handoff.md

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

7069 bytes, as of the version from 2026-11-19 23:07 (current). Recorded here, not run.

# Build handoff: T766

Table: https://numberdb.org/T766, Primitive odd modular periods of rational
newforms. Owner: zeta3. Run: 20261119T222759Z; configured engine
codex-cli, model gpt-6.1-sol, effort xhigh. The populated table is a private,
unoffered draft. The controller owns the independent scientific and argument
critique, any bounded repair, final family review, guarded offer and queue tick.
No publication, offer, human review action or unrelated table edit was made.
The initial history message cites proposal #349 and contributes-to wanted
#262 exactly as supplied; neither #303 nor #309 is invented as an origin.
No issue was closed. The queue was already claimed by this campaign.

Revision after fill: 5470, digest
444bc11be39f6a647db2468e9d31b22e5c60cd496779d014f8bcc5446c9d4dd9.
The final attachment list and live review token are recorded in
final-metadata.json after the final upload. The website remains authoritative.
Selection fingerprint (not an independent verdict):
1fb1c093f739944050e3e3eee32abbec65c409bacbb8f3fd561a991bf1721765.

Quantity: the signed real Omega_f^- in r_f^-=Omega_f^-*P_f^-, with upward
identity-coset path 0 to i infinity, parity average (r(X)-r(-X))/2,
primitive integral positive-leading polynomial, normalized a_1=1, trivial
character, exact level N, even weight k>=4, rational Hecke field and nonzero
odd part. Parameters remain N,k,i in that order; i counts all real embedded
newforms lexicographically before rational filtering, displayed as i_f.
There is no extra phase, parity, conductor factor or magnitude convention.
Delta has P=4X^9-25X^7+42X^5-25X^3+4X and
Omega=10*L(Delta,2)/(4*(2*pi)^2).

All 18 selected forms have one 100-significant-digit real entry: six at
level one (weights 12,16,18,20,22,26), twelve at weight 12 across exact levels
2 through 10. The domain is the natural family; complete is no. Noncanonical
finite cutoffs and every discrete parameter are justified in
argument-selection.md, with no continuous coupling or decimal grid. Exact
space decomposition includes empty spaces and all excluded nonrational forms.

Generation: PARI mfperiodpol at 160 and 220 decimal working digits plus
32 guard bits (564/763 total bits), with exact Sage modular-symbol primitive
normalization. Rigour is conservatively heuristic (agreement-checked).
Sage 10.9, PARI 2.17.1, Python 3.12.3, numberdb 0.1.10 are recorded in
dependency-versions.json. Source caches and verified hashes are in
source-hashes.json, with concise inspected passages in research.md.

Completed build checks:

- All 21 cusp/new-space dimensions agree with independent formulas, all 33
  embedded indices agree between Sage and PARI and the exact companion roster,
  and every selected primitive shape matches live published T696 (revision
  5103, digest 5ec929375909ff122b536c298311d5b8c5639e122b9434da1f613bd16aab8d49).
- For each rational form, 600 exact Fourier coefficients agree between Sage
  and PARI. A separate eta-product convolution checks 100 Delta coefficients.
  Exact Fricke normalization is checked against PARI, including the factor
  N^(k/2-1) in Sage's raw operator.
- A separate explicit split Mellin implementation using Sage Fourier
  coefficients and ball arithmetic checks all 111 odd coefficients at both
  150-digit/450-term and 190-digit/600-term settings. Positive tail bounds
  derive from the normalized-newform coefficient bound (Theorem 2.6 cited
  in research.md). Every nonzero coefficient-derived scale lies inside its
  stored decimal interval. Worst residual <0.48 stored ULP; narrowest
  independent scale accuracy 520 bits. Exact zeros are checked as well.
- Three known nonzero exponential-integral controls, Delta's exact primitive
  shape, wrong-sign/doubled-scale controls on every form, and rejection of
  NaN/wide balls exercise the checker. No numerical result from T765 was
  reused as evidence. No unresolved mathematical disagreement was found.
- Existing dry_run.py and check.py pass: native RealInterval/Fraction return
  types, finite non-singleton intervals, imports, suspicious exactness and
  size. Serialized values occupy about 3 KB, longest 105 characters; native
  interval displays occupy about 9.1 KB with a 472-character maximum.
- Authenticated verify(sample=None) matches 18/18, with no missing/extra
  values. Fresh independent 190-digit/600-term Mellin checks pass on every
  stored value. Live ordered form identities match T765, T695 and T696.
  See postfill.py, postfill-evidence.json and postfill.log.
- The authenticated audit has only a 331-character definition-length
  advisory. audit-assessment.md explains why the single defining sentence
  is retained; no genuine fault was reported. All four suggested live tags
  apply and are retained without departures.
- Server-rendered preview sections were browser-inspected: 18 number cells,
  all six intended Similar tables records/relations, every section, correct
  citation targets and the displayed PARI program; no math or browser error.
  Screenshots and presentation-evidence.json are preserved. Direct private
  page access requires an authorized website session unavailable here (404
  without one); the controller/critic should inspect it with such a session.
  The composite is not claimed to be the directly visited private page.
- history.html and history-evidence.json preserve #349/#262 citations and
  codex-cli/gpt-6.1-sol attribution. Initial citations were never dropped.

The single full offline computation is checks.log, with resumable per-form
CHECKPOINT records and final JSON artifacts. computation-provenance.json binds
source hashes. Generator and checks are attached, together with all JSON input
artifacts needed for reproduction. Run on an ordinary Sage installation with
numberdb and all attachments beside the scripts:

    sage -python checks.py
    sage -python generate.py

In this deployment use agents/sage.sh, supplying the named source/JSON files
as in the retained run commands; use only an API-key pipe for draft reads.
checks.py writes no website state; the generator defaults to verification.
No removing=True update was used.

Skill evidence: the coordinated archive
agents/runs/20261119T222759Z-build-skills/manifest.json is attached unchanged
as manifest.json, SHA256
7629363ce07d2e1340528294fe639e6fc7a5b18da808ec52558687fc1db91534.
Both valid local/live bundles and every resource hash match; bundle SHA256
b5c3c7985b7c9e178b4899d083a9e3bc4284989dff76f05e5b0e52835d47eed2.
Core SHA256 b79b7fbc9fa3ce63f2d642cccdac12b0d214757ab01db66c03a2d84264a58e5f.
Core plus all six task-relevant references were read. No source, skill, prompt,
shared toolchain or campaign receipt was modified. Environment-only findings
are in environment-run.md, not proposed as contributor skill changes.

These are builder checks, not the independent revision-bound scientific
verdict or argument-review sidecar. Those are still required before offering.
The sibling T765 was not edited; its reverse link is a controller/owner
follow-up if desired, not permission for this actor to amend it.