numerical-evidence.md

back to table · edit · history · where entries came from · files · download

4535 bytes, as of the version from 2026-11-17 14:09 (current). Recorded here, not run.

# Numerical evidence for T738

All 17 selected values were computed with mpmath 1.4.1 at 150 and 200 decimal
digits, using the same unregularized function definition at both precisions.
The generator returns nonzero-width RealIntervalFieldElement objects through
numberdb.agreeing; 100 significant digits are retained. Python 3.12.3,
numberdb 0.1.10 and the builder's Sage runtime were used. computation.jsonl
and prepared-values.json retain both answers and actual elapsed times.
The total generating time was 6.945 seconds; the slowest entry took 4.867 s.

Builder separate-formula checks (not the independent critique):

* All 11 inputs with |z|<1 were checked using Arb complex balls at 512 bits.
  Exact rational upper parameters and arguments were enclosed. For the
  positive quadratic argument, an enclosure of 27*((-1+sqrt(5))/2)^15 was
  propagated through hypergeometric evaluation, independently of the
  generator's subtracted radical formula. Balls were finite, real-compatible,
  and had over 400 bits of relative accuracy.
* All 6 outside-disk negative inputs were checked with the Barnes formula
  DLMF 16.5.1 at 115 decimal working digits, on Re(s)=-min(alpha)/2, with the
  conjugate-symmetric integral split at imaginary heights 0, 1/2, 1, 2, 4,
  8, 16, 32, 64 and 128. It uses gamma-function quadrature, rather than
  mpmath's _5F4 continuation. This is a separate formula, sharing mpmath's
  arithmetic and gamma library. No certified quadrature or tail bound is
  claimed. The maximum relative discrepancy in these/disk comparisons was
  2.780e-116.
* Every value was additionally checked with the Euler–Abel rank-four integral
  at 35 decimal working digits. The maximum relative discrepancy was
  1.503e-36. This lower-precision check tests the rank reduction, coordinate
  and normalization; it alone does not certify 100 digits. It shares mpmath
  hypergeometric machinery, but evaluates rank four under an integral.
* The Barnes checker reproduces the nonzero elliptic value at z=-4 using
  1/AGM(1,sqrt(1-z)) as an independent control, beyond 105 digits. It rejects
  a factor-two normalization error and an altered decimal. The exact origin
  controls the Abel normalization. Readback checks additionally mutate stored
  last places and exchange conjugates; their executable output is attached.

The deterministic dry_run.py/check.py checks passed for all 17 native values:
finite enclosures, arithmetic types, decimal exactness, source imports, prose
and sizes. The longest native value was 154 characters; the native entries
block was 3.5 KB. No value is asserted exact from its digit pattern.
The coefficientwise Abel relation was checked exactly through n=31 for all
eight templates, and both reduced radical identities and defining-polynomial
identities were checked in exact rational pair arithmetic.

check_values.py and check-results.json contain actual comparisons and controls. dry-run.log and
deterministic_checks.py retain the exact checks. verify.log records package
reproduction against the stored draft, which is not an independent oracle.
All numerical values have heuristic (agreement-checked) rigour; interior
Arb certification is not promoted to a table-wide claim. The future critic
must independently inspect sources and numerical claims before any offer.

Skill evidence: ../20261117T124903Z-build-skills/manifest.json, copied here as
skill-manifest.json; status ready, matching core SHA256
cb9ef957600e7b7e302c71f5a0e2df2b560aef670cfc507f72a65d0dc4eeb517 and bundle
889383d95f882353fc09d210af003abe6edcf0331f27c377d7acdabafa344208.
source-hashes.json records the manifest and source artifact hashes.

Stored-data checks: verify(sample=None) matched 17/17, with zero differing,
missing or extra entries. Each actual stored decimal interval contains its
separate-formula reference; the maximum error is 0.464 ulp. Adding three ulps
to any stored center is rejected, as is swapping the two real conjugates.
The combined (template,z) display group preserves the source-row enumeration
including both quadratic entries at the end. readback-results.json and
readback_checks.py retain these checks.

generate-original.py has the hash recorded in prepared-values.json and is the
exact source used for the first checkpoints. The current generate.py only adds
an installation command to its docstring; it was recomputed for the full
stored-data verification. The final selection and live generator hash are in
build-fingerprint.json. No independent critique sidecar has been written here.