diff --git a/CHANGELOG.md b/CHANGELOG.md
new file mode 100644
index 0000000..935b26f
--- /dev/null
+++ b/CHANGELOG.md
@@ -0,0 +1,51 @@
+Change log for muTopOpt
+=======================
+
+v1.0.0 (16Sep26)
+----------------
+
+First release. muTopOpt does FFT-accelerated finite-element topology
+optimization of periodic metamaterials: it designs a unit cell whose
+homogenized stiffness (or conductivity) matches a prescribed target, by the
+method of Jödicke et al., *Topology optimization of metamaterials with
+FFT-accelerated micromechanical solvers*.
+
+What it does
+
+- Stress-matching objective against a target effective stiffness, given either
+  as bulk and shear moduli or as Young's modulus and Poisson's ratio, with
+  phase-field regularization and no explicit volume constraint. A conductivity
+  analogue (`FluxTargetProblem`) shares the same machinery
+- Exact sensitivities by the discrete adjoint method. The finite-difference
+  gradient check in `test/test_gradient.py` is the correctness gate for the
+  whole pipeline and runs serially and under MPI
+- Dimension-agnostic: the same code paths run 2D (3 load cases) and 3D (6)
+- Two density discretizations: element-wise (per-pixel, FD-Laplacian penalty)
+  and nodal finite-element (element-consistent H¹ seminorm), the latter acting
+  as an implicit sensitivity filter so the optimizer can merge or dissolve
+  features instead of locking in the initial topology
+- Two outer optimizers, both MPI-distributed through NuMPI: a bound-constrained
+  L-BFGS and a trust-region Newton-CG with exact Hessian-vector products from
+  the second-order adjoint. The trust region is the default where available,
+  since its acceptance test compares against a computable predicted reduction
+  and so cannot drown in inner-solve noise the way a line search does
+- Adaptive inner CG tolerance coupled to the outer optimizer, and
+  precision-aware tolerance defaults
+- Restart from a previous run's output, Fourier-resampled if the grids differ
+- NetCDF output, flushed per frame, carrying the full invocation and the
+  versions that produced it
+
+Scale
+
+No stiffness tensor and no strain or stress field is ever stored: the operator,
+the preconditioner and the sensitivity are all matrix-free and fused, and all
+fields share one ghosted, MPI-decomposed, optionally device-resident layout.
+A 512³ design in single precision fits on a single 128 GB unified-memory
+accelerator -- measured above 60 GiB resident during the first L-BFGS
+iterations of an MI300A run -- and runs entirely on device.
+
+Requirements
+
+`muGrid` provides the FFT engine, the domain decomposition, the fused operators
+and the preconditioners. This release needs a muGrid that provides
+`NodalMomentOperator` (see the pin in `pyproject.toml`).
diff --git a/muTopOpt/__init__.py b/muTopOpt/__init__.py
index 4da0db0..abf45d3 100644
--- a/muTopOpt/__init__.py
+++ b/muTopOpt/__init__.py
@@ -41,7 +41,7 @@
     rho, info = optimize_bounded_lbfgs(problem, initial_density(homog.nb_pixels))
 """
 
-__version__ = "0.0.1"
+__version__ = "1.0.0"
 
 from .conduction import HomogenizationConductivity, SimpConductivity
 from .conduction_problem import FluxLoadCase, FluxTargetProblem
diff --git a/pyproject.toml b/pyproject.toml
index 71a2532..c7dd79b 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -17,6 +17,11 @@ dependencies = [
     "numpy",
     # >= 0.112.0 for the dtype-threaded preconditioner factories, which let a
     # float32 Homogenization run the whole solve in single precision.
+    #
+    # NOTE: the consistent nodal double well (muTopOpt.nodal.ConsistentDoubleWell)
+    # calls muGrid's NodalMomentOperator, which is newer than this pin. Raise the
+    # pin to the first muGrid release that ships it before tagging v1.0.0 --
+    # muGrid 1.2.0 does not have it.
     "muGrid>=0.112.0",
     "NuMPI",
 ]