diff --git a/src/stage1/recurse_fwd.jl b/src/stage1/recurse_fwd.jl
index fa8a99fe..dbe43ed2 100644
--- a/src/stage1/recurse_fwd.jl
+++ b/src/stage1/recurse_fwd.jl
@@ -13,6 +13,7 @@ struct ∂☆new{N}; end
 # we split out the 1st order derivative as a special case for performance
 # but the nth order case does also work for this
 function (::∂☆new{1})(B::Type, xs::AbstractTangentBundle{1}...)
+    @info "∂☆new{1}" B supertype(B) typeof(xs)
     primal_args = map(primal, xs)
     the_primal = _construct(B, primal_args)
     tangent_tup = map(first_partial, xs)
@@ -23,8 +24,9 @@ function (::∂☆new{1})(B::Type, xs::AbstractTangentBundle{1}...)
         tangent_nt = NamedTuple{names}(tangent_tup)
         StructuralTangent{B}(tangent_nt)
     end
+    the_final_partial = maybe_construct_natural_tangent(B, the_partial)
     B2 = typeof(the_primal)  # HACK: if the_primal actually has types in it then we want to make sure we get DataType not Type(...)
-    return TaylorBundle{1, B2}(the_primal, (the_partial,))
+    return TaylorBundle{1, B2}(the_primal, (the_final_partial,))
 end
 
 function (::∂☆new{N})(B::Type, xs::AbstractTangentBundle{N}...) where {N}
@@ -41,7 +43,8 @@ function (::∂☆new{N})(B::Type, xs::AbstractTangentBundle{N}...) where {N}
             tangent_nt = NamedTuple{names}(tangent_tup)
             StructuralTangent{B}(tangent_nt)
         end
-        return tangent
+
+        return maybe_construct_natural_tangent(B, tangent)
     end
     return TaylorBundle{N, B}(the_primal, the_partials)
 end
@@ -50,6 +53,28 @@ _construct(::Type{B}, args) where B<:Tuple = B(args)
 # Hack for making things that do not have public constructors constructable:
 @generated _construct(B::Type, args) = Expr(:splatnew, :B, :args)
 
+
+maybe_construct_natural_tangent(::Type, structural_tangent) = structural_tangent
+for BaseSpaceType in (Number, AbstractArray{<:Number})
+    @eval function maybe_construct_natural_tangent(::Type{B}, structural_tangent) where B<:$BaseSpaceType
+        try
+            # TODO: should this use `_construct` ?
+            # TODO: is this right?
+            unwrap_tup(x::Tangent{<:Tuple}) = ChainRulesCore.backing(x)
+            unwrap_tup(x) = x
+            field_tangents = map(unwrap_tup, structural_tangent)
+            B(field_tangents...)
+        catch       
+            error(
+                "`struct` types that subtype `$($BaseSpaceType)` are generally expected to provide default constructors (one arg per field), and to be usable as their own tangent type.\n" *
+                "If they are not please overload the frule for the constructor: `ChainRulesCore.frule((_, dargs...), ::Type{$B}, args...)` " * 
+                "and make it return whatever tangent type you want. But be warned this is off the beaten track. Here be dragons 🐉"
+            )
+        end
+    end
+end
+
+
 @generated (::∂☆new{N})(B::Type) where {N} = return :(zero_bundle{$N}()($(Expr(:new, :B))))
 
 # Sometimes we don't know whether or not we need to the ZeroBundle when doing
diff --git a/test/forward.jl b/test/forward.jl
index f8040639..0a43feee 100644
--- a/test/forward.jl
+++ b/test/forward.jl
@@ -1,6 +1,7 @@
 module forward_tests
 using Diffractor
-using Diffractor: TaylorBundle, ZeroBundle, ∂☆
+using Diffractor: TaylorBundle, ZeroBundle, DNEBundle, ∂☆
+using Diffractor: first_partial, primal
 using ChainRules
 using ChainRulesCore
 using ChainRulesCore: ZeroTangent, NoTangent, frule_via_ad, rrule_via_ad
@@ -173,6 +174,32 @@ end
 end
 
 
+@testset "custom number types" begin
+    struct CustomNumber <: Number
+        val::Float64
+    end
+
+    double_and_custom_num(x) = CustomNumber(2.0*x)
+    let var"'" = Diffractor.PrimeDerivativeFwd
+        @test double_and_custom_num'(100.0) == CustomNumber(2.0)
+    end
+end
+
+@testset "custom array type" begin  
+    struct MyLittleStaticVector{N, T} <: AbstractVector{T}
+        val::NTuple{N, T}
+    end
+    Base.size(::MyLittleStaticVector{N}) where N = (N,)
+    Base.getindex(x::MyLittleStaticVector, ii::Int) = x.val[ii]
+    
+    once_twice_three_times(x) = MyLittleStaticVector((x, 2x, 3x))
+    @assert once_twice_three_times(10.0) == MyLittleStaticVector((10.0, 20.0, 30.0))
+    
+    🥯 = ∂☆{1}()(DNEBundle{1}(once_twice_three_times), TaylorBundle{1}(10.0, (1.0,)))
+    @test primal(🥯) = MyLittleStaticVector((10.0, 20.0, 30.0))
+    @test first_partial(🥯) == MyLittleStaticVector((1.0, 2.0, 3.0))
+end
+
 @testset "taylor_compatible" begin
     taylor_compatible = Diffractor.taylor_compatible