Repository navigation
Improve conditional κ to 0.0016805 with selective dirty borrowing and physical owner repacking - #2260
Conversation
|
Maintainer replay at The registered complete command currently stops in All other expected receipt fields match. Replayed Please supply the original generated payload (or explain/reproduce the original hash), then change cross-environment replay to compare its decompressed/canonical payload hash, preserving the original source commitments. A negative control that changes a factor's family/axis while retaining aggregate counts should still reject. This is a reproducibility hold, not a mathematical rejection. The optional Lean kernel claims have not yet been independently rerun in this maintainer pass. Follow-on cross-check: #2265 already records this same mismatch and publishes a decompressed-payload pin. Its |
2 participants
Improve conditional κ to 0.0016805 with selective dirty borrowing and physical owner repacking
This adds a new exact finite complex supplier and obtains κ=3361/2000000=0.0016805 under the project's retained all-size and analytic interfaces. It exceeds the highest competing open conditional claim in the recorded review (PR #2220, approximately 0.0016797549561255445) by approximately 7.45044e-7, or 0.0443543%. The gain changes the actual paid construction rather than refining digits of the same supplier. Selected-result pointers remain unchanged.
Starting from Cinna Davis's PR #2226 anticipatory borrowing architecture, the new 44-alias selection retains 436 helper roles. The package supplies both complete native orientations, exact dirty restoration/private cleanup, actual helper and data endpoint adapters, and a new physical owner packing for all 72 calls of the T8 parent. Its complete profile certifies complex saving b=0.0016834 in two independent rational engines, including the conservative retained bad-residue surcharge. The retained bit rail and seven-level ordinary bootstrap then pass all 47 strict outer inequalities at the stated κ.
The changed controller has 31,392 literal helper assignments and 7,150 actual 73-column charts. An independent implementation checks 387,072 family-tagged Gaussian factors, keeping every retained translation and exact phase. Fresh exact-dyadic work/stock/precision accounting charges 7,745 physical families, magnitude guard 244λ<512λ and row coefficient 16467<20161. This does not use cross-family rank reassignment or introduce an intermediate rounded-child assumption.
Validation: complete isolated optimized Python replay; 5,992 exact native jobs; all 660 public columns and both 76-CLEAR owned-private postludes; physical chronology/charts and factor partition; finite invoice; two independent exact moment engines; all 47 strict outer rows; local/ownership/phase/direction/cross-family/arithmetic fault controls. Eleven optional universal lemmas additionally pass fresh Lean 4.33.0/Mathlib v4.33.0 kernel and axiom checks, using standard axioms only. The full multiplication result is not formalized in Lean. Local Python was 3.11.9; Python 3.12/hosted CI is not claimed in advance.
Reproduce from the repository root:
The self-contained 70-file package includes exact inputs, replay scripts, proofs, immutable source/file pins, preserved licenses/notices and CI registration. It imports no personal Lean tree. Scope remains conditional on the retained all-size exact compiler, physical fixed-tape execution, completed ordinary/bit interfaces, recovery and analytic transfer. Substantial OpenAI Codex assistance by one primary and three agents; no independent human review claimed. The frozen package records the review through PR #2251 at 2026-10-15T02:15:55Z. A final fully paginated refresh through PR #2258 at 02:32:48Z found no higher κ claim; each newly changed PR body was read. This is metadata screening and targeted relevant code review, not a claim to have read every diff/discussion.