# T715 build evidence and handoff notes

Private populated draft: https://numberdb.org/T715 (Central_binomial_sums).
36 entries: p=1..9, v=-1/4,-1/16,1/16,1/4, ordered (p,v).
The n=0 term is retained and signed arguments are used without extra factors.
No table was published or offered; no review action was taken.

The generator uses 944-bit real balls and explicit error bounds to supply
100 significant digits. Maximum measured generator radius: 9.0165807e-131.
All 36 values were independently enclosed by 450-bit analytic quadrature plus
a proved exponential tail; maximum radius: 6.3633586e-119. Every complete
independent enclosure lies inside its stored decimal interval. Known controls
and all five deliberately invalid controls passed. Reproduction reports
36/36 matched, zero differing/missing/extra. The read-back computation is
stored-check-results.json, and the proofs are in computation.md. Both main
algorithms use Arb, so this is implementation/formula independence, not
library independence. A separate mpmath hypergeometric evaluation checks all
18 interior cases. Companion comparisons check normalization only.

agents/table-build/dry_run.py and check.py passed all 36 entries, native return
types, finite enclosures, exactness-pattern, import, prose and size checks.
See deterministic-checks.log. The generated ball block is 8.4 KB; the actual
100-digit decimal records are smaller. Versions: Sage 10.9, Python 3.12.3,
numberdb 0.1.10, mpmath 1.4.1 (runtime-versions.json).

The authenticated API audit has one remaining heuristic finding: the rational
parameter denominators are 4 and 16. This is assessed, not suppressed. Both
convergence endpoints are intrinsic; +/-1/16 are selected by CGLXZ's published
Apery-like identities. argument-selection.md gives exhaustive groups and the
noncanonical order bound. The independent critic must assess this rationale.
The audit's link issue was repaired. Render inspection also found an unsupported
Data properties.digits field; it was removed, retaining the supported precision
statement in rigour details and the generator contract. No numerical values
or selected identities changed in these presentation repairs.

Authenticated live HTML was read and rendered in the shared Chromium runtime.
Definition, parameter order, formulas, 100-digit rows, Programs, citations,
companion links and completeness were inspected. MathJax rendered 46 math nodes,
with no math errors, page errors, unresolved HREF/CITE markers or unknown
properties. Screenshots and live HTML are in the run work directory. The
Similar tables links point to T697, T177 and T698, all inspected live.
T713 is still offered and private; its log-sinh relation remains a future
cross-link plan, not an inaccessible link on this table. The planned harmonic
central companion should link through the H_(2n+1) shift when it exists.

The revision history was read back: creation and subsequent edits retain
https://github.com/numberdb/numberdb-data/issues/318 and the explicit partial
contribution to https://github.com/numberdb/numberdb-data/issues/263. The recorded
agent/tool is codex-cli and model is gpt-6-astra. This does not close wanted #263.
The queue remains controller-owned/claimed until the review workflow advances.

Canonical skill evidence: skill-manifest.json, copied unchanged from
agents/runs/20261111T025123Z-build-skills/manifest.json. Both archived bundles
were hash-verified and match, bundle SHA-256
 d6661c2f21613524889fa3a57d3c95f8330344ca62e5a856211c08a282b7dbab.
The core and six task-relevant references were read from the local archive only.

This report is builder evidence, not an independent scientific verdict or
argument-review sidecar. The controller must arrange those fresh independent
checks before offering the same draft. The generator has no local custom
source dependency beyond generate.py; checks.py, extra_checks.py and their
frozen comparison inputs are attached for reproducibility. No live skill,
pipeline source, published table, deployment or supervisor receipt was changed.