# Agent 6 — adversarial admission decision Reviewed mathematical baseline: `5434d96f0034f842b0229f1d69dbbf3fb7fea270` (merged PR73). Working audit HEAD: `971c8bd02c834cf31ab98ea2a50060289fbece40`. The coordinator refreshed the remote again during this review; the scientific baseline was unchanged. I read the current agent/proposal gate, register and continuation scopes, current checkpoints/freeze and evidence ledger, the shared brief, relevant PR72/73 and NS100/101 records, and reviews 01–05. The coordinator's full review is recorded in the brief. This is an independent paper check of the stated dependencies and calculations, not a replay of all historical proofs, computations or external curvature certificates. Both original-evidence recovery groups remain OPEN. **Wall check: Same open gap — ZERO-GEOMETRY, NS100/101, PR72/73.** What changes: an explicit smooth-completion error bound, an analytically rejected local certificate, a positive Bessel-reference margin and a direct first-Laguerre control screen can be assessed. What remains unchanged: the complete original arithmetic lower-sign estimate is absent. The exact target is `L1[X](r)=X'(r)^2-X(r)*X''(r)>=0` for every real r, equivalently positive definiteness of the complete associated kernel C. This first-Laguerre target alone is not RH, G2 or the cofinal signed-arithmetic floor. ## Decision and severity **Admit zero original-theta sign candidates. Stop before any scan.** The notes contain useful supporting estimates and scoped rejections. They do not, individually or together, provide a complete original-theta sign method. This verdict follows from the dependency checks below, not from agreement between agents. | Finding | Severity and disposition | |---|---| | No proved original lower comparison or lower-reference margin with the required complete error comparison | **Admission blocker.** Already disclosed in the notes; not a newly found error in their claims. No computation or row follows. | | Global regularity of review 02's quotient needs pointwise `p'(t)<0`, not just the bare phrase strict log-concavity | **Minor formulation correction, resolved.** The coordinator changed the premise to an explicit conditional assumption. The diagonal rejection does not depend on it. | | Review 03 controls `X-X_M`; review 04's margin belongs to a different Bessel reference | **Mandatory scope restriction.** Substituting the Bessel margin into the smooth-completion transfer would introduce an unproved comparison. No note actually makes that substitution. | | Review 05 proves a sufficient parameter range in the NS101 family, not the fixed historical parameter 2 | **Mandatory quantifier restriction.** The final note states this correctly. No fixed-2 claim is admitted. | No unresolved major algebra, factor, sign or boundary defect was found in the particular supporting calculations checked below. This does not validate every theorem or source mentioned in the repository. ## 1. Review 03: the strip estimate survives the audit Keep its notation: `a=pi/4`, `v=pi/8`, `B=pi/sqrt(2)`, `d=pi/(4*sqrt(2))`, `beta=pi/(2*sqrt(2))`, and ```text w(z)=1/(1+exp(-2*a*sinh(2z))), Delta_M(z)=w(z)*T_M(z)+w(-z)*T_M(-z). ``` The second term cannot be omitted. The identity for Delta uses the complete theta reciprocity, analytically continued inside `|Im z|=0`, and `|Im z|<=v`, the stated division at u-star gives either `|q|<=1/2` or ```text |continuous exponent argument of q| <= sqrt(pi^2/8+log(2)^2) < pi/2. ``` Thus `|1+q|>=1/2` in the first part and `>=1` in the second. The constants in the two weight bounds are valid. The closed strip has no weight poles; holomorphy and real smoothness are separate facts, and both are addressed. For `y=exp(2u)>=1`, the two tail estimates reduce to ```text n^2*y >= (n^2+y)/2, (d_eta*y+b*n^2/y)/2 >= sqrt(d_eta*b)*n >= beta*n. ``` Here `b=pi*cos(2*Im z)>=B` and `d_eta=a*cos(2*Im z)>=d`. Both inequalities check. The remaining half of the second exponent retains `d*y/2`; only a nonnegative term is discarded. Summation over all `n>M` therefore gives exactly the declared geometric factor `exp(-beta*(M+1))/(1-exp(-beta))`, with the factor 4 on one half-strip