diff --git a/src/algorithms/ctmrg/c4v.jl b/src/algorithms/ctmrg/c4v.jl
index bfefcae34..093506ac4 100644
--- a/src/algorithms/ctmrg/c4v.jl
+++ b/src/algorithms/ctmrg/c4v.jl
@@ -71,7 +71,16 @@ end
 # as expensive as `svd_vals` on CPU but much cheaper on GPU, and gives effectively
 # the same information. For the QR projector, off-diagonal elements are still
 # present until CTMRG converges, so we need a fallback for the non-diagonal case.
-corner_spectrum(C::AbstractTensorMap, ::C4vCTMRG) = eigh_vals(C)
+# gauge fix sign of spectrum such that the eigenvalue of largest magnitude is positive
+function corner_spectrum(C::AbstractTensorMap, ::C4vCTMRG)
+    S = eigh_vals(C)
+    if -minimum(S) > maximum(S)
+        for (_, b) in blocks(S)
+            b .= .-reverse(b) # spectrum of -C: negated and, to stay ascending, reversed
+        end
+    end
+    return S
+end
 
 """
     convergence_tensors(env::CTMRGEnv, alg::C4vCTMRG) -> (corners, edges)
diff --git a/test/ctmrg/c4v_sign_branches.jl b/test/ctmrg/c4v_sign_branches.jl
new file mode 100644
index 000000000..10a5dca95
--- /dev/null
+++ b/test/ctmrg/c4v_sign_branches.jl
@@ -0,0 +1,10 @@
+using PEPSKit
+
+@isdefined(TestSuite) || include("../testsuite/TestSuite.jl")
+using .TestSuite
+
+is_buildkite = get(ENV, "BUILDKITE", "false") == "true"
+
+if !is_buildkite
+    TestSuite.ctmrg_c4v_sign_branches(Vector)
+end
diff --git a/test/testsuite/TestSuite.jl b/test/testsuite/TestSuite.jl
index 018a68a82..db2c7b000 100644
--- a/test/testsuite/TestSuite.jl
+++ b/test/testsuite/TestSuite.jl
@@ -121,6 +121,12 @@ module CTMRGFlavors
 end
 using .CTMRGFlavors
 
+module CTMRGC4vSignBranches
+    include("ctmrg/c4v_sign_branches.jl")
+    export ctmrg_c4v_sign_branches
+end
+using .CTMRGC4vSignBranches
+
 module CTMRGGaugeFix
     include("ctmrg/gaugefix.jl")
     export ctmrg_gaugefix_asymmetric, ctmrg_gaugefix_c4v
diff --git a/test/testsuite/ctmrg/c4v_sign_branches.jl b/test/testsuite/ctmrg/c4v_sign_branches.jl
new file mode 100644
index 000000000..3f8311d50
--- /dev/null
+++ b/test/testsuite/ctmrg/c4v_sign_branches.jl
@@ -0,0 +1,61 @@
+using Test
+using Random
+using LinearAlgebra
+using TensorKit
+using PEPSKit
+using Adapt
+using PEPSKit: ctmrg_iteration, compute_gauge_fix_gauge, fix_relative_phases, ScramblingEnvGaugeC4v
+
+# For half-integer U(1) charges, (-1)^{2Q} on the ket bonds negates the sandwich, so fixed
+# points come in pairs, one of which flips (C, E) to (-C, -E) every iteration. Both should converge.
+
+sd = 20260924
+H = heisenberg_XXZ(ComplexF64, U1Irrep, InfiniteSquare(); spin = 1 // 2)
+V = U1Space(0 => 1, 1 // 2 => 1, -1 // 2 => 1)
+Venv = U1Space(0 => 2, 1 // 2 => 2, -1 // 2 => 2, 1 => 1, -1 => 1)
+tol = 1.0e-10
+maxiter = 400
+
+# (-1)^{2Q} on a U(1) space
+function charge_parity(W)
+    X = id(W)
+    for (c, b) in blocks(X)
+        b .*= (-1)^Int(2 * c.charge)
+    end
+    return X
+end
+
+# sign by which one gauge-fixed iteration multiplies the edge
+function iteration_sign(network, env, alg)
+    env′, = ctmrg_iteration(network, env, alg)
+    signs, = compute_gauge_fix_gauge(env′, env, ScramblingEnvGaugeC4v())
+    env′ = fix_relative_phases(env′, signs)
+    return sign(real(dot(env.edges[1], env′.edges[1])))
+end
+
+function sign_twin(env)
+    E = env.edges[1]
+    g = charge_parity(space(E, 2))
+    @tensor Ê[χ_out D_ket D_bra; χ_in] := E[χ_out D_ket′ D_bra; χ_in] * g[D_ket; D_ket′]
+    return CTMRGEnv(env.corners[1], Ê)
+end
+
+function ctmrg_c4v_sign_branches(AT)
+    return @testset "C4v CTMRG converges on both sign branches of a half-integer U(1) state ($AT)" begin
+        Random.seed!(sd)
+        peps = InfinitePEPS(randn, Float64, physicalspace(H)[1, 1], V)
+        symmetrize!(peps, RotateReflect())
+        peps = adapt(AT, InfinitePEPS(map(a -> a / norm(a), peps.A)))
+        network = InfiniteSquareNetwork(peps)
+        alg = C4vCTMRG(; tol, maxiter, projector_alg = :C4vEighProjector)
+
+        env, info = leading_boundary(initialize_random_c4v_env(peps, Venv), peps, alg)
+        @test info.converged
+
+        env_twin = sign_twin(env)
+        @test iteration_sign(network, env_twin, alg) == -iteration_sign(network, env, alg)
+        env_twin, info_twin = leading_boundary(env_twin, peps, alg)
+        @test info_twin.converged
+        @test expectation_value(peps, H, env_twin) ≈ expectation_value(peps, H, env) atol = 1.0e-8
+    end
+end