# Numerical evidence for T764 Builder evidence only; an independent critique and argument-selection verdict are required before the controller offers the private draft. All 112 exact polynomial values passed the deterministic dry run, including native coefficient types, exactness, decimal-pattern and prose checks. The longest polynomial has 21 characters. The dry-run measured block is 7.4 KB. The executable checks.py completed in 3.121 seconds: 112/112 complete factors agreed with exact PARI newforms and independently implemented BCM finite-character sums. Newform selection used exact source coefficients 0..16 and compared 0..99; none of the primary truncated sums was used to select a newform. Primes 37 and 41 were unused in selection and lie outside the stored range: all 28 held-out comparisons passed. Eight literal all-half quadratic-character convolutions passed. The four exact zero traces are template 3.4 at p=11,17,23,29. Known nonzero source factors, reversed-sign and omitted/wrong character corrections passed their controls. All real/imaginary Gauss-sum radii were finite and below 10^-80 at 384 bits; the largest measured selected-entry radius was 1.55e-109. The imaginary balls contain zero; each real ball isolates exactly one integer. Integrality comes from the canonical rational-data trace, not from decimal appearance. The exact trace recovery uses the published congruence and weight bound; no numerical error or 100-digit padding applies to these integer polynomials. Software: Sage 10.9, PARI 2.17.1, Python 3.12.3, numberdb 0.1.10. Commands: agents/sage.sh run-dry.py generate.py dry_run.py check.py; agents/sage.sh checks.py generate.py newform-reference.json check.py. Builder mounting is deployment-specific; standalone commands are in code. Results: check-results-offline.json; full execution: checks-offline.log. Source and normalization evidence: source-manifest.json, ltyz-passages.txt, BCM Theorem 1.3 (arXiv:1505.02900v2), and newform-reference.json. The latter contains exact coefficients and source URLs/hashes. The level-216 q-expansion was transcribed from the displayed exact LMFDB page because its download/API endpoints returned challenge HTML; PARI independently reproduced all 100 transcribed coefficients. Other labels used exact trace downloads or the official API with dimension=1 checked. Some Sage documentation downloads returned HTTP 403; the inspected screening archive is retained with its hash. Skill manifest: ../20261119T014035Z-build-skills/manifest.json. All core/reference hashes were verified; matching local/live bundle SHA-256 b5c3c7985b7c9e178b4899d083a9e3bc4284989dff76f05e5b0e52835d47eed2. Core SHA-256 b79b7fbc9fa3ce63f2d642cccdac12b0d214757ab01db66c03a2d84264a58e5f. No sibling's stored values serve as independent evidence. The BCM implementation uses the same published formula as T674's checker, inspected as a component lead, and is independent of the primary truncated-series algorithm. This is computational validation, not a new proof of modularity or a critic's verdict. ## Stored readback and final presentation The authenticated draft readback reproduced all 112 entries with verify(sample=None): 112 matched, zero differing, missing or extra. The full newform and finite-character-sum checks were repeated on those stored values; all 112 factors and 28 extra-prime controls passed. See check-results-stored.json and checks-stored.log. Later edits only removed a positional word and replaced the displayed Programs example with an exact modular implementation tested in a fresh Sage process; Numbers are unchanged. The reusable generator's final documentation clarifies that preview reads the target and needs the author's key for a private draft. Its arithmetic and enumeration are unchanged and the final source was dry-run checked again. The audit has one advisory: the 468-character definition is longer than the corpus median. It retains the precise extracted constituent, normalization, realization and no-twist convention, so no unsupported mathematical claim or formatting fault was identified in that finding. Browser evidence covers all 112 numerical cells (56 in each of two server preview segments), all prose sections, upright parameter display and all four Similar tables links. Thirteen server-rendered segments returned HTTP 200 and no detected MathJax error. A single full preview returned HTTP 414; intermediate long previews also exceeded request-header limits. The run has no authenticated website browser session, so a full live-page visual check remains explicitly pending for the independent critic/controller. This is not a scientific verdict or permission to offer. See browser-results.json, inspected-render-results.json and the retained screenshots.