Skip to content

Exact native gauges: 698 → 674 scalar allowance, with a zero-scratch option - #4499

Merged
Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
Dugongue:research/exact-unit-gauges-20261021
Oct 22, 2026
Merged

Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
Dugongue:research/exact-unit-gauges-20261021

Conversation

@Dugongue

Copy link
Copy Markdown
Contributor

Four authenticated native cleanup blocks cost 766 scalar allowance per orientation in main. The rank-factor compiler in #4494 reduces that to 698. This package adds an exact power-of-two gauge optimizer: it reaches 686 with no additional owned fields, or 674 with one reusable owned field by selecting the cheaper of gauges and the inherited rank-one sum macro. Both inverse-transpose orientations are checked, including every scale/unscale and owned-field clear.

The new reusable result is a sufficient exact replacement for any disjoint common-chart dyadic block, with a feasible signed circulation certifying the optimum within the declared gauge family. This is classical min-cost-flow duality specialized to the native scalar price; no unrestricted circuit-optimality claim. The rank-one alternative and source-window fixtures are credited to #4489/#4494. See research/exact-unit-gauges/ATTRIBUTION.md for upstream compiler contributors and MathProof discovery precedents.

Validation: source-byte pins and actual contiguous native window extraction; 320 exact symbolic-column checks across both modes/orientations; 100 deterministic random cases; 10 bounded exhaustive cross-checks; rejection controls for altered flows, missing restoration/clears, illegal scalars and shifted selectors. Registered replay, isolated optimized replay, committed-byte checks and a Git-free exported replay pass. Standard-library Python only; tested on Windows, Linux execution not available.

python3 -I -O -B research/exact-unit-gauges/verify.py
python3 scripts/research_package.py --package research/exact-unit-gauges

Scoped partial contribution, no new κ, no new recursive jobs, no productive-temporary credit and no selected-frontier change. Full rebuilt stock/norm/coefficient and multiplication-bound replay remains necessary before using the savings in a global bound. Canonical main was refreshed at the submission checkpoint and remained 2a13eb1827784206202db90e5dc8327cf8a250db. Recent visible PRs through #4495 were checked; cached overlap coverage is incomplete, with remote cache/API failures recorded explicitly. No novelty claim follows from that incomplete coverage.

Prepared for Dugongue with substantial OpenAI assistance; Apache-2.0. The package contains the proof, executable programs, source pins, certificates and open next questions.

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