# T732 mathematical and numerical method This is builder evidence, not an independent critique. Canonical instructions: ../20261117T035346Z-build-skills/manifest.json; matching local/live bundle SHA-256 984965681e2f8925c511a121c118efcde59581c7b0d906cac8b641bf6925e55a. The inspected primary definitions are Watkins v2, Section 2 (Proposition 2.2, Theorems 2.6–2.8, Definition 2.9), and Ogg (1974), p.449. Source bytes, locations and SHA-256 are in sources.json in the run workspace. The table's orientation, differential, cusp, field trace, exceptional units and parallelogram convention are those stated in the family proposal. No source's rank-one height algorithm or unverified coordinate reconstruction is used. For every (N,D,beta), enumeration of positive forms with N|A and B=beta mod 2N stops only after the reduced forms exhaust independently enumerated primitive proper classes. The complete correspondence is preserved in numerical-evidence.json and final-entry-evidence.json. Thus the trace is neither an average nor a stabilizer-weighted orbit sum. Independent squarefree-radicand and prime-splitting enumeration checks all 262 arguments. Let I be the integral structure of the cuspidal sign-zero weight-two modular-symbol space. At genus one it is rank two and is H_1(X_0(N),Z). Gamma_0(N) Farey generators give loops represented by {infinity,gamma infinity}. A pair with determinant ±1 in I is selected; this establishes a full lattice, not merely two linearly independent periods. For gamma=(a,b;c,d), c>0, take z=(-d+i)/c. Then gamma z=(a+i)/c. Evaluating the q-primitive difference gives the loop period, with no untracked cusp translation. Conjugation is computed exactly on I; a primitive integral +1 eigenvector determines omega_1. A unimodular complement, its imaginary sign, and exact conjugation normalize omega_2. Its real part is exactly delta*omega_1/2. All integral vectors, loop matrices, cusp ratios, models and numerical period comparisons are preserved. The exact curves in heegner.py supply a_n through point counts/Euler factors. Modularity places their normalized forms in S_2(Gamma_0(N)); their conductor is checked, and exact modular-symbol coefficients agree through the Sturm bound. The period construction uses f_N directly, so there is no assumption that any arbitrary isogenous model has Manin constant one. Independent invariant-differential periods of the stated models agree numerically with the constructed basis to 120 decimal places; that comparison is not used to turn approximate periods into certified balls. The elementary tail bound uses |a_n|<=d(n)*sqrt(n) and d(n)<=2*sqrt(n), whence |a_n|/n<=2. With a certified upper rho<1 on |q|, the primitive tail is <=2*rho^(M+1)/(1-rho). A floating-point estimate chooses a starting M only; a ball inequality certifies the final M. Input square roots, exponentials, separate powers, summation and lattice-coordinate reduction propagate rounding and input uncertainty. At least 480 bits are used. Finite, narrow results and final 100-digit decimal encodings are checked on every row. Exact real coordinates are justified as follows. Fricke has differential eigenvalue -1 at every selected N (checked exactly). Its base-point term is c_0=Phi_N(0), with c_0=kappa*omega_1 and kappa computed by the integral period mapping. Conjugation and Fricke both reverse the orientation and permute the entire ideal class group. Therefore 2*Re(z)-h(D)*c_0 is a real lattice period, so 2*Re(z)/omega_1-h(D)*kappa is an integer. A ball of radius below 1/4 containing one integer isolates it. The resulting rational real-period ratio is proved by this identity, not guessed from decimal agreement. It also determines an exactly zero real component after reduction when applicable. No imaginary vanishing is inferred from a small midpoint. Exact half-coordinate ties in rectangular lattices are decided by this rational relation; no unresolved ball boundary is accepted. The direct check integrates the normalized cusp form along the radial q-path, equivalent to a vertical upper-half-plane path. It constructs a_n independently from sign-one Hecke modular symbols and integrates the entire polynomial q*sum_{n=1}^M a_n*(q*t)^(n-1), 0<=t<=1, using Arb/acb_calc. The same elementary primitive tail bound encloses omitted terms. Rectangular splitting stabilizes evaluation. Every sampled trace is compared before and after reduction. The lattice enclosures, CM representatives, modularity theorem and Arb backend are shared; the coefficient computation and numerical integration differ. The saved sample plan covers each N's first field, largest class orbit, closest boundary, and all eligible D=-3,-4 orientations. Conjugate-orientation checks cover every row. Checkpoints preserve the long quadrature work. A defect in this snapshot client's form='ball' formatter rounds the displayed center and radius independently without enlarging the radius. The original diagnostic ball strings are therefore not submitted as claimed intervals. The decoder explicitly adds one printed-center ulp (zero for an exactly printed zero midpoint) and one printed-radius ulp when recovering those checkpoints. This encloses both rounding errors and retains over 103 checked decimal places. Final table values and final comparison output use the ordinary decimal encoding with its one-ulp interval semantics. The unchanged direct quadrature can be rerun in a fresh directory without the optional checkpoint attachment. The client also writes an exact-zero real ball as [0,0] in a complex string, but its precision counter does not recognize that component. The generator uses the public ComplexInterval type with exact rational real component 0 and rigorously outward binary-rational imaginary endpoints for these entries. This preserves the proof and its uncertainty without patching the client or returning uncertified strings. The standard deterministic dry run is rerun on the final generator. Controls test polynomial integration of t^2, the zero cusp, full-period reduction, the exact level-11 modular-symbol cusp ratio 1/5, and an independent critical integral at the Fricke fixed point. Negative controls require rejection of omitted classes, averaging, sign reversal, lost cusp translation, an unreduced full period, a wrong sublattice, NaN and a wide finite ball. Actual outcomes and samples belong in builder-check-results.json; no independent-actor verdict is asserted here. Certified algebraic identification through (j(tau),j(N*tau)) and exact group-law reconstruction was not attempted. Neither the numerical agreement nor a point lying on a model is claimed to establish that extra identification. No height normalization, BSD assertion or new relation is inferred.