diff --git a/src/assemblers/microkernel.jl b/src/assemblers/microkernel.jl
new file mode 100644
index 0000000..d4746f8
--- /dev/null
+++ b/src/assemblers/microkernel.jl
@@ -0,0 +1,144 @@
+# This file is a part of JuliaFEM.
+# License is MIT: see https://github.com/JuliaFEM/JuliaFEM.jl/blob/master/LICENSE.md
+
+#=
+Microkernel contract for the DOF-based assembler.
+
+The DOF-based assembler walks one DOF row at a time and asks the kernel
+for a single scalar `K[i, j]`. To keep the assembler kernel-agnostic and
+zero-allocation, every kernel must opt in by implementing three pieces:
+
+1. `qpoint_buffer_eltype(kernel)` — what type of value the kernel needs
+   stored once per quadrature point per element. For continuum mechanics
+   this is the elasticity tensor; for heat conduction it would be the
+   conductivity tensor; for beams a stiffness scalar; etc.
+
+2. `update_qpoint_buffer!(buffer, material_workspace, kernel)` — fill
+   that buffer once per element from the per-element material workspace.
+   This is called in Pass 1 of `assemble!` (element loop) so it must be
+   allocation-free. Piecewise-element scalars (e.g. `ElementWiseScalarDiffusion`)
+   use an extra `eid` argument only on the internal
+   `_dof_based_fill_qpoint_buffer!` dispatch path in `dof_based_coo.jl`.
+
+3. `evaluate_entry(kernel, geometry_cache, qpoint_buffer, layout_i, layout_j, elem_id)`
+   — the actual microkernel. Returns the single scalar `K[i, j]` for the
+   local DOF pair `(i, j)` described by two `DOFLayoutEntry` values.
+   `elem_id` is the volume element index (needed for facet-oriented kernels).
+   Called inside Pass 2 of `assemble!`, in a hot loop, so it must also
+   be allocation-free.
+
+Together these three methods let the DOF-based assembler dispatch on any
+`AbstractKernel` without baking in continuum-specific assumptions, while
+keeping the inner loop fully type-stable thanks to the compile-time
+`local_dof_layout(E)` table that produces the `DOFLayoutEntry` arguments.
+=#
+
+"""
+    qpoint_buffer_eltype(kernel::AbstractKernel) -> Type
+
+Element type of the per-quadrature-point buffer the kernel needs. The
+DOF-based assembler allocates `Vector{qpoint_buffer_eltype(kernel)}` of
+length `n_ips` per element and fills it once via `update_qpoint_buffer!`
+in Pass 1.
+
+A kernel must define this method; there is no default.
+"""
+function qpoint_buffer_eltype end
+
+"""
+    update_qpoint_buffer!(buffer, material_workspace, kernel::AbstractKernel)
+
+Populate the per-quadrature-point buffer for one element from the
+per-element material workspace. Called once per element in Pass 1 of the
+DOF-based assembler.
+
+Must be allocation-free.
+"""
+function update_qpoint_buffer! end
+
+"""
+    evaluate_entry(
+        kernel::AbstractKernel,
+        geometry_cache,
+        qpoint_buffer,
+        layout_i::DOFLayoutEntry,
+        layout_j::DOFLayoutEntry,
+        elem_id::Int,
+    ) -> Float64
+
+Compute the scalar stiffness contribution `K[i, j]` for the local DOF
+pair `(i, j)` on a prepared element (`elem_id` indexes `mesh.connectivity`).
+
+Each `DOFLayoutEntry` describes one local DOF as
+`(field_idx, entity_local, component)`, produced at compile time by
+`local_dof_layout(::Type{Element{K, P, S, N}})`. This contract lets a
+kernel inspect both DOFs (e.g. to dispatch on the field pair for
+multi-field problems) while keeping the call site uniform.
+
+`geometry_cache` provides `∇N`, `detJ·w`, and node coordinates.
+`qpoint_buffer` is the kernel-specific buffer filled by
+`update_qpoint_buffer!`.
+
+Must be allocation-free in the inner loop.
+"""
+function evaluate_entry end
+
+"""
+    evaluate_mass_entry(
+        kernel::AbstractKernel,
+        geometry_cache,
+        qpoint_buffer,
+        layout_i::DOFLayoutEntry,
+        layout_j::DOFLayoutEntry,
+    ) -> Float64
+
+Compute the scalar mass-matrix contribution `M[i, j]` for the local
+DOF pair `(i, j)`. Same call shape and constraints as `evaluate_entry`,
+but for the `(N_i, ρ N_j)` bilinear form instead of `(B_i : C : B_j)`.
+
+The default implementation returns `0.0`, so a kernel that has not opted
+into mass-matrix support transparently produces a structural-zero `M`
+through `apply_M!` / `assemble_M!`. Continuum and heat kernels override
+this; thermo-elastic / new-physics kernels can override or inherit
+zero-mass.
+
+Reads basis values from `geometry_cache.N_data` (the SoA batch added in
+the geometry-cache refactor) and the per-IP weights from
+`geometry_cache.detJ_w`. Material density / heat capacity is carried on
+the kernel itself rather than per-IP, so this microkernel does not
+look at `qpoint_buffer` for the linear case (the argument is still
+present so a future variable-density material drops in without changing
+the assembler).
+
+Must be allocation-free in the inner loop.
+"""
+@inline evaluate_mass_entry(
+    ::AbstractKernel,
+    geometry_cache,
+    qpoint_buffer,
+    layout_i,
+    layout_j,
+) = 0.0
+
+"""
+    reference_fields(kernel::AbstractKernel) -> (fields_ref::NamedTuple, empty_state::NamedTuple)
+
+Per-quadrature-point material *reference* values used to seed the
+per-element material workspace at the start of every assembly pass.
+
+For *stateless, constant-tangent* materials (linear elasticity, linear
+heat conduction) every IP can simply receive `fields_ref` directly,
+eliminating any per-IP constitutive call in Pass 1 and keeping the
+assembler allocation-free. For materials with state (plasticity etc.)
+this same hook can return a sensible "current-step zero" reference and
+the per-IP update is done elsewhere.
+
+The returned `NamedTuple`'s field names must match what the kernel's
+`update_qpoint_buffer!` reads from the material workspace, e.g.
+`(σ, 𝔻)` for `ContinuumKernel`, `(q, k)` for `HeatKernel`.
+
+A kernel must define this method; there is no default — the previous
+`(σ=…, 𝔻=…)` hardcode lived inside `DOFBasedCOOCache` and made the
+assembler accidentally continuum-only.
+"""
+function reference_fields end