This directory preserves the frontier work produced during the 2026-09-06 continuation of hunts/prime_pair_error before the next agent phase.
Reconciliation, 2026-09-13
The missing source files from PRs #196, #199 and #200 have been restored: factorial_direct_bn/, factorial_full_cost/, certificate_route_test/BASELINE_REVIEW.md and its review/ directory, and joint_correction_candidate/ with its original archive and review. The four associated tests are restored too. These reviews cover only the objects they name, not the later adaptive packages. See ../../RECOVERY_2026_09_13.md for exact source commits and the distinction between archival integrity and mathematical review.
Canonical status
- The parent hunt remains the canonical public record for established/reviewed work.
- PR #189 has been repaired, retargeted to
main, and merged; its independent referee corrections are now canonical. - The conversation-era frontier drafts preserved here are not promoted results unless a later independent review says otherwise.
- None of the new drafts currently improves the completed upper bound on total CHHL
E(N).
Exact handoff archive
archive/zeta_frontier_handoff_2026-09-06.zip is the lossless handoff archive, committed directly as a binary. archive/SHA256SUMS carries its hash, and tests/test_frontier_archive.py checks on every run that the file is present, that its SHA-256 matches the value below, and that every member listed in the archive's own SHA256SUMS.json is inside it with the recorded hash and size.
Verify from the repository root:
sha256sum -c hunts/prime_pair_error/frontier/2026-09-06/archive/SHA256SUMSExpected SHA-256:
61f4901f8659d13cd2c795b560475b1313db666650da78dabeccd7e03c1807de
Correction, 2026-09-06. The first version of this section described archive/ as a set of base64 chunks (part-*.b64) written through the GitHub text API, with a reconstruction command. That upload never landed: archive/ did not exist on main or on any branch, and the documented command produced an empty file whose hash is the hash of zero bytes. The archive was re-supplied from the originating session's attachment and committed as the binary above; the hash it was expected to have is the hash it has. The test exists so that a documented archive can no longer be missing without the suite saying so.
The archive contains the original reverse-engineering experiment, Möbius branch, q3 follow-up, zero-energy feasibility draft, sharp-transfer draft, multiscale draft, direct central arithmetic attack, runnable checkers/results, and source text used for the referee/Möbius handoffs. Its own MANIFEST.md and SHA256SUMS.json inventory the members. One line in that manifest is stale by construction: it records PR #189 as closed unmerged, which was true when the archive was sealed and is not now (see PROVENANCE.md). The archive is evidence and is not edited to catch up.
Next phase
Before promoting any conversation-era analytic lemma, independently review it. The immediate mathematical frontier is the central prime-counting obstruction: obtain an unconditional upper-bound mechanism that rules out a coherent R(N)=psi(N)-N excursion larger than the RH scale, without merely restating RH or using an aggregate identity that annihilates zeta-zero modes.
Factorial-certificate pilot (preserved and reviewed 2026-09-06)
factorial_certificate_pilot/ preserves a separate downloaded attachment, factorial_certificate_pilot.zip, hash-pinned in its own archive/SHA256SUMS and by tests/test_factorial_pilot_archive.py. It is a Chebyshev-type factorial upper certificate for psi(N) with an LP-chosen seed: 87 exact-rational certificates, all reproduced. The independent review is factorial_certificate_pilot/REVIEW.md; the review's own checker and its output live in factorial_certificate_pilot/review/. It is a pilot, not a prime-counting record and not an RH result, and it does not touch the A/B referee record above.
Cumulative research checkpoint and the three newest packages (preserved 2026-09-06)
checkpoint/ is the research checkpoint: checkpoint/CHECKPOINT.md is the narrative (chronology, claim grades, failed methods, open questions), checkpoint/README.md says where every original byte lives, and checkpoint/archive/zeta_research_checkpoint_2026-09-06.zip is the cumulative archive itself, committed unchanged and hash-pinned by its sidecar and by tests/test_research_checkpoint_archive.py. It bundles all twelve original research ZIPs, including the two already committed above, and its own verifier.
The three newest packages it carries are imported beside it as runnable working directories, byte for byte: certificate_structural_step/ (omitted-prime obstruction and the period-30030 extension), certificate_refinement_rule/ (positive repair rule, five applications, full error budget) and certificate_route_test/ (fixed-recipe ceiling and the combined-weight repair; its aggregate_results.json is the current candidate baseline). They are archived, unreviewed source packages, not promoted results, and none of them bears on RH. Rerun them into new output files; the recorded outputs are the originals.
Adaptive correction block freeze, and the research pause (preserved 2026-09-06)
adaptive_freeze_v1/ is the checkpoint taken when Thomas paused the research. Its README.md is the repository's account of what landed; adaptive_freeze_v1/provenance/FREEZE.md is the originating session's narrative, copied unchanged. The archive itself is adaptive_freeze_v1/archive/zeta_adaptive_freeze_v1.zip, committed unchanged, hash-pinned by its sidecar and independently by tests/test_adaptive_freeze_archive.py:
a264da2d6da781a3133ab75557f22842fe2581e210c7a628c0fddeff0c653681 (4,185,414 bytes)
It bundles seventeen original research ZIPs, which include the three already committed above, and its own verifier. Three further packages are imported beside the earlier ones as runnable working directories, byte for byte from the nested originals: joint_support_analysis/ (why the inherited repair dictionary freezes the weight at 20), repair_cost_rule/ (the repair-potential bound and the accepted direct steps at 20 and 21) and adaptive_correction_block/ (23 masked reciprocal-carry corrections, 82 positive repairs).
All three are unreviewed drafts. The reviews in PR #199 and PR #200 cover only the earlier objects they name: the combined-weight baseline and the joint correction candidate. Neither covers the support-obstruction diagnostic, the repair-cost rule or the adaptive block, and the producing session's own checks.json records its result as a separate-code self-check rather than a review. Storing them does not promote their mathematics, the reported reduction is a fraction of one construction's leading excess rather than progress on RH, and nothing here improves the completed upper bound on total CHHL E(N). The rejected third step at 20, the direction test that is not the block's baseline, and the envelope-oriented alternative candidate are all preserved alongside the accepted state.