research.md

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

4417 bytes, as of the version from 2026-11-17 00:36 (current). Recorded here, not run.

# T729 build research

Checked sources are archived as griffon.html, brunjes.html and kloosterman.html; sources.json records their URLs and SHA-256 hashes. Griffon arXiv:1612.08721v2, Definition 2.2 and Theorem 2.3 express the surface middle polynomial as the ambient factor times the full character-orbit product. Remark 2.4 multiplies each eigenvalue by the inverse-coefficient character: for (-1,1,...,1), this is chi_0(-1). Brünjes arXiv:math/0301186, Section 4 assumes p>d and roots of unity in the base field; Theorem 4.6 gives one-dimensional primitive character lines. For n=4,d=3 at a nonsplit prime, apply it over F_(p^2) and use the cyclic-block determinant 1-beta*T^2 over F_p. Example 4.7 gives eigenvalues -56-21*zeta_3, 49 (multiplicity 20), -35+21*zeta_3 at p=7. Kloosterman arXiv:math/0703120v2, Lemma 6.1 supports monomial permutation; Definition 6.8/Corollary 6.9 use complement/inverse conventions and were not copied as our definition.

Normalization checked: even dimension n>=2, d>=3, p prime not dividing d; model X_1^d+...+X_(n+1)^d=X_0^d; geometric Frobenius, det(1-T*Frob), no Tate twist, full primitive middle cohomology including primitive algebraic classes. beta=chi_0(-1)J_Q equals the higher sum at y_1+...+y_(n+1)=1 by x_j=-x_0*y_j. This cancellation explains why the production higher sum already contains the model twist. The exact Gauss quotient independently computes that higher sum. If a prefix character becomes trivial in the binary recurrence, its next multiplier is -Q, not the usual nontrivial-product rule.

Live discovery: cached catalogue includes authenticated drafts, then screen.already_here/already_asked, two api/lookup expanded expected polynomials, and refreshed plausible metadata. There was no identical family; T143/T214/T665/T666/T674/T722/T723 have different quantities or contexts, saved as full documents. Initial lookup syntax errors were corrected with the text parameter and retained in the logs. No existing table was changed. Existing tags number theory, polynomial and L-function were confirmed; proposed Fermat varieties and Frobenius tags remain absent and were not created. Proposal #325 explicitly contributes to wanted #271; #294/#299/#322 are research context, not direct origins. #271's current issue body was checked and archived as issue-271.json.

Checks actually run: 48 exact production rows and all their coefficient native types; all 48 full polynomials via separate exact additive Gauss sums with a different finite-field multiplicative generator and power-residue character lookup; character-free projective counts over F_p and F_(p^2) for all 48. Known controls: cubic surface 7,45 over F_2,F_4; quartic one-minus/all-plus 12,0 over F_5; published cubic-fourfold full polynomial at p=7. Negative controls reject negation, a multiplicity change, reversal and the wrong even-degree model. Degree, constant, exact weight norms, signed reciprocity and coefficient bounds pass throughout. Both methods share Sage finite fields/cyclotomic arithmetic. This build's check is not an independent actor verdict. No arithmetic discrepancy remains.

Selection: argument-selection.md gives all 48 triples and every free index. The full block is about 100 KB; the largest degree-105 polynomial is 10177 characters. dry_run.py's old hardcoded display advisory at 1300 characters is recorded, not treated as an arithmetic failure or used to remove justified rows. The matching current skill uses degree near 100 as a starting point and block readability judgment; this family explicitly anticipated a modest degree-105 excess. Expanded values are long, but all exact factors are retained in computed.json for inspection.

No new theorem about Hodge classes, full cycle lattices, transcendental motives or lifting is asserted. Companions provide model context, not independent arithmetic data. No unresolved convention alternative remains for the selected entries; cubic-fourfold p=2 and higher degree/dimension ranges are uncomputed, not bad fibers or factors equal to one.

Skill evidence: ../20261117T002128Z-build-skills/manifest.json, core SHA-256 981b5f1d6f597ce7396556b3c1067a3b6220cf91beed5c2d9f45d610ce994a2c; matching local/live bundle SHA-256 984965681e2f8925c511a121c118efcde59581c7b0d906cac8b641bf6925e55a. All archived resource hashes were verified before work. Larger artifacts and final byte hashes are listed in evidence-index.json.