numerical-evidence.md

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

4110 bytes, as of the version from 2026-10-27 19:42. Recorded here, not run.

# Numerical evidence for T559

Build run: 20261027T193344Z-build. Draft only; no offer, review, or publication.

Policy provenance: agents/runs/20261027T193344Z-build-skills/manifest.json,
attached as skill-manifest.json. The archived local and live instructions both
have SHA-256 e37efb7e2ededf6c1d7134c673f152f648f26e0267c556aebdf3ec65765d8b4c;
the canonical source-snapshot file has the same digest.

## Arithmetic and coverage

All 75 pairs of nonnegative indices with d=m^2+n^2-mn+n <= 60 are included,
in increasing (d,m,n) order. Values are exact elements of QQ[x], not rounded
100-digit approximations. Exact rational coefficients exceed the campaign's
precision objective without an error radius. Sage 10.9 and numberdb 0.1.10
were used through agents/sage.sh in the isolated builder.

The generator uses the physicists' Hermite recurrence and manual fraction-free
Bareiss elimination with row pivoting. Every polynomial division has zero
remainder. It compares the determinant's leading coefficient to
2^(sum k_j) product_(i<j)(k_j-k_i), including the sign of the ordered list,
then divides over QQ. Empty, singular, row-swap and polynomial determinant
controls pass. The actual Hermite inputs H_0,...,H_23 agree with T103.

The repository dry_run.py completed without findings. Shared exactness(),
measure(), names_its_rings() and agrees_with() were run. The longest value has
868 characters, at (m,n)=(2,8), of degree 60. The dry-run entries block is
26.6 KB. No disguised decimals or float coefficients occur.

## Independent reconstruction

checks.py constructs no determinants and imports no generator arithmetic. It
starts from Q=1 and solves for each successive monic polynomial P using

P''Q - 2P'Q' + PQ'' - 2x(P'Q-PQ') + 2(deg P-deg Q)PQ = 0.

This follows from the harmonic-oscillator Darboux eigenstate in Clarkson,
Theorems 3.6–3.7 and equation (3.15). For Q monic of degree e and P monic of
degree d, the coefficient multiplying the unknown coefficient p_j in degree
j+e is 2(d-j), which is nonzero for j<d. Thus descending coefficient solution
is unambiguous; the entire polynomial residual must vanish at the end.

Known-answer controls reconstruct H_k/2^k for k=0,...,25 and compare them
with the closed factorial coefficient sum, with exact rational division.
Four low-degree Okamoto polynomials also pass. A deliberately wrong sign is
rejected; incrementing any computed constant coefficient disagrees with the
independent construction. Every one of the 75 polynomials agrees coefficient
by coefficient. Degree, monicity and parity pass for all 75. The Darboux
identity passes for 75 m-neighbor and 75 n-neighbor pairs, computing required
neighbors beyond the stored degree bound independently. The Programs snippet
also agrees on all 75 entries.

## Source discrepancy resolved before writing formulas

Clarkson arXiv:0901.0139v2, equation (2.14a), prints a coefficient
2(m-2n+2) for the ordered pair (Omega_(m+1,n), Omega_(m,n)). At the positive
indices m=n=1, the monic polynomials are P=x^4+x^2-1/4 and Q=x^2+1/2.
The printed expression has residual -2PQ, not zero. Constant normalization
cannot repair a homogeneous bilinear identity. Their degree difference is 2,
so equation (3.15) requires coefficient 4, whereas (2.14a) prints 2.
The generic Hermite-Wronskian identity in Theorem 3.7 supplies the consistent
coefficients 2(2m-n+1) and 2(2n-m+2) in the two forward directions.
Both have been checked on every selected index. This isolates a source
indexing inconsistency; no numerical disagreement was resolved by selecting
one output without explanation. The table uses the general identity, not
(2.14). The explicit definition and boundary strings agree with the family
proposal and the measured degree formula.

## Read-back validation

The reproducible read-back procedure is checks.py with
NUMBERDB_CHECK_STORED=1 and draft credentials, followed by
GeneralizedOkamoto().verify(sample=None). The builder repeats all coefficient
comparisons and identities on the API's stored values before its handoff.
The final run artifacts record those results and the site audit.