numerical-evidence.md

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

3668 bytes, as of the version from 2026-11-18 17:13 (current). Recorded here, not run.

# T754 numerical evidence

Exact integer polynomials are stored; decimal precision and uncertainty padding do not apply. There is no floating reconstruction. Generator values are native Sage integer polynomials and every coefficient is checked to have Sage Integer type. Sage was initialized through numberdb.sage before named ring imports.

Actual invocation (builder only):

    agents/sage.sh offline.py generate.py checks.py agents/table-build/dry_run.py agents/table-build/check.py

All paths were the run's files mounted by agents/sage.sh. offline.log retains the complete invocation output, including flushed per-prime checkpoints. checkpoints.json and computed-values.json preserve the computed results; no long-running work remains. dry-run.log contains the unchanged repository deterministic checks: 612 entries, no exactness/import/prose faults, longest polynomial 20 characters, measured block 35.2 KB.

Production: finite-field quartic character sums in Q(i); two exact binary sums assemble Griffon's four-character zero-sum normalization; coefficient twist and p-orbit factors produce degree two. Complexity is linear in the field order per prime, cached across coefficient models. Exact conversion into ZZ is required.

Distinct mathematical check: checks.py, runnable under ordinary Python, independently enumerates arguments, locates primary Gaussian primes by norm and congruence, computes the modular quartic residue of the coefficient product and obtains the trace of the twisted Gaussian square. It checks both conjugate Gaussian-prime choices. Every coefficient of all 612 outputs agrees. The inert formula checks 330 rows; split regimes fill the remaining 282. These implementations share the mathematical source and the characteristic-zero summand, but not the character-sum arithmetic. This is builder-authored evidence for a subsequent independent critic, not that critic's verdict.

Positive controls: all-plus at p=5 gives 1+6*T+25*T^2; (1,1,2) gives 1-8*T+25*T^2; one-minus gives 1-6*T+25*T^2; all-plus at p=3 gives 1-9*T^2; all-plus at p=13 gives 1-10*T+169*T^2. Direct affine fourth-power counting, with the zero vector removed and division by 4, gives 0,12,64 projective points over F_5 for the three respective sign/twist controls. Point counts fix the surface convention, not the rank-two factor in isolation.

Negative controls rejected: altered linear coefficient (612/612), wrong Tate scale in quadratic coefficient (612/612), determinant reversal (612/612), positive inert determinant (330/330), missing nontrivial twist (148 distinguishable rows). A missing twist is deliberately not claimed distinguishable where the correct twist fixes the polynomial. check-results.json contains every comparison and the actual control results, with no failures. Direct double sums over fields of orders 5,9,13,49 agree with the binary reduction; the fields cover both prime and quadratic-extension branches.

Every row also satisfies constant term 1, degree 2, determinant +/-p^2, |linear coefficient|<=2p and signed reciprocity. versions.json records Sage 10.9, Python 3.12.3, numberdb 0.1.10, cypari2 2.2.4, gmpy2 2.3.0, numpy 2.4.4 and sympy 1.14.0. The measured offline phase after imports took about 1.1 seconds.

Skill evidence: skill-manifest.json, copied from ../20261118T160132Z-build-skills/manifest.json; matching complete bundle b5c3c7985b7c9e178b4899d083a9e3bc4284989dff76f05e5b0e52835d47eed2. Source locations and artifact hashes are in research.md and source-hashes.json. Stored-value verification and audit will be recorded separately in the handoff. No scientific/argument acceptance verdict or offer is made by the builder.