diff --git a/Project.toml b/Project.toml
index d422312..acf4ccf 100644
--- a/Project.toml
+++ b/Project.toml
@@ -10,6 +10,7 @@ CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
 DifferentiationInterface = "a0c0ee7d-e4b9-4e03-894e-1c5f64a51d63"
 FlexiChains = "4a37a8b9-6e57-4b92-8664-298d46e639f7"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
+OrderedCollections = "bac558e1-5e72-5ebc-8fee-abe8a469f55d"
 Random = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 
@@ -39,6 +40,7 @@ ForwardDiff = "1"
 LinearAlgebra = "1"
 LogDensityProblems = "2"
 Mooncake = "0.5.26"
+OrderedCollections = "1"
 Random = "1"
 Statistics = "1"
 Zygote = "0.7.10"
diff --git a/ext/DynamicPPLExt.jl b/ext/DynamicPPLExt.jl
index 7800d81..cc33ad4 100644
--- a/ext/DynamicPPLExt.jl
+++ b/ext/DynamicPPLExt.jl
@@ -136,13 +136,12 @@ end
 function ParallelMCMC._construct_flexichain(
     ::Type{VarName},
     vals::AbstractMatrix{<:Real},
-    internals::AbstractMatrix{<:Real},
+    internals::NamedTuple,
     ::Any,
-    internal_names::Vector{Symbol},
     model::DensityModelLDF,
 )
-    pwss = map(zip(eachrow(vals), eachrow(internals))) do (val, internal)
-        stats = NamedTuple{Tuple(internal_names)}(internal)
+    pwss = map(enumerate(eachrow(vals))) do (i, val)
+        stats = map(v -> v[i], internals)
         DynamicPPL.ParamsWithStats(val, model.logdensity.ld, stats)
     end
     return AbstractMCMC.from_samples(VNChain, hcat(pwss))
diff --git a/src/ParallelMCMC.jl b/src/ParallelMCMC.jl
index 7a9a557..4ad005e 100644
--- a/src/ParallelMCMC.jl
+++ b/src/ParallelMCMC.jl
@@ -4,6 +4,7 @@ using AbstractMCMC
 using CUDA
 using FlexiChains
 using LinearAlgebra
+using OrderedCollections: OrderedDict
 using Random
 using Statistics
 
diff --git a/src/interface.jl b/src/interface.jl
index 7523297..41d1de7 100644
--- a/src/interface.jl
+++ b/src/interface.jl
@@ -205,20 +205,23 @@ for TKey in (Symbol, VarName)
     )
         N = length(samples)
         D = model.dim
+        # Follow the sampler's working precision instead of pinning `Float64`, so a
+        # Float32 (e.g. GPU) run yields a Float32 chain rather than a silently widened one.
+        FP = typeof(sampler.epsilon)
 
-        internal_names = [:logp, :accepted]
-
-        vals = Matrix{Float64}(undef, N, D)
-        internals = Matrix{Float64}(undef, N, 2)
+        vals = Matrix{FP}(undef, N, D)
+        logp = Vector{FP}(undef, N)
+        accepted = Vector{Bool}(undef, N)
 
         for i in 1:N
             s = samples[i]
             vals[i, :] .= s.x
-            internals[i, 1] = s.logp
-            internals[i, 2] = s.accepted ? 1.0 : 0.0
+            logp[i] = s.logp
+            accepted[i] = s.accepted
         end
 
-        return _construct_flexichain($TKey, vals, internals, param_names, internal_names, model)
+        internals = (logp=logp, accepted=accepted)
+        return _construct_flexichain($TKey, vals, internals, param_names, model)
     end
 end
 
@@ -543,7 +546,7 @@ function _parallel_mala_update_progress!(
 end
 
 function _copy_trajectory_rows!(
-    vals::AbstractMatrix{Float64}, first_row::Int, S::AbstractMatrix, ncols::Int
+    vals::AbstractMatrix{<:Real}, first_row::Int, S::AbstractMatrix, ncols::Int
 )
     rows = first_row:(first_row + ncols - 1)
     S_host = Array(view(S, :, 1:ncols))
@@ -563,10 +566,10 @@ function _sample_parallel_mala_chain(
     progressname="Sampling",
 ) where {TKey}
     D = model.dim
-    internal_names = [:logp]
+    FP = typeof(sampler.epsilon)
 
-    vals = Matrix{Float64}(undef, N, D)
-    internals = Matrix{Float64}(undef, N, 1)
+    vals = Matrix{FP}(undef, N, D)
+    logp = Vector{FP}(undef, N)
 
     progress = _parallel_mala_progress(progress, progressname)
     x0 = _parallel_mala_initial_x(rng, model, sampler, initial_params)
@@ -586,7 +589,7 @@ function _sample_parallel_mala_chain(
                 rows = first_row:(first_row + nkeep - 1)
 
                 _copy_trajectory_rows!(vals, first_row, S, nkeep)
-                internals[rows, 1] .= view(logps, 1:nkeep)
+                logp[rows] .= view(logps, 1:nkeep)
 
                 x0 = copy(view(S, :, nkeep))
                 nsteps += nkeep
@@ -599,15 +602,15 @@ function _sample_parallel_mala_chain(
         end
     end
 
-    return _construct_flexichain(TKey, vals, internals, param_names, internal_names, model)
+    internals = (logp=logp,)
+    return _construct_flexichain(TKey, vals, internals, param_names, model)
 end
 
 function _construct_flexichain(
     ::Type{TKey},
     vals::AbstractMatrix{<:Real},
-    internals::AbstractMatrix{<:Real},
+    internals::NamedTuple,
     param_names::Any,
-    internal_names::Vector{Symbol},
     model::DensityModel,
 ) where {TKey}
     # Wrap user-supplied names in `Parameter`. This allows people to specify, e.g.,
@@ -616,19 +619,14 @@ function _construct_flexichain(
     to_parameter(vn::VarName) = FlexiChains.Parameter(vn)
     to_parameter(s::Symbol) = FlexiChains.Parameter(TKey <: VarName ? VarName{s}() : s)
 
-    # vals and internals are both `iters x params`
-    # FlexiChains expects `iters x chains x params`
-    arr = hcat(vals, internals)
-    arr = reshape(arr, size(arr, 1), 1, size(arr, 2))
+    N, D = size(vals)
 
-    # Wrap parameter names in the format that FlexiChains expects. Ref:
-    # https://pysm.dev/FlexiChains.jl/stable/arrays/#api-fromarray
     if param_names === nothing
         param_names = model.param_names
     end
     wrapped_param_names = if param_names === nothing
         # Wasn't defined either as `param_names` or `model.param_names`, so make some up
-        (to_parameter(:x) => (size(vals, 2),),)
+        (to_parameter(:x) => (D,),)
     else
         map(param_names) do n
             if n isa Pair
@@ -641,11 +639,23 @@ function _construct_flexichain(
         end
     end
 
-    all_names = (
-        wrapped_param_names...,
-        FlexiChains.Extra.(internal_names)...
+    arr = reshape(vals, N, 1, D)
+    param_chain = FlexiChains.FlexiChain{TKey}(arr, Tuple(wrapped_param_names))
+
+    isempty(internals) && return param_chain
+
+    #=
+    The internals have mixed element types (e.g. `Bool` for `accepted`/`is_warmup`,
+    `FP` for `logp`), so they cannot share a single stacked array without widening.
+    Store each in its own column via the dict-of-arrays constructor, which preserves
+    the natural element type (see issue #49), then merge with the parameters.
+    =#
+    extras = OrderedDict{FlexiChains.ParameterOrExtra{<:TKey},AbstractArray}(
+        FlexiChains.Extra(name) => internals[name] for name in keys(internals)
     )
-    return FlexiChains.FlexiChain{TKey}(arr, all_names)
+    extra_chain = FlexiChains.FlexiChain{TKey}(N, 1, extras)
+
+    return merge(param_chain, extra_chain)
 end
 
 function _sample_parallel_mala_blocks(
@@ -870,18 +880,18 @@ for TKey in (Symbol, VarName)
     )
         N = length(samples)
         D = model.dim
+        FP = typeof(sampler.epsilon)
 
-        internal_names = [:logp]
-
-        vals = Matrix{Float64}(undef, N, D)
-        internals = Matrix{Float64}(undef, N, 1)
+        vals = Matrix{FP}(undef, N, D)
+        logp = Vector{FP}(undef, N)
 
         for i in 1:N
             vals[i, :] .= samples[i].x
-            internals[i, 1] = samples[i].logp
+            logp[i] = samples[i].logp
         end
 
-        return _construct_flexichain($TKey, vals, internals, param_names, internal_names, model)
+        internals = (logp=logp,)
+        return _construct_flexichain($TKey, vals, internals, param_names, model)
     end
 end
 
@@ -1075,21 +1085,24 @@ for TKey in (Symbol, VarName)
         filtered = discard_warmup ? filter(s -> !s.is_warmup, samples) : samples
         N = length(filtered)
         D = model.dim
+        FP = typeof(sampler.epsilon_init)
 
-        internal_names = [:logp, :accepted, :step_size, :is_warmup]
-
-        vals = Matrix{Float64}(undef, N, D)
-        internals = Matrix{Float64}(undef, N, 4)
+        vals = Matrix{FP}(undef, N, D)
+        logp = Vector{FP}(undef, N)
+        accepted = Vector{Bool}(undef, N)
+        step_size = Vector{FP}(undef, N)
+        is_warmup = Vector{Bool}(undef, N)
 
         for i in 1:N
             s = filtered[i]
             vals[i, :] .= s.x
-            internals[i, 1] = s.logp
-            internals[i, 2] = s.accepted ? 1.0 : 0.0
-            internals[i, 3] = s.step_size
-            internals[i, 4] = s.is_warmup ? 1.0 : 0.0
+            logp[i] = s.logp
+            accepted[i] = s.accepted
+            step_size[i] = s.step_size
+            is_warmup[i] = s.is_warmup
         end
 
-        return _construct_flexichain($TKey, vals, internals, param_names, internal_names, model)
+        internals = (logp=logp, accepted=accepted, step_size=step_size, is_warmup=is_warmup)
+        return _construct_flexichain($TKey, vals, internals, param_names, model)
     end
 end
diff --git a/test/test-AbstractMCMC-Interface.jl b/test/test-AbstractMCMC-Interface.jl
index c033993..033dc27 100644
--- a/test/test-AbstractMCMC-Interface.jl
+++ b/test/test-AbstractMCMC-Interface.jl
@@ -152,9 +152,49 @@ gradlogp_iface(x) = -x
         # logp values should be finite
         @test all(isfinite, chain[:logp])
 
-        # accepted values should be 0 or 1
+        # accepted is stored with its natural `Bool` type (issue #49)
         acc = chain[:accepted]
-        @test all(a -> a == 0.0 || a == 1.0, acc)
+        @test eltype(acc) === Bool
+        @test all(a -> a == false || a == true, acc)
+    end
+
+    @testset "chain element type follows sampler precision" begin
+        # Scalar param names so `chain[:a]` is a scalar-valued column whose eltype is the
+        # numeric type directly (a vector-valued param would nest inside a `Vector`).
+        model = DensityModel(logp_iface, gradlogp_iface, 2)
+
+        @testset "Float32 sampler → Float32 chain" begin
+            sampler = MALASampler(0.1f0)
+            @test sampler.epsilon isa Float32
+            chain = sample(
+                MersenneTwister(1),
+                model,
+                sampler,
+                50;
+                chain_type=SymChain,
+                progress=false,
+                param_names=[:a, :b],
+            )
+            @test eltype(chain[:a]) === Float32
+            @test eltype(chain[:b]) === Float32
+            @test eltype(chain[:logp]) === Float32
+            @test eltype(chain[:accepted]) === Bool
+        end
+
+        @testset "Float64 sampler → Float64 chain" begin
+            sampler = MALASampler(0.1)
+            chain = sample(
+                MersenneTwister(1),
+                model,
+                sampler,
+                50;
+                chain_type=SymChain,
+                progress=false,
+                param_names=[:a, :b],
+            )
+            @test eltype(chain[:a]) === Float64
+            @test eltype(chain[:logp]) === Float64
+        end
     end
 
     @testset "sample() with custom param_names" begin
diff --git a/test/test-Adaptive-MALA.jl b/test/test-Adaptive-MALA.jl
index d4c5a61..5bac2b0 100644
--- a/test/test-Adaptive-MALA.jl
+++ b/test/test-Adaptive-MALA.jl
@@ -201,13 +201,16 @@ end
     @test FlexiChains.Extra(:is_warmup) in extras_names
 
     @test all(isfinite, chain[:logp])
-    @test all(x -> x == 0.0 || x == 1.0, chain[:accepted])
+    # accepted / is_warmup keep their natural `Bool` type (issue #49)
+    @test eltype(chain[:accepted]) === Bool
+    @test eltype(chain[:is_warmup]) === Bool
+    @test all(x -> x == false || x == true, chain[:accepted])
     @test all(s -> s > 0, chain[:step_size])
 
     # Warmup samples appear at the start
-    @test chain[:is_warmup][1] == 1.0
-    # Post-warmup samples have is_warmup == 0
-    @test chain[:is_warmup][end] == 0.0
+    @test chain[:is_warmup][1] == true
+    # Post-warmup samples have is_warmup == false
+    @test chain[:is_warmup][end] == false
 end
 
 @testset "step_size is constant after warmup" begin