Skip to content

Certify a one-layer full-cover QQ16 F0-to-B1 exact grid - #3270

Merged
Mister-Meeseeks merged 3 commits into
CrocSwap:mainfrom
petalas:research/worker3-batch3-20261017T1956-4608d343
Oct 17, 2026
Merged

Mister-Meeseeks merged 3 commits into
CrocSwap:mainfrom
petalas:research/worker3-batch3-20261017T1956-4608d343

Conversation

@petalas

@petalas petalas commented Oct 17, 2026

Copy link
Copy Markdown
Contributor

Summary

bounded dyadic inputs across the whole owner cover
  -> literal F0, carrying all 712 field bounds
  -> literal B1, including native interiors and 528 cover moves
  -> sufficient quantum h*2^-71 and signed width P+215

Supplies the B1 numerical-reference obligation left outside PR3226's F0-only grid proof. Exactly one 73-axis tensor layer is covered, with N = 2^73 per owner. The 512 normalization describes the local kernel. The full payload is not allocated or executed; native bridge and exact endpoint/restoration semantics remain inherited premises.

Symbolic per-owner bounds are 483082 N + 12 W-bit word operations, plus 2112 N conceptual cover transfers and 1426 N + 12 numeric words. Physical routing, code/control storage, recursive tariffs, an admitted entry contract and the F17 consumer remain unpriced. No multi-layer, sharpness, full functional/native admission, supplier, selected-frontier or kappa claim.

Evidence

  • Before: PR3226 proves an exact grid for F0; PR3235's two-stage completed-error theorem leaves finite register widths and native interiors open.
    After: PROOF.md gives a uniform component induction through the literal F0→B_1_2 word, including actual carried state and 73-dimensional cover bijections.
  • Registered Python 3.12 replay passes all three commands: 53,118 events, 270,288 primitive-bound observations, separate rational/integer agreement on every trace and operation count, and seven optimized behavioral groups. Ruff and formatting pass.
  • The reviewed 18-file read-only, Git-free export passes all three direct scripts under optimized Python 3.13. Final publication adds only the exact REVIEW.json receipt: all 18 common files are identical, and all 19 publication-byte/export checks pass. Required predecessor bytes are retained and authenticated; replay needs no pending-PR Git objects. Exact input/provenance usage mapping rejects both input-reclassification mutants.
  • The assigned internal read-only peer review and proportional source-authentication recheck are recorded in REVIEW.json. This is not independent human or proof-kernel review. Full-payload execution, multi-layer powering, inherited bridge/endpoint replay, selected cold replay, Lean, Docker and hosted CI are unrun.

Package: research/qq16-f0-b1-exact-grid/, with exact sources, receipts, reproduction commands, limits and Open next questions. Canonical investigation. Base 463113d675af7410ac91d53a4fed51ec6494eb9a, priority revision 6; submission guidance and scientific inputs are unchanged.

Credit commdata2338/PR3226/3235, Sricharan Vagvala/PR3101, Nick Petalas/PR2965/3250/3258, Ali Romman's joint-query construction, Tekalign's native conventions, Ross Richardson, Arham Tahir and the retained supplier/bridge lineage. Frozen predecessors and the selected record are unchanged.

Merge Danger

Door: two-way

Blast Radius: research

One isolated reference-contract package with opt-in bounded replay. No shared code, frozen inputs or selected-result changes.

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