerp`, `equals`.
 - All mutating functions accept `out` parameter for in-place updates to reduce allocations.
 - Deterministic clamps: every operation ends with `fround` (float mode) or `fixed` operations.
+- Rust parity: `rmg_core::math::Vec3` currently implements add/sub/scale/dot/cross/length/normalize; `Vec2`/`Vec4` remain TODO.
 
 ### Mat3 / Mat4
+
 - Column-major storage (`Float32Array(9)` / `Float32Array(16)`).
 - Methods: `identity`, `fromRotation`, `fromTranslation`, `multiply`, `invert`, `transformVec`.
 - Deterministic inversion: use well-defined algorithm with guard against singular matrices (records failure and returns identity or throws based on config).
+- Rust parity: `rmg_core::math::Mat4` exposes `multiply` and `transform_point`; identity/fromRotation/invert are pending.
 
 ### Quat
+
 - Represented as `[x, y, z, w]`.
 - Functions: `identity`, `fromAxisAngle`, `multiply`, `slerp`, `normalize`, `toMat4`.
 - `slerp` uses deterministic interpolation with clamped range.
+- Rust parity: `rmg_core::math::Quat` implements identity/fromAxisAngle/multiply/normalize/to_mat4; `slerp` remains TBD.
 
 ### Transform
+
 - Struct bundling position (Vec3), rotation (Quat), scale (Vec3).
 - Helper for constructing Mat4; ensures consistent order of operations.
+- Rust parity: transform helpers are still tracked for Phase 1 (not implemented yet).
 
 ### Bounds / AABB
+
 - Useful for physics collision; stores min/max Vec3.
 - Provides deterministic union/intersection operations.
 
@@ -69,6 +80,7 @@ type VecLike = Float32Array | number[];
 ## PRNG Services
 
 ### Engine PRNG
+
 - Based on counter-based generator (e.g., Philox or Xoroshiro128+).
 - Implementation in TypeScript with optional WebAssembly acceleration later.
 - Interface:
@@ -84,12 +96,15 @@ interface PRNG {
 - `state` serializable for replay.
 - `jump` used for branch forking: clone generator with deterministic offset.
 - `seed` derived from combination of world seed + branch ID + optional subsystem tag.
+- Rust parity: `rmg_core::math::Prng` implements seeding, `next_f32`, and `next_int`; state/jump APIs are follow-up work.
 
 ### Deterministic Hashing
+
 - Provide `hash64` function (e.g., SplitMix64) for converting strings/IDs into seeds.
 - Ensure stable across platforms; implement in TypeScript to avoid native differences.
 
 ### Integration Points
+
 - Scheduler passes `math.prng` on `TickContext`.
 - Codex’s Baby `CommandContext` exposes `prng.spawn(scope)` for per-handler streams.
 - Timeline branch creation clones PRNG state to maintain deterministic divergence.
@@ -97,6 +112,7 @@ interface PRNG {
 ---
 
 ## Utility Functions
+
 - `clamp(value, min, max)` – deterministic clamp using `Math.min/Math.max` once (avoid multiple rounding).
 - `approximatelyEqual(a, b, epsilon)` – uses configured epsilon (float32 ~1e-6).
 - `degToRad`, `radToDeg` – using float32 rounding.