Skip to content

Research construction: direct-source E8 with 120 fewer helpers - #697

Merged
Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
swbiggart:research/seed26-direct-source-e8
Oct 16, 2026
Merged

Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
swbiggart:research/seed26-direct-source-e8

Conversation

@swbiggart

@swbiggart swbiggart commented Oct 13, 2026 •

Copy link
Copy Markdown
Contributor

Research construction: a smaller depth-one complex RAM supplier. No new integer-multiplication κ.

This concrete variant of PR647 lets the actual source registers feed the E8 pair/gather core. Each source stays at its port frame until its last pair read, then grows to FULL for the gathers. This removes 120 live input-copy helpers and 720 global recursive children, while preserving the full dirty-restored forward/backward interface.

The complete new profile is m=27, W=7088, R=6848, H={1:40680,2:240,7:20160,8:1080,9:48}, with deficit 24. It admits a Gaussian-RAM saving η=1/14100, with exact rank-mass slack above 11/100. A rational lower bound shows the predecessor's profile fails at that same value, so this is a construction improvement rather than a finer rounding claim. The predecessor's explicit certificate used 1/15000. Existing community suppliers remain stronger overall.

The proof includes source access, helper-only old-response evaluation, inverse-adjoint backward execution, cleanup, center copies in both directions, private workspace, phase adapters, the retained three-stage cover, ordinary padding inputs and the complete all-size RAM cost recurrence. Global scalar ADDs fall from 139456 to 116248. The source's exact complex RAM model is retained; bit precision, tapes, multiplication assembly and a new Fourier transfer are outside scope.

Reproduce (Python 3.9+, standard library; do not use -O):

python3 -B research/e8-direct-source/verify.py --check research/e8-direct-source/RESULTS.json
python3 -B research/e8-direct-source/test_verify.py

Expected: PASS_E8_DIRECT_SOURCE: 7088 columns in both directions; 120 fewer helpers; 7 negative controls, then three tests OK.

Validation:

  • Every one of 7088 independent live columns in both directions, every chronological frame, 3601 local phase projectors, and the final rank-two phase lowering.
  • Exact full invoices and rational moments, seven negative controls, plus 48 separate modular physical executions with actual Gaussian phase/center transport.
  • Receipt and tests pass locally on Python 3.9.6, 3.11.16 and 3.13.16. Optimized Python is explicitly rejected because inherited phase utilities use assertions.
  • Source pins, staged immutable bytes, selected-record integrity, workflow syntax, whitespace and clean archive replay checked. Scoped Linux CI passed on Python 3.11 and 3.13 at submitted head 3219863d868b4230eeb4bf2948758eff45efffcf, including the full receipt, three tests and clean-tree checks. This is separate from broader repository CI.
  • No full repository-suite pass claimed. The previous full run stopped at missing Boost headers in the unchanged selected bit supplier; that expensive suite was not repeated for this isolated package.

Base: 66849e1475c930fd7732b1a0067ffee7e1e95f60.
PR647: 956bb581e7f97328e1b53baa5aca3d2490f2f607.
Assignment seed 26, complex supplier; source port 26 supplies the adversarial read controls. No pinned coordination issue existed. The local claim has one result update. Overlap checked through #692 (512 open PRs); the predecessor head is unchanged. PR682 audits the old tensor profile and does not implement this rewrite.

Credits: Artus / @artuskg's PR647, its OpenAI Pro provider proof, Codex verification and all retained E8/community predecessors. The predecessor's phase/binary-arithmetic utilities and notice are vendored unchanged and authenticated; this reuse is explicit, not described as a wholly independent checker. Prepared by Scott Woolsey Biggart (@swbiggart) with substantial OpenAI Codex assistance, under Apache-2.0. No human-only review or formal proof is claimed. No existing file, upstream source or selected-frontier pointer changes.

@Mister-Meeseeks

Copy link
Copy Markdown
Contributor

100-PR batch review at 3219863d868b4230eeb4bf2948758eff45efffcf.

This remains open after scoped triage. Next review step: Uniform RAM supplier/interface proof needs dedicated review.

This is a specific continuation point, not a rejection or a request to redo already recorded checks. Completed local evidence is retained separately from outstanding whole-head obligations.

Batch review, attribution and evidence. Thank you for the contribution.

@Mister-Meeseeks
Mister-Meeseeks merged commit 1c1de25 into CrocSwap:main Oct 16, 2026
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