Skip to content

About

No description, website, or topics provided.

Resources

Stars

5 stars

Watchers

0 watching

Forks

Repository files navigation

Ordered binary matrix removal

Candidate 3×3 obstruction · Main theorem checked in Lean · October 12, 2026

Dated progress in ordered binary matrix removal, from OpenAI’s 66×66 pattern through ProveIt’s 5×5, Erebos Labs’ 4×4 and this project’s 3×4 to 3×3. Time is proportional in UTC and pattern entries use a logarithmic axis.

We give a construction for the fixed ordered binary pattern

0 1 1
1 0 1
1 1 0

The 3×3 pattern has nine entries, down from twelve in our previous 3×4 construction and sixteen in the earlier external 4×4 results. Arbitrarily large matrices can contain extremely few copies of this pattern while still requiring many edits to eliminate it: no polynomial removal bound holds. Copies preserve separate row and column orders and match both zeros and ones. Repairs may change any entry in either direction.

Read the paper, Markdown manuscript, and Lean theorem. For every c,C,ε₀>0, the theorem gives a fixed 0<ε<ε₀ such that, for every size threshold N, some square host of order n≥max(1,N) requires at least εn² edits to become pattern-free and has fewer than c ε^C n⁶ copies.

What is verified

The full qualitative theorem for the 3×3 pattern is checked in Lean. The source-bound audit covers 92 project modules, 573 named theorems and 2,163 declarations, including inherited infrastructure. A separate AI reviewer found no gap in the written proof. Independent mathematical peer review remains pending; we make no global minimality or first-ever claim.

The written and formal proofs use two versions of the host, with m=2^h:

Version Anchor size per mode and axis Host order n Copies Repair lower bound
Written proof and separate-agent review m (8h+2)m exactly 7m⁴ m²
Lean certificate 2m (14h+2)m at most 2mn⁴ m²

Both establish the same no-polynomial-bound theorem. The exact 7m⁴ count and the smaller host's quantitative parameters are outside the Lean certificate. See formal scope, the audit, separate written-proof review, and model alignment check. The original proof snapshot is preserved unchanged for its review hashes; the manuscript gives the current verification status.

Baseline and attribution

The tree, threshold entries, protected sampling and propagation originate in OpenAI's 66×66 construction and formalization. ProveIt supplied earlier 5×5 work; Erebos Labs published a 4×4 theorem and Lean proof before our 4×4 PR. Zinan Wang's 4×4 manuscript also gives the exact copy count found in our earlier quantitative refinement, with a different mode order.

Our 3×3 construction uses one anchor on each axis, suffix column anchors, vertical-dummy ones, and a new all-copy classification. See BASELINE.md and ATTRIBUTION.md for pinned sources, verification stages and the bounded novelty search. The preceding 3×4 certificate and 4×4 refinement remain available.

The chart uses dated GitHub records, with proportional UTC spacing. Erebos Labs' v1.0.0 release was published October 9, 2026 at 12:40:51 UTC. The line follows the smallest entry count among the selected records, not a complete priority history. See timestamp sources and scope.

Reproduce

python3 scripts/run_checks.py
# With the pinned Lean/Lake toolchain installed:
python3 scripts/run_crown_formal_checks.py
python3 verification/check_crown_formal_model.py --output .context/crown-model-check.json

The default exact checks use Python 3.9+ and the standard library. They include two seven-depth censuses, a separately written reviewer checker, repair guards, Boolean propagation, and adverse variants. Optional SAT discovery and repair searches are documented separately. Finite checks support the proof; the Lean certificate establishes the universal theorem. See verification scope, RESULT.md, RUNLOG.md, and HANDOFF.md.

Build the paper with python3 scripts/build_manuscript.py; add --regenerate-tex with Pandoc 3.6.4 installed to regenerate LaTeX first. The Markdown manuscript is the mathematical source; the template and preamble control the title, abstract, typography, and pagination. The pdfLaTeX build requires the packages listed in the preamble, including microtype, parskip, needspace, xurl, and bookmark. After editing the Markdown or template, regenerate LaTeX; python3 scripts/build_manuscript.py --check-tex checks that the generated source is current. PDF links open repository files online. The bibliography supplies the manuscript's [@key] citations. Pandoc's built-in citeproc generates linked numeric citations and references using the citation style; unresolved citations fail the build. Regenerate LaTeX after changing either bibliography file as well. The improvement record records development, not priority. Prepared by Ryan Shea with Codex assistance. Licensed under Apache 2.0; see NOTICE.

About

No description, website, or topics provided.

Resources

Stars

5 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages