# T739 numerical evidence (builder, not an independent verdict) Canonical skill: ../20261117T132607Z-build-skills/manifest.json, also attached as skill-manifest.json. Both complete archived bundles were hash-validated; local/live bundle SHA256 is 889383d95f882353fc09d210af003abe6edcf0331f27c377d7acdabafa344208. Source URLs, downloaded bytes and SHA256 are in source-hashes.json. Runtime versions and the installed HGM module hash are in versions.json. Production: generate.py, SageMath 10.9, numberdb 0.1.10. All 120 polynomials are native integer polynomials; each first/second trace uses explicit modulus p^N>10*p^(2*r). The rank-five weight-four Weil bound is 5*p^(2*r), so its centered residue has a unique possible integer trace. Newton identities and the local functional equation reconstruct all six coefficients. Prepared results and each modulus exponent are checkpointed in prepared-values.json. No approximate decimal is promoted to an exact value. Independent method within this builder: check_values.py implements BCM Theorem 1.3 via finite-field Gauss sums, not Sage HGM trace functions. Gauss sums are DFTs evaluated by Bluestein convolution in native 192-bit Arb complex ball polynomials. Every real and imaginary component must be finite with radius <1/4; imaginary balls contain zero, and ceil(lower)=floor(upper) isolates one integer. Across all 240 stored first/second traces, the largest real radius was 2.146e-43, substantially smaller than the integer-isolation threshold. Exact rational inputs include all specialization reductions and normalizing factors. Controls: the DFT Gauss sums on F_25 agree with separate direct sums for all 24 characters. DPVZ row1 p3, row9 p5, row13 p13 full polynomials agree exactly. The source's row13 p11 display has a sign error: its quadratic is printed 1-30*T+14641*T^2, implying trace151. An exact BCM computation in Q(zeta_110) (exact_control.py/results) gives trace91. The definition-consistent factor is 1+30*T+14641*T^2; independent complex balls, Sage and PARI agree. Both coordinate choices were explicitly tested in investigation.log; inversion does not explain the discrepancy. The printed p11 polynomial is rejected, not counted as an unchanged published-control pass. The table's Comments disclose this finding. For all 120 entries, Euler's criterion independently gives coefficient sign -Legendre(1-z,p), derived from Watkins 11.1.1 at tilde_t=1/z. All complete polynomials also agree with PARI hgmeulerfactor at 1/z. PARI and Sage have related Gross--Koblitz ancestry; that agreement supplements, and does not replace, the complex-character-sum implementation. Twelve direct higher-extension sums check coefficients provided by duality: r=3 on the first admissible prime for each of eight templates, r=3 on template 2.8 at p31, r=4,5 on template2.2.2.2.2,z=-1/4,p3, and r=4 on that fiber at p7. This sample covers both determinant signs and the largest selected prime. Full arguments and radii are in check-results.json. Universal identities derive from the source formula; finite agreement alone is not their proof. Negative controls reject changed trace sign, coefficient mutation, omitted linear factor, changed determinant sign and reversed coordinate. The asymmetric coordinate control is row9,p5; reciprocal pairs in the first four rows are not used to diagnose inversion. Actual stored-value check: stored-document.json was read from NumberDB, flattened to stored-records.json, parsed as Z[T], and fed to the same independent BCM/character/determinant/PARI checker. stored-check-results.json records 120/120 checked, no failures, both checker controls true, and all twelve higher checks. The document agrees exactly with the intended upload. Separately, Generator.verify(sample=None) reports 120/120 matched (verify.log); this is reproducibility, not independent scientific evidence. The existing agents/table-build/dry_run.py and check.py passed arithmetic-type, finite-enclosure, suspicious-exactness, imports, prose and size checks. The block is 15.7 KB and the longest entry is 84 characters. No pipeline source was edited. Reproduce in this environment using agents/sage.sh with named input files: prepare.py generate.py (JSON records captured by capture.py), check_values.py prepared-values.json, exact_control.py, run_deterministic.py generate.py agents/table-build/dry_run.py agents/table-build/check.py, stored_check.py check_values.py stored-records.json, and verify.py generate.py with the API key piped on stdin. The standalone attachments also document ordinary Sage laptop invocations. The shared toolchain was not changed. This evidence authorizes no offer or independent acceptance. Asai recognition beyond the published noncircular controls remains deferred: no new Hilbert-side Hecke dataset or held-out-prime matching is claimed. Global bad factors and critical L-values are outside this table. A later fresh critic must independently adjudicate the source sign discrepancy and every scientific/argument-selection claim.