EXACT CLAIM Consider two independent uniform bits with qubit outputs ρ±^(1)=(I+aZ±√(1−a²)X)/2 and ρ±^(2)=(I+cZ±√y X)/2, where 0≤a,c≤1 and 0≤y≤1−c². Thus the first channel has pure conditional outputs. Entropies are in bits. Define h(t)=h₂((1+t)/2), J_c(y)=h(c)−h(√(c²+y)), and Δ=s₁+s₂−s₁s₂−H(X₁⊕X₂|B₁B₂). For a,c<1, put A=1−a², C=1−c², q=y/C, f(t)=atanh(t)/t with f(0)=1, L(t)=f(t)/(2 ln 2), T(a)=Af(a)/h(a), and F(q)=Σ_{n≥1}q^(n−1)/[n(2n−1)]. Then Δ ≥ h(a)yL(c)[T(a)−F(q)]. Consequently the erasure upper bound holds whenever F(q)≤T(a). In particular, uniformly for 63/64≤a<1, 0≤c<1, and 0≤y≤(1−c²)/2, Δ ≥ h(a)yL(c)/120. This strip is strict when y>0. If a=1 or c=1, Δ=0. Also T(a)→2 ln 2 as a→1; hence for each fixed q<1 the sufficient condition holds for all a sufficiently close to 1, uniformly in c. This is a partial one-pure qubit result, not a resolution of the arbitrary-output conjecture. FULL FROZEN PROOF Write ℓ=ln 2. First establish the two analytic estimates used in the claim. 1. Parity information. For temporarily arbitrary equal-purity qubit signals b,d, let z=b²d² and let D=((I+aZ)/2)⊗((I+cZ)/2). The parity-conditioned states are D±√z X⊗X/4. They are unitarily equivalent and their average is D. Set K(z)=S(D)−S(D+√z X⊗X/4). Thus H(X₁⊕X₂|B₁B₂)=1−K(z). The matrix splits into two 2-by-2 blocks, each with fixed trace w and eigenvalues (w±√((p−r)²+z/4))/2, where p,r are its diagonal entries. Its entropy is w h(√((p−r)²+z/4)/w)−w log₂w. The expansion 1−h(t)=Σ_{n≥1}t^{2n}/[2n(2n−1)ℓ] shows that each block's entropy loss is convex as a function of z. This remains true at the physical boundary by continuity. Consequently K(z)≥zK′(0). For a,c<1, direct differentiation of these blocks gives K′(0)=[(a f(a)·a−c f(c)·c)/(a²−c²)]/(2ℓ), with coincident arguments interpreted continuously. Equivalently, K′(0)=(1/(2ℓ))∫₀¹ dt/[(1−a²t²)(1−c²t²)]. The two factors in this integrand are increasing functions of t. Chebyshev's integral inequality therefore gives K′(0)≥f(a)f(c)/(2ℓ). For completeness, Chebyshev here follows by integrating the nonnegative expression (u(t)−u(s))(v(t)−v(s)) over the unit square. We have proved K(z)≥z f(a)f(c)/(2ℓ). 2. Individual information. Integrating the derivative of J_c yields the nonnegative expansion J_c(y)=Σ_{n≥1} yⁿ/(2nℓ) ∫₀¹ t^{2n−2}/(1−c²t²)ⁿ dt. For y9/8 for 63/64≤a<1 using elementary certified constants. Let δ=1−a, p=δ/2, and R=ln(2/δ). Natural binary entropy satisfies −p ln p−(1−p)ln(1−p)≤p(R+1), because −(1−p)ln(1−p)≤p. Consequently T(a)≥ℓ(1+a)/a · ln((1+a)/δ)/(R+1) ≥2ℓ[R+ln(1−δ/2)]/(R+1). Since δ≤1/64, ln(1−δ/2)≥−(δ/2)/(1−δ/2)≥−1/127>−1/100. Also ℓ=2Σ_{k≥0}(1/3)^{2k+1}/(2k+1)>56/81>69/100, so R≥ln128=7ℓ>483/100. The function (R−1/100)/(R+1) is increasing. These inequalities yield T(a)≥(69/50)(482/583)=16629/14575>9/8. All quantities multiplied in this argument are positive. For 0≤q≤1/2, separating the first two terms of F and using n(2n−1)≥15 for n≥3 gives F(q)≤1+q/6+q²/[15(1−q)]≤1+1/12+1/30=67/60. Hence T(a)−F(q)≥9/8−67/60=1/120. Inserting this in step 3 proves the quantitative strip. For a<1, c<1, and y>0, its right-hand side is positive. 5. Endpoints and scope. If a=1, the first signal vanishes and s₁=1; the parity bit is independent of both outputs, so Δ=0. If c=1 then y=0, with the same conclusion because s₂=1. Finally, as δ=1−a tends to zero, natural binary entropy at δ/2 is (δ/2)[ln(2/δ)+1+o(1)], while Af(a)=δ(1+o(1))ln(2/δ). Thus T(a)→2ℓ. The entropy series gives F(1)=2ℓ, and its positive coefficients imply F(q)