# Builder numerical evidence: T736 Archived skill manifest: ../20261117T101917Z-build-skills/manifest.json; matching bundle 984965681e2f8925c511a121c118efcde59581c7b0d906cac8b641bf6925e55a. The copied skill-manifest.json is attached. Source passages, software versions, selection and exact outputs are retained in this directory. This is builder evidence and is not a scientific or argument-review verdict. All 142 selected ideals pass build_check.py: 90 split primes, 3 inert primes, 49 composites. Exact character-free quadratic sums and finite-field Jacobi sums infer a cubic sum from each candidate sextic value. The cubic trace conversion is checked by its exact cube and its real part S_p/2, where S_p is the unweighted cubic exponential sum. The three candidate cube roots have different real parts at every selected split prime; both wrong roots fail. The inert cubic value is q. CRT checks every composite with both unsimplified sextic residue factors. This does not import generate.py or sibling numbers; Sage exact cyclotomic arithmetic and Arb are shared dependencies. Elementary derivation of the Gauss–Jacobi identity: expand G(alpha)G(beta) as a double sum over x,y and group by t=x+y. The t=0 contribution vanishes because alpha*beta is nontrivial. For t nonzero put x=tu,y=t(1-u); the inner sum is (alpha*beta)(t)J(alpha,beta). This proves the identity for every finite field and this additive character, including the inert q^2 fields. Here eta=psi^3, chi=psi^2 and eta*chi=bar(psi). Complex conjugation and x -> -x give the stated psi(-1) factor. No universal assertion is inferred from testing. Independent inert known answer: psi is trivial on F_q^* when q=5 mod6. On x=u+v*omega the different-additive character for a rational generator is exp(plus-or-minus 2*pi*i*v/q). For v=0 the nonzero character sum is q-1; for each v nonzero it is the same value A, and total character orthogonality gives A=-1. Summing the additive factors gives q-1-A=q. This checks all three inert values without their production sextic sums. Local-controls check 10 prime-power/shift combinations (including nonzero ones), every conjugate ideal's E-primary sign correction, the nonunit mask at exponent six, the standalone Programs expression, and T142 at (7,3). T142 is a cross-context check, not the evidence for this ideal-indexed family. The same-prime use of the coprime CRT formula is deliberately rejected. 512-bit Arb Horner evaluation of the exact finite polynomial has minimum 484 bits of relative component accuracy. Six rational values remain exact. Every stored decimal must contain a separate 640-bit evaluation of the phase-validated exact algebraic representation. Controls reject both wrong cubic roots, missing CRT phases, the wrong norm-7 sign/additive convention, a 10^-60 shift, NaN and a broad interval. Dry-run checks use the repository's unmodified dry_run.py and check.py. Full stored-value checks follow upload; results will be in stored-check-results.json. No check output is invented. Reproduce offline: sage -python build_check.py entries.json and sage -python local_controls.py (place all named attachments together). For full live readback, pipe your authorized NumberDB key to sage -python check_stored.py. Default generate.py only verifies the table.