# T5 — verdict for the Pascadi candidate **NO.** For the two-sequence bilinear class that Thm 1.2 belongs to, verified at the source abstract (arXiv:2511.08445v2, "bilinear (Type II) sums of Kloosterman sums with composite moduli `c` … For sums of length √c … saving `c^(−1/12)` for products of two primes of the same size"). The obstruction is algebraic, appears at three independent places, and no rearrangement of D1 supplies a product form. ## 1. The attempt Write the target as `Σ_{m,n} α_m β_n K(m,n;c)`. With second index `n := (h₁,h₂,u₁,u₂)` and `t` the variable introduced by the completion, ``` Obj = (1/c) Σ_{t mod c} Σ_n c_n · F̂_n(t) · S(t, r(n); c), c_n = c_{h₁} c̄_{h₂}, r(n) = −θ(h₁ℓ₂ − h₂ℓ₁) (r ≠ 0 retained), ``` which would need `α_t` independent of `n`, `β_n` independent of `t`, and `S(t, r(n); c)` of the theorem's kernel form. ## 2. First exact obstruction — one index map serves both sequences ``` F̂_n(t) = Σ_{m∈I_m} a_m Φ_{u₁,h₁}(m) conj(Φ_{u₂,h₂}(m)) e_c(−tm), Φ_{u,h}(m) = ψ(h/(um)), ψ(x) = e(x z₀/g) − e(x z/g) [exact] ⇒ F̂_n(t) = Σ_m a_m ψ(h₁/(u₁m)) conj(ψ(h₂/(u₂m))) e_c(−tm). ``` Its `n`-dependence is not removable: `ψ` is non-constant on the range of `h/(um)`, so every `t` with some `a_m ≠ 0` inherits it. The kernel's numerator is the *same* `n`: `r(n) = −θ(h₁ℓ₂ − h₂ℓ₁)`. Hence `n ↦ (coefficient of the first sequence, argument of the kernel)` is a single map, and no `α_t β_n` exists. This is exactly "a variable of the first sequence depends on a variable of the second". ## 3. Second obstruction — the `t`-sum is the Fourier dual of `m`, not a second sequence ``` Σ_{t mod c} e_c(−tm) S(t, r; c) = c · e_c(r m̄) · 1_{(m,c)=1} [exact] ``` Taking `(t, m)` as the two sequences returns the pre-completion sum identically: the completion is a finite Fourier transform, so `t` and `m` are conjugate, never independent. Completion cannot manufacture the missing sequence. ## 4. Third obstruction — the coprimality ties the first sequence's support to `n` The `m`-domain is `{m : (m,c)=1}` with `c = j_eℓ₁ℓ₂ = c(n)`. A bilinear hypothesis needs coprimality expressible in the pair; here the *domain* of the first sequence depends on the second index. The standard repair `1_{(m,c)=1} = Σ_{d|(m,c)} μ(d)` does not remove it: it introduces a third index `d` supported on divisors of `c(n)`, i.e. it relocates the dependence and raises the rank, and a two-sequence theorem cannot consume a three-index kernel. ## 5. Why no known separation rescues it — and the return to 103/100 The object does carry one exact factorization, in the **side** index: ``` F(m) = ψ(h₁/(u₁m)) · conj(ψ(h₂/(u₂m))) ``` a product of a left-side function of `(u₁,h₁,m)` and a right-side function of `(u₂,h₂,m)`, both read at the same `m`. D1's first Cauchy over `(m,q)` is precisely the device that consumes this side-factorization — it is what produces the `Q^(1/2)`, `Q` and `Q^(−1/4)+A^(−1/4)` factors. The variant that takes absolute values side-wise, `(R,q) ≤ q`, discards exactly that structure, and `structured-dispersion-estimate.md` §1 records that this variant returns **exactly the generic 103/100** — the step its own stop rule refuses. Therefore * a separation in the `(m ↔ n)` split (what Thm 1.2 needs) and the `(side 1 ↔ side 2)` factorization (what D1 exploits) are **different** factorizations; * buying the first requires destroying the second; * the price of destroying it is the generic exponent, not a constant. The closure is structural. No numeric improvement of any two-sequence saving changes it, because the failure is at the level of the coefficient *class*, not its size. ## 6. The one thing that would change the verdict (a different class, not PARTIAL) The needed object is not an identity or a missing bound: it is a theorem for **matrix-coefficient** Kloosterman forms `Σ_{t,n} A(t,n) S(t, r(n); c)` in which `A` has exactly the rank-one side structure above. Nothing in the corpus's recorded interface list lies in that class and no separation known here reaches it. That is the exact form the missing object would have to take, should the owner want the literature question asked — it is a *different* theorem class, so this row is **not** PARTIAL. **Access gap.** arXiv:2511.08445v2 abstract read at source (v1 11 Nov 2025, v2 21 Jun 2026); the HTML full text is 2.0 MB and the Theorem 1.2 display itself was not read this pass. The verdict is for the two-sequence class the abstract names; if that display were a matrix-coefficient form, §2–§4 would need re-examination.