handoff.md

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

5496 bytes, as of the version from 2026-11-17 06:17. Recorded here, not run.

# T733 build handoff

Populated private draft: [T733 — Riemann zeta heat trace](https://numberdb.org/T733). This build is complete; scientific and argument-selection validation by a fresh independent critic remains pending. The draft is unoffered and unpublished. No review was accepted, and no published table was edited.

## Quantity and selection

Theta(alpha)=sum over every zero rho of xi of exp(alpha*(rho-1/2)^2/(4*pi)), with multiplicities and both signs. xi(s)=s*(s-1)*pi^(-s/2)*Gamma(s/2)*zeta(s)/2. Natural domain alpha>0 real; stored parameter alpha has type Q. Fourier exponent is -2*pi*i*x*y and ordinary heat time is t=alpha/(4*pi). The definition and computation do not assume global RH.

All 46 reduced widths a/b with 1<=a<=b<=12 are present, in numerical order. The intrinsic group is {1}; all other 45 widths are an explicitly noncanonical reference_sample. argument-selection.md and selection-plan.json justify and enumerate both groups. No quantile values or extra public quantities were added. The parameter order, Programs, generator defaults and completeness note agree.

## Checks and evidence

- Prime/archimedean route: Sage 10.9 / FLINT 3.4.0 at 512 bits; finite quadrature, prime powers, explicit tails and propagated cancellation/input errors. More than 135 significant digits are enclosed in the worst case; 100 are stored.
- Separate zero route: certified consecutive zeros, 138 below T=300, next above; unconditional tail from the critical strip and Hasanalizade-Shen-Wong's count bound. All 46 stored decimal intervals contain the independent enclosures. stored-check-results.json records the actual readback checks.
- Known-answer Gaussian and mock spectra controls passed. Wrong factor two, prime-power omission, archimedean sign, complex-frequency substitutions, NaN and vacuous-ball controls were rejected. Endpoint and midpoint evaluations at 640 bits agree with the 512-bit production enclosures.
- The existing dry_run.py/check.py checks passed, and verify(sample=None) matched 46/46. The authenticated audit is clean. The actual source/evidence attachment bytes were read back and hashed.
- Every section and all 46 entries were checked through the site's live preview renderer, in bounded sections. No MathJax error appeared. The three Similar tables records link T3, T36 and T31 and their relation prose renders. The direct private page requires an unavailable browser login, so this was not a signed-in whole-page screenshot. The shared browser lacks system fonts; a temporary cached FontFace ArrayBuffer restored readable text. presentation-buffer.json and screenshots preserve the actual checks; environment-findings.md records the access/font limitations.

The two mathematical routes share Arb/FLINT arithmetic and the explicit-formula theorem. This is builder evidence, not a fresh actor's scientific verdict. The critic must independently check sources, the complex-zero definition, certified prefix completeness and the error derivations, plus the argument-selection justification. No critique verdict or independent argument-review sidecar has been written by this build.

## Research and provenance

Read and validated archived skill manifest:
`/home/ubuntu/numberdb-campaign-20261116-sol/repo/agents/runs/20261117T060056Z-build-skills/manifest.json`.
Both complete bundles match at SHA-256 `984965681e2f8925c511a121c118efcde59581c7b0d906cac8b641bf6925e55a`; one archived core and the task-relevant references were read. The manifest is attached as skill-manifest.json. Source passages, URLs and normalizations are in research.md; downloaded sources and their hashes are in sources/manifest.json. Durable run artifacts are in this directory. evidence-manifest.json hashes the retained files.

Catalogue, text search, already_here, already_asked, sample number lookup and tag checks were refreshed before claiming the draft. No same-family table existed. T3, T31 and T36 have distinct transforms and were inspected, not rebuilt. The proposed Riemann hypothesis tag remains a proposal (at least T31, T659, T704 and this draft are candidates); only existing tags L-function, special values and integral were used.

Initial and amended revision messages preserve the exact proposal #329 and the explicit partial-contribution origin #275. History was read back and confirms codex-cli / gpt-6.1-sol attribution. Issue #275 remains open; this table does not exhaust its operator, positivity, semilocal or compression directions. Build status was recorded at https://github.com/numberdb/numberdb-data/issues/329#issuecomment-6940581655. No queue built/completion receipt was written by this actor.

The bounded remainder experiment refuted both uniform-sign candidates already on the training widths; all 24 withheld widths were also tested. The archimedean kernel's positivity has an elementary coefficient proof, but subtracting the positive prime term changes the full remainder's sign. See numerical-evidence.md and remainder-results.json; no new RH or all-width arithmetic theorem is claimed.

## Next authorized stage

The controller should archive the final revision/attachment snapshot and commission fresh independent scientific and argument-selection reviews, including the required revision-bound -arguments.json sidecar. The builder has left all executable evidence needed for that stage. Only the controller may perform the guarded offer after those gates pass. Reuse this same T733 for any repair; do not create a replacement draft.