# T749 numerical evidence (builder evidence)

Quantity and conventions: W(k)=<1/(1-H/k)>, H=8 cos(theta_1) cos(theta_2)
cos(theta_3), probability measure on [0,pi]^3, real k>=8. The graph has the
eight diagonal steps with unit adjacency weights and per-vertex normalization.
The step discount is 8/k. We store W, not W/k or a return probability.

The selected exact arguments are all integers 8..32. See argument-selection.md
for the intrinsic, documented-application and exploratory groups. The selection
and its disjoint exhaustive membership are asserted by check_green.py.
The archived canonical skill is ../20261118T075646Z-build-skills/manifest.json;
its exact copy is attached as skill-manifest.json, matching local/live bundle
SHA-256 d6e85d49873fac5bc748d068b0288cf242b0f9aebf3a9a0fff70295019227420.

## Generating method and rigorous precision

The actual runtime is SageMath 10.9, Python 3.12.3, numberdb 0.1.10,
mpmath 1.4.1, NumPy 2.4.4 and SciPy 1.17.1; versions.json preserves details.
Every Sage job ran through agents/sage.sh, on the configured builder.

The generator uses Arb elliptic K at the exact-real parameter
alpha=(1-sqrt(1-64/k^2))/2, enclosing the rational argument before the square
root. For k=8 it explicitly sets alpha=1/2 so no uncertain square root at zero
is taken. The domain gives 0<alpha<=1/2, and the elliptic function is real
analytic there. ComplexBallField.elliptic_k is used as the Arb implementation;
mathematical reality follows from this domain, not from discarding an unknown
imaginary part. Squaring and pi normalization carry input and rounding errors.

At 100 significant decimal digits, numberdb.bits(100,losing=128)=461 bits.
All 25 returned values are native RealBall objects and have 456 relative
accuracy bits. The standard client's to_text serializes them to 100 significant
digits with its one-unit-last-place uncertainty. The smallest value is above
one; every written ULP is exactly 10^-99. computed.json records all pre-upload
balls, native types and written intervals; no values are padded or exactified.

The copied, unmodified agents/table-build/dry_run.py and check.py inspect every
return type, finite bounds, imports, suspicious decimal exactness, prose and
size. compute.log preserves their output: all clean, 25 entries, 155 characters
for the largest raw ball, 4.8 KB raw block. The final serialized entries are
smaller. Missing mpmath in the outer Python interpreter did not cause a fallback
computation; the builder runtime supplies it. The source snapshot's older local
client lacks list_tables/table_metadata; discovery used the documented HTTP API.

## Independent checks of the whole selection

check_green.py imports no generator code. For each of the 24 exterior values,
it independently forms a_n=binomial(2n,n)^3/2^(6n) with exact integer binomials
and uses x=64/k^2. It sums a_n*x^n in 600-bit balls. The coefficient bound
0<a_n<=1 gives a positive tail bounded by x^(N+1)/(1-x), since x<1.
The exact rational tail is checked below 10^-155; N is doubled until this holds.
The computed ball includes both rounding and the entire positive tail, by
centering the omitted range [0,tail] at tail/2. The endpoint is independently
evaluated by Gamma(1/4)^4/(4*pi^3), whose derivation from DLMF is preserved in
source-passages.md. The independent enclosures have at least 516 accuracy bits.
All 25 lie strictly inside the *whole written decimal intervals*, using exact
rational endpoints, not merely midpoint agreement or broad overlap.

An additional implementation integrates the original angular kernel. Integrating
the third angle analytically gives k/pi^2 times the integral over [0,pi]^2 of
(k^2-64*cos(s)^2*cos(t)^2)^(-1/2). Reflect to [0,pi/2]^2, split into two
triangles and put (s,t)=(r,r*v). The Jacobian r removes the integrable inverse-r
corner at k=8. The stable denominator is
sqrt(k^2-64+64*(sin(r)^2+cos(r)^2*sin(r*v)^2)); at k=8,r=0 the transformed
integrand has limit 1/sqrt(1+v^2). The measure multiplier is 8/pi^2.
SciPy dblquad checks **all 25 arguments**, including the edge and cutoff.
Observed maximum absolute midpoint difference is 2.220446049250313e-16,
with estimated integration error at worst about 2.03e-12. The stated check
threshold is 10^-11; these double-precision estimates are heuristic, not the
source of the proven 100-digit claim.

extra_checks.py independently evaluates the squared 2F1(1/4,1/4;1;x) formula
at 180 decimal digits on all 25 written intervals. Its differences are below
the written ULP. The exact series check, rather than a high-precision
hypergeometric routine, supplies the rigorous exterior corroboration.

## Checker controls and identities

Actual eight-step walk enumeration gives counts 0,8,0,216 at lengths 1..4;
therefore W(k)=1+8/k^2+216/k^4+... . The zero-adjacency original-kernel control
returns the known nonzero value 1. At k=9, deliberately omitted factor 8,
wrong z power, modulus/parameter swap, unnormalized W/k, always-zero and
last-place-shift candidates are rejected. The exact written-interval comparator
accepts the known endpoint and rejects a 10^-98 shift. NaN and huge balls are
rejected before any containment assertion. See offline-check-results.json and
the later stored-check-results.json for actual controls and numeric outcomes.
A sign reversal of the kernel is an equality by symmetry, not a negative control.

The relation k M'(k)=W(k) follows by differentiating the common logarithmic
integral for k>8. On any compact exterior k interval the derivative integrand
is bounded, so differentiation under the integral is valid. The limit
M(k)-log(k)->0 follows by dominated convergence. The endpoint ordinary integral
converges because the denominator vanishes quadratically at finitely many
corners, while volume is three-dimensional. These are analytic arguments,
not claims proved by agreement on 25 numbers. The sources and the derived
normalizations are preserved in source-passages.md.

This is the builder's check evidence. It is not the fresh independent scientific
verdict or argument-selection sidecar required before the controller offers
T749. Readback checks and the final revision identity are in handoff.md.