and factor 8 in the full-line moment constants C_j. These constants are finite and independent of r and M. No numerical evaluation is needed for the inequalities to have a precise meaning. Contour shifting is legitimate for the complete Delta: its strip majorant decays at both horizontal ends and on the closing vertical pieces. The polynomial factor after differentiating the transform is bounded by `(|Re z|+v)^j`, giving ```text |Dhat_M^(j)(r)| <= C_j exp(-v*|r|-beta*(M+1)), j=0,1,2. ``` The differentiation is of the actual transform, not of its nonsmooth absolute-value upper envelope. The estimates also hold at r=0. Real-axis integration by parts for the differential operator has zero endpoint terms: Delta and its differentiated corrections decay there. Consequently ```text E_M=X-X_M=-(r^2+1/4)*Dhat_M/4 ``` has exactly the epsilon bounds printed in review 03. In particular the second derivative contains `2*C_0+4*|r|*C_1+(r^2+1/4)*C_2`, all divided by 4. Those terms are necessary and present. The quadratic error bounds also check. Expanding either in `(X_M,E_M)` or `(X,E_M)` gives the three linear terms and both possible quadratic absolute allowances. Review 03's original-theta A_j constants dominate the contour-shifted moments by the complete theta series, with evenness used for the other half-strip. Thus its fully specified bound (13) is valid without unknown X_M values. The quantifier is genuinely ```text for every integer M>=1 and every real r, the fixed-M derivative bounds hold. ``` One may therefore select any integer M(r) pointwise afterward. One may not call the derivatives of `r -> X_(M(r))(r)` the fixed-M derivatives. No derivative-of-M error was found in the note. In particular, for `M(r)>=ceil(|r|)+M_0`, its exponent rates are `pi/4+beta` and `pi/4+2*beta`, with polynomial degrees at most 2 and 4. Both exceed `pi/2`, as asserted. This is an absolute approximation estimate; it establishes no lower scale for L1. I also checked the two endpoint/shape claims. The homogeneous mode at negative infinity contributes -1 to the weighted boundary bracket, so the preserved normalization is 1/4. For fixed M the slower artificial tail is negative after applying D, with leading term `-a^2*exp(9u/2-a*exp(2u))`. Neither its real-space negativity nor its cancellation by the complete tail decides the transform's L1 sign. **Surviving support:** a specified smooth reciprocal approximation with uniform complete error bounds. **Missing premise:** an independently proved `L1[X_M](r)>=B_M(r)` or the corresponding Btilde comparison on every required height, including any zero-margin locations. Mere nonnegativity of L1[X_M] would not pay a positive absolute error budget. ## 2. Review 02: the local certificate is correctly rejected For fixed shifts, the simultaneous derivative is `D=partial_u+partial_v`, not a derivative of their difference. The exact entrywise integration by parts gives ```text R_h=w-Dh-(p(u)+p(v))*h, w=(u+v)^2/4. ``` The integral boundary term is zero under the displayed hypotheses. The sufficient condition is positivity of every finite residual matrix for arbitrary real nodes and complex coefficients. Entrywise positivity, finitely many minors or the original diagonal energies cannot replace this condition. Positive phi factors preserve matrix inertia. For `h_*=w/(p(u)+p(v))`, direct differentiation gives ```text R_*(u,u)=-u/p(u)+u^2*p'(u)/(2*p(u)^2) =u*[u*P1(2u)-P(2u)]/[2*P(2u)^2]. ``` Both clock factors, `p(u)=2P(2u)` and `p'(u)=4P1(2u)`, are correct. The complete theta tail gives, for `z=pi*exp(2u)`, ```text p=-2z+9/2+O(1/z), p'=-4z+O(1/z), R_*(u,u)=(u-u^2)/(2*pi*exp(2u))+O(u^2*exp(-4u)). ``` The Gaussian gap from n=1 to n=2 permits the stated differentiated relative remainder; this is not differentiation of an arbitrary big-O. The leading term is negative eventually, so the local residual cannot be positive semidefinite. The fixed Gaussian predecessor also fails by its displayed original-tail diagonal. These are analytic failures of these certificate choices, with no numerical threshold asserted. The initially phrased global removability needs `p'(t)<0` at every real t: strict concavity as a bare definition can permit isolated second derivative zeros. This was raised and corrected forward in review 02. Under the explicit condition, the quotient's expansion at the crossing is valid and its double-exponential weighted endpoint terms vanish. Crucially, the large-positive-diagonal calculation is defined and rejects the candidate even without any global curvature theorem. There is no remaining dependence on an un-replayed curvature certificate for this rejection. The negative residual is not a negative value of the integrated Q. A one-translate integral remains C(0)>0; compensation elsewhere is allowed. No exclusion of every divergence correction or nonlocal method follows. ## 3. Review 04: the reference margin is real, the transfer is absent I read [Gasper, arXiv:0801.2996v1](https://arxiv.org/html/0801.2996v1), especially (2.6), directly. Its y-squared coefficient supplies review 04's margin after the frequency rescaling. With `B_a(r)=K_(ir/2)(a)`, the transform kernel is `exp(-a*cosh(2u))` and the chain-rule factor in L1 is 1/4. The positive square integral converges at both endpoints and is strictly positive for each real r, including reference zeros. This does not claim a constant positive frequency floor. The independently derived crude lower bound also checks. With `nu=r/2`, `V=max(2a,nu^2,1)`, integration by parts gives `integral u^2 exp(-v*cosh u)du<=K_0(v)/v`. The elementary cosine bound then gives `K_(i nu)(v)>=K_0(v)/2` for `v>=nu^2`. On `[V,2V]`, the stated elementary K_0 bound and integration length V yield ```text m_a(r) >= [log(2)/(16V)] exp(-4V-2). ``` The constants and domains are consistent. Its very fast decay does not make it a practical original-theta comparison automatically. The perturbation formula in review 04 is exact for real-valued functions on the real axis. Its sufficient lower comparison (T) may omit epsilon_1 squared because `(E')^2` is nonnegative. Review 03 bounds an absolute difference and correctly retains that allowance. The two formulas are not in conflict. Neither divides by a reference at a zero. The source's shifted Pólya sum is a different reference: the extra term in (3.3) cannot be dropped on the basis of its conjectured positivity. The separate real-zero argument does not supply an error from original theta to that reference. I checked this scope directly in Gasper; I did not independently re-audit every other paper in review 04's bounded source table. **Non-composition warning:** review 03 controls `X-X_M`, while this lane would need `X-kappa*B_a` or the Pólya difference. No estimate for `X_M-kappa*B_a`, or lower margin for X_M, is supplied. Thus combining the two supporting results still leaves an explicit additional unproved bridge. A smooth kernel, common endpoint normalization, a real-tail match or a generic real-zero reference theorem does not fill it. ## 4. Review 05: direct control failure with the correct quantifiers I expanded the product independently. For `q_beta(r)=3+2cos(beta*r)` and `X_beta=q_beta*X/B_beta`, ```text B_beta^2 L1[X_beta] =q_beta^2 L1[X]+2*beta^2*(2+3cos(beta*r))*X^2. ``` At `r=pi/beta`, this is `L1[X]-2*beta^2*X^2`. The proof then uses only the complete positive moments `m0,m2`, with `k=m2/m0>0`. For `beta^2>=pi^2*k`, the moment bounds give `X>=m0/2` and ```text L1[X]/X^2 <= 4*r^2*k^2+2*k <=6*k<2*beta^2. ``` The advertised explicit upper bound `-(beta^2-3k)*m0^2/(2*B_beta^2)` follows with the correct inequality direction, since its coefficient is negative and `X^2>=m0^2/4`. There is no assumption of the original L1 sign or of any zero's location. The real division is safe because this argument proves X positive at the selected point. Consequently every `beta>=max{2,pi*sqrt(m2/m0)}` yields a direct negative L1 value while retaining the NS101 family's critical-strip restriction on its added zeros. The constants are exact full moments, not estimated finite prefixes. This is a sufficient analytic parameter range; it is not a numerically located threshold or a separately established assertion for the historical beta=2 member. This control retains the general reciprocal positive-mixture structure, smooth even boundary matching, endpoint normalization and generic pair identities. It changes the original lattice and fails the normalized original Jacobi IVP. Hence it directly rejects the general-properties inference to first-Laguerre positivity, while leaving an estimate using the distinguished original arithmetic open. No inference from off-axis zeros to negative first Laguerre has been used. Review 01's separate smooth log-concave countermodel likewise checks: the Gaussian Fourier multiplier gives the printed negative value at pi, and complete moment domination preserves that strict sign after small positive double-exponential damping. It matches different hypotheses; it must not be silently combined with NS101 to claim one model satisfying every property of both. Neither model has the original Jacobi data. ## 5. Final rendered-report claim and scope check I read the coordinator's `theta-six-agent-2026-10-05-v1.html` report. Its central claims match the reviewed notes: the new bound controls `X-X_M`, the Bessel reference is a different object, a zero margin cannot absorb a positive absolute allowance, and the reciprocal-family threshold does not decide the historical beta=2 member. The conditional global quotient regularity and unconditional diagonal rejection are separated. The report admits no original-theta sign method and makes no first-Laguerre to RH inference. One minor wording correction was requested and verified: the initial phrase that every calculation retains the original kernel was too broad for deliberately changed references and controls. The corrected text says the target retains the complete original kernel and explicitly identifies approximations and controls. This is resolved. No further claim/scope correction is required by this check. I did not independently repeat the coordinator's manuscript-build or browser-layout validation. ## 6. Exact stop and bounded next step The required all-coefficient original comparison remains ```text ||D1||^2 <= 2||E||^2+Re ``` for every finite set of real translates and every complex coefficient vector. Equivalently, the original complete L1 is nonnegative at every real height. There are no source/parity exceptions that shrink this quantifier. The existing fixed-prefix closure is not uniform in a growing cutoff; the new smooth estimates do not reopen that closed fixed family. The control record is now sharper, and the approximation budget is explicit. Neither is a positive estimate of the original mixed form. NS100's actual negative slice continues to exclude only per-slice sign; Davenport–Heilbronn has different arithmetic/completion and does not match the original IVP; NS74/83 are NB controls with different domains. These mismatches are not control passes. No generic inference is admitted on the strength of a missing match. No current proposal meets all five admission needs: a complete exact estimate of the original target, an arithmetic estimating method, its uniform range, a complete dependency and matched controls. The original Jacobi equations identify an object but have not estimated the missing signed comparison. Listing an additional unproved inequality does not make it a method. **Budget:** zero for scans, model fitting, larger cutoffs or certificate replays. A future supplied candidate may receive one bounded paper admission review only if it states an independently estimable signed inequality or a complete reference/error comparison, with the actual original arithmetic step and all-height scope. Such success would prove the first-Laguerre target only; failure would stop that specified method. There is no recommendation to try an adjacent variation automatically. Only this assigned internal note was written. No numerical or symbolic scan, research row, manuscript change, commit, push or outreach was made. Other agents' edits were preserved. The one minor phrasing correction was communicated to and applied by the coordinator in review 02. **Final wall check: Same open gap.** Zero sign candidates admitted. Supporting estimates survive in their stated scopes; the complete original first-Laguerre lower sign, RH, G2 and the actual cofinal signed-arithmetic lower bound remain open.