Skip to content

Bind the retuned ordinary rate to the complete T74 layer - #3576

Merged
Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
kumagi:research/t74-retuned-ordinary-layer
Oct 20, 2026
Merged

Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
kumagi:research/t74-retuned-ordinary-layer

Conversation

@kumagi

@kumagi kumagi commented Oct 18, 2026

Copy link
Copy Markdown
Contributor

PR3562's stronger ordinary rate cannot be substituted directly into PR3477's T74 proof: tau=332353/332929 is below the complete T74 moment threshold, reversing the old summation inequality. This package closes that consumer step under the retained source/physical premises. It binds Profile(1,577) to an opt-in T74 factory and proves conditional layer time O_q(V d^.998273), retaining the full child histogram, exponent-1000 stopping, root domain/reservations and storage/endpoint contracts. The layer saving improves from .001695 to .001727; this is not a new multiplication exponent.

The complete moment at sigma=.998271 passes two exact engines. Internal work uses the tau<sigma tree branch; stopped leaves cost exponent .998272729, leaving exact margin 271/10^9. Both the direct ordinary-moment substitution and one extra rank-one child fail. A separate rational toy tree checks the reversed inequality, strict stop and leaf invoice.

The production constructor preserves the retained native construction AST except its profile. A private T74 binding alias avoids a real short-module-name collision with the native catalogue. The complete entry point has constructor-recording controls; all 35 files from PR3562's exact backend export are byte-identical in the authenticated T74 archive closure. PR3562's theorem/interface/manifest are retained without depending on its unmerged Git object at replay time.

Validation:

  • Registered Python 3.12 checks with -I -O -Werror -B, including nine factory rejection controls and direct/fast reduced root returns.
  • Normal/optimized exact rate checks and metadata/source pins. Publication 1630afea0fc1083717b5b2799b8de7a819a6ea02 passes committed-byte/source-history checks for 32 files. Both Git-free export commands reproduce identical receipts; all 33 regular export files (including manifest) remain unchanged after replay.
  • Reduced root checks: four direct packets and one fast packet, five balanced recursive returns. Ordinary run counts are [0,0]; these verify object installation and root behavior, not the fitted ordinary branch or production GL64/T74 execution.

Remaining limits: ordinary/C0, native/common-label, hosting, height and physical refinement remain inherited conditional premises; the all-size PR3562 theorem is not independently reproved. This retains d=floor(p^.997) and does not import PR3493's smaller dimension plan or complete its analytic/error/packing invoice. No selected-frontier or shared-default changes.

Started from canonical main 20ecf2891d80b5acf18d980b2164fb44ea7faa90, priority revision 6, taking the explicit PR3562 handoff. The pre-submission refresh found a concurrent actual-width/deeper-stop effort; this independent contribution retains the old root/stopping and does not duplicate that broader scope. Coordination and overlap: canonical claim/result. The independent-moment scope change is recorded in TASK.json; no new lowering search or supplier admission is claimed.

Credit: Jose Manuel/PR3562,3477,3385 and physical-backend lineage; Marius-Paul Dumitrean/PR2979 and Jeremy Kothe's predecessor work; Douglas Colkitt and Nick Petalas/PR3342 exact interval engines; David Arias, Adam Gall, Gareth Davidson and the retained ordinary/native/compiler/hosting notices. Hiroki KUMAZAKI (kumagi), with substantial OpenAI Codex assistance in proof, code, checks and prose. Apache-2.0.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants