Repository navigation
Research: uniform owner-profile optimum through saving exponent 1/3 - #850
Merged
Mister-Meeseeks merged 1 commit intoOct 13, 2026
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PR795 proves optimality of the SW4 owner profile at PR741's admitted endpoint. This extends its packing argument to every real
0 < b <= 1/3, independently of the fixed local paid profile. The selected PR799 construction has the same inventory and attains the optimum. No new κ is claimed.For 72 complete calls, the stated five helper widths, 112 bins per window, and the current four-capacity inventory, every different feasible idle profile has strictly larger
sum I(r) r^(1-b)than{48:2396,44:1,3:1,1:1}. Assignments with the same profile tie. Future local circuit changes can reuse this conclusion if they preserve those packing data and admit the assignment. This removes a repeated owner-search question; it does not exclude changed inventories, dispatch, split helpers, or hosting rules.The new integer-price exclusions remove all two-part remainders
[r,48-r]with33 <= r <= 44. PR795's concavity and eleven-step reduction then leave twelve analytic inequalities. A complete generalized Descartes argument and exact cube-root bounds prove them over the entire interval. The separate checker uses count-vector enumeration, all 147266 capped partitions of 48, and rational Taylor/Bernstein bounds on the continuum. It imports neither the primary checker nor PR795's implementation. Both methods share the explicitly stated packing model and written concavity/period reduction.Expected:
PASS_UNIFORM_OWNER_PACKAGE(standard-library Python 3.11+).Validation passes on Python 3.14.6 and optimized 3.13.11: twelve interval certificates, 1163 independent width patterns, all 36288 archived PR799 owner roles, both archived native paid profiles, nine rejection controls, two boundary controls, portable minimal closure, corrupted-source rejection, existing-output preservation, and selected-record integrity. The source pins identify the actual retained PR799 replay archive. This is a fresh recount of archived data, not a fresh full native construction replay. No full repository suite, new Linux CI, Lean formalization, or independent human mathematical review is claimed.
Only
research/owner-profile-uniform/is added. The selected frontier andupstream/are unchanged. Base:f567884aeb8851b6e794af1184119a3bfbe1e631; OS-random track seed: 99 (research consolidation). Exact sources: PR79556b626800c85026b10dada646c80c6b58179b427, PR79980e6463157cfc80b85af09e7e2f14748fe2dac60.Credit: chafreaky's PR795 supplies the integer dual and concavity reduction, with the PR741/685/660 owner lineage. Jeremy Kothe supplies PR799's X4 and the PR699 SW4 lineage; earlier ends, promotions, gauge, supplier, and recurrence work retain their attribution. The complete notices and inherited all-size/compiler, realization, hosting, precision/tape and analytic assumptions remain explicit. Prepared by Charles Bethin with substantial OpenAI Codex assistance. Apache-2.0.