# Builder numerical evidence for T717 Run: 20261115T174305Z; engine codex-cli; model gpt-6.1-sol; effort xhigh. This is builder evidence, not the independent scientific or argument-selection verdict. The controller must arrange both before any offering. The archived core and all six task-relevant references were read from the matching local bundle. Manifest: /home/ubuntu/numberdb-campaign-20261115-sol/repo/agents/runs/20261115T174305Z-build-skills/manifest.json. Manifest SHA-256: c5ee30e128246f511fb7c0e8e37f8b7d2419d5fad25227688a6111fdf9686049. Bundle SHA-256: d6661c2f21613524889fa3a57d3c95f8330344ca62e5a856211c08a282b7dbab. The identical manifest is attached as skill-manifest.json. Both archived local/live complete bundles were checked against their recorded resource hashes. All 70 values are native RealBall objects. The minimum measured relative accuracy of the 100-digit generator enclosures is 446 bits. The finite range is exactly the 70 identities justified in argument-selection.md. certification.md derives the explicit complete-tail bounds, including the positive endpoint. computed.json preserves the exact decimal encodings and full ball strings, not merely reported counts. Python: 3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0]. The computation used numberdb 0.1.10; further dependency versions are recorded in the stored-value verification output. checks.py independently computed all 70 references: 60 interior values use exact Python math.comb and mpmath summation; endpoint r=1,2 use the DK angular formulas, quadrature and closed forms; endpoint r=3,4 use gamma/polygamma summands with Levin acceleration. Every endpoint was recomputed at 140 and 180 decimal digits and agreed beyond 120 digits. The native implementation, formula and acceleration differ from the Stirling/EM ball generator. The direct interior check shares the defining series but independently implements its coefficients and arithmetic. All 9 applicable r=2 identities were also tested, including the positive boundary. Every reference falls inside the decimal interval actually encoded for the intended stored entry. beta-checks.py adds 6 independent beta-integral comparisons, covering r=1,...,5, the negative sign and the largest positive interior argument. It also checks the first-order differential relation in all 20 (r,u) regimes of the s=1 slice. These are in addition to the direct series comparisons, not checks of a sibling table. Levin and quadrature known-answer controls independently reproduce pi^4/24. The comparator rejects a deliberately wrong answer shifted by 10^(-50), and the main checker also rejects NaN. Finite, narrow generator balls are required independently. The actual check-results.json and beta-results.json preserve comparisons, residuals, controls, references and method limitations. The references are numerical checks, not themselves rigorous enclosures; the stated rigour comes from the separately bounded generator. The repository's existing dry_run.py and check.py were executed unchanged; their source copies and dry-run.log are attached. This checks native arithmetic types, finiteness, size, suspicious exactness and imports/prose. All 70 stored entries were read back. Their 140-digit generator enclosures fit inside the stored 100-digit intervals, and verify(sample=None,digits=140) matched 70/70 read-only. The displayed Programs example was executed and matched its independently computed reference. readback-results.json records these checks, dependencies and the comparison against the preserved independent reference checkpoint. Reusing that checkpoint is explicitly distinct from a fresh scientific critic verdict. Reproduction on this campaign's builder: - agents/sage.sh compute.py generate.py - agents/sage.sh checks.py computed.json - agents/sage.sh beta-checks.py checks.py computed.json - agents/sage.sh run_dry.py dry_run.py check.py generate.py The runner mounts /work read-only. Checks emit JSON on stdout before optional local file writes; the host captures it in this artifact directory. Each reference is checkpointed immediately. The initial failed report write and interrupted acceleration probe remain in the run artifacts; neither was silently treated as successful numerical evidence.