research.md

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

4968 bytes, as of the version from 2026-11-18 04:29 (current). Recorded here, not run.

# T743 research and handoff leads

Checked 2026-11-18 under codex-cli / gpt-6.1-sol / xhigh. Only the selected
Family 3.5B is being built. proposal.json maps proposal #336 and contributes
to #263; #262, #290 and #335 are context, not direct origins. The creation
message preserves both exact URLs (claim.json). #263 was read live.

Checked original source: Bertin–Zudilin, DOI 10.1007/s40316-016-0068-4,
journal HTML Introduction and The second family; arXiv:1601.07583 PDF
Introduction, section 2 pp. 4–7 and final p. 14. The original polynomial is
Q_k=(x²+x+1)y²+(x⁴+kx³+(2k−4)x²+kx+1)y+x²(x²+x+1).
Ordinary logarithmic measure, natural log, normalized Haar torus, convergent
log singularities. No principal value or modified measure.

Write u=k−2 when reading source comparisons. On the torus use s=x+x⁻¹,
a=s+1, b=s²+ks+2k−6. The source's inverse monomial substitution preserves
measure. Roots have product 1. The real numerator contribution is
log((|b|+sqrt(b²−4a²))/2) for positive discriminant and log|a| otherwise.
Subtracting log|a| globally has integral zero, but dropping its local
contribution only on the negative-discriminant region is incorrect.
Contact factorization and Q5 factorization have executable exact checks.

The source has a wrong sign in an intermediate derivative of a logarithm;
this build uses no such derivative. Conjectural Q_{−1} L-values are not used
as oracles or asserted as current open problems. q5=4G/pi follows by exact
factorization and an invertible monomial transformation to the diamond
family; G is Catalan's constant. It supplies a nonzero independent control.

Live discovery: catalogue.json (one paginated fetch), candidates.json,
discovery.json, preclaim-metadata.json and preclaim-search.json. Existing
T664 is a different three-variable Q family; T649 and T658 have different
polynomials. T742 is the distinct P family, now offered_for_review, so the
earlier dossier's private/unoffered status was stale. It is read only here.
lookups.json checks q(2), q(−1), q(5): no public matches for the first two,
T282 and unrelated T153 at the shared last value. Equality of one value is
not equality of families. Fresh metadata preceded creation of T743.

The seven proposed tags were checked in the exact live registry. All exist,
including draft-only Mahler measure and hyperelliptic curves (both public
counts zero). Retain all seven: each applies to this integral, generic
genus-two geometry, elliptic quotient/regulator context and special values.
No statistical-mechanics tag is inferred. No tag aliases or other tables edited.

The preceding gzipped screening report was attempted through connected GitHub
and returned 404; unavailable, not silently treated as inspected. The live
proposal and original paper are retained. Source files reused from the family
source archive were independently inspected; sibling numerical values are
not used as computation inputs. SHA-256s live in evidence-hashes.json.

Common conventions and cross-links: HREF{T742} for q(k)=p(k−2), k>=6;
HREF{T282#4} for q5. The R family is unbuilt; its future link and reciprocal
P link remain controller/family-review work, not authority to edit T742.
No unresolved definition alternative remains. See argument-selection.md for
the mixed coverage and honest noncanonical cutoff.

Skill: ../20261118T025446Z-build-skills/manifest.json; matching full bundle
d6e85d49873fac5bc748d068b0288cf242b0f9aebf3a9a0fff70295019227420.
The copied skill-manifest.json keeps this evidence with the draft.

Primary precision measurement: k=6 is the limiting contact case. Its
140/180-dps agreement hull prints about 113 significant digits; replayed
width is 2e−14 of the stored 100-digit last place. All others retain about
140. Cancellation at the coincident discriminant contact s=−2 is the likely
cause; it has not been isolated by a separate factored-discriminant benchmark.
An extra, unrequested 125-digit margin assertion in checks.log therefore
failed. The check was corrected to the actual 100-digit claim with two
guard digits (width <0.01 stored ulp), and measured widths retained in
checks-second.log. No values or generator were changed. The independent
135-dps original-polynomial check still must pass this difficult case.

Final checked outcome: all 73 separate 135-dps checks passed, including k=6,
and all 73 live stored intervals contain those independently computed values.
Known-answer and deliberate-error controls passed. verify(sample=None)
passed on hash-bound preserved primary intervals; see postfill-results.json.
The fresh recomputation attempt lost its final output when the exec session
returned 143; its child was tracked to exit and no fresh verdict is claimed. The audit is
clean, rendered formulas/radicals/relations and history citations were read.
Independent critic and controller offering remain pending. The GitHub progress
reply was refused with 403; issue-reply.md preserves its text.