This repository gives a source-faithful, computer-assisted proof of Erdős Problem 848 for every positive integer, together with the exact certificates, independent checkers, source audit, failed approaches, and reproduction instructions used to verify it.
For
The bound is sharp: the class
- Mathematical theorem: the all-$N$ result is proved by a gapless union of an exact finite coloring certificate, explicit structural interval certificates, and an audited high-range theorem.
-
Computational evidence: every decisive finite stream has an independent
exact checker, authenticated inputs, complete interval semantics, and
negative controls. One fresh, uninterrupted run of the six-stage root
orchestrator has also passed and received an independent receipt audit.
The canonical evidence is recorded in
certificates/. -
Formal evidence: the pinned ART-006 Lean development has a literal,
diagonal-inclusive endpoint and a coherent noncircular source/interface
audit. Its own publication state reports a closed theorem with only
propext,Classical.choice, andQuot.sound. -
Remaining project-level gate: this repository has not independently run
ART-006's complete 30,636-module provider build and trust-zero replay from
source on the required Windows capacity. No qualifying host is currently
available to the maintainer; the reviewed but unexecuted route is tracked in
issue 2 and
diagnostics/STATUS.md. The remaining work is formal and clean-reproduction assurance, not an uncovered value of$N$ .
The precise status of every dependency is recorded in
docs/proof-ledger.md and
docs/proof-dag.yaml. No static check, preflight, or
partial replay is presented as a completed theorem gate.
- Final proof in Markdown
- LaTeX source
- Compiled PDF
- Structural-certificate appendix and verification challenge
- Exact problem and statement audit
- Internal adversarial review record
- Fresh six-stage root replay and independent receipt audit
- Complete reproduction guide
| Closed range | Method | Decisive evidence |
|---|---|---|
| Exact compatibility-graph colorings at every benchmark endpoint | Independent C++ stream replay and exact leaf-primality certificate | |
| Least-witness structural rank bounds, including the full |
37 exact rows, independent transcript primality audit, all other outsider branches | |
| Exact-rational short-shift envelopes | Authenticated ART-005 rows and semantic mutations | |
| Exact-rational rank envelopes | 1,255 complete multiplicative blocks | |
| Explicit analytic theorem | Source-pinned Sothanaphan proof and exact-rational numerical audit |
The first two ranges overlap; every later pair shares its displayed endpoint.
Their union is all positive integers. The canonical range-ledger SHA-256 is
b28760bca88b3f4a356f5212f5aa3711df00ee527606058a9aefc193e715ebe1.
Start with the quick controls. The root-manifest mutation suite deliberately fails closed unless the decisive high-range source is present and authenticated, so retrieve that one ignored input first:
mkdir -p sources/cache
curl -L --fail --silent --show-error \
'https://drive.usercontent.google.com/download?id=1ujhm4_WYpgRV_rd1rJXIfHyvx16COEKe&export=download&confirm=t' \
-o sources/cache/sothanaphan-2.64e17.pdf
python3 -B - <<'PY'
from hashlib import sha256
from pathlib import Path
path = Path("sources/cache/sothanaphan-2.64e17.pdf")
expected = "8162113a571dc2283fc77de1cdf36e7abf424eeec952aa27cf82a4f44b3a796f"
assert sha256(path.read_bytes()).hexdigest() == expected
print("HIGH-RANGE SOURCE AUTHENTICATED")
PY
python3 -B computations/test_all_n_manifest.py
python3 -B computations/test_all_n_resume.py
python3 -B computations/check_prefix_certificate.py \
certificates/prefix-10000.json \
--expected-sha256 \
693ce882fb3f3786caf8eb502dd0677f42a4ed687adaa71a613b27ad7ef49727
python3 -B computations/test_prefix_checker.py \
certificates/prefix-10000.json
python3 -B computations/exhaustive_small_prefix.py --limit 100The full package requires two external repositories at exact immutable
revisions, plus Java, a C++20 compiler, Python, and gmpy2==2.2.1 backed by
GMP 6.3.0. REPRODUCE.md contains the exact clone, hash,
build, replay, operational-resume, mutation, TeX, and Lean commands.
The primary entrypoints are:
# Theorem-grade six-stage computational replay. Use a fresh empty directory
# outside this repository and do not pass --resume.
python3 -B scripts/check_certificate.py --work-dir /absolute/external/path
# Bounded Lean-runner controls, then a static source audit. Neither compiles Lean.
python3 -B lean/test_completion_gate.py
python3 -B lean/run_completion_gate.py --source-audit-only--resume is available only for operational recovery. Unsigned local
checkpoints cannot prove execution provenance, so a resumed run never emits the
theorem-grade completion PASS and cannot discharge CD0.
The full Lean completion gate intentionally refuses underprovisioned or unsupported hosts. Its documented minimum is Windows x86-64, 64 GiB physical RAM, 200 GiB free storage, and a 32 GiB guarded Lean ceiling.
| Path | Purpose |
|---|---|
proof/ |
Human-readable all-$N$ proof, TeX, and PDF |
certificates/ |
Canonical certificates and authenticated replay receipts |
computations/ |
Independent exact checkers, generators, and negative controls |
scripts/ |
Primary all-$N$ and finite replay orchestrators |
lean/ |
Literal final theorem, source lock, axiom audit, and guarded completion gate |
diagnostics/ |
Bounded host/cache/source-build experiments and their non-promotion boundaries |
docs/ |
Statement/source audits, proof DAG, ledger, failed lemmas, computation semantics, and handoff |
sources/ |
Source manifest; downloaded third-party PDFs/pages are intentionally untracked |
This project deliberately separates exploration, mathematical proof,
computation, structural review, detailed review, and formal verification.
It does not accept numerical evidence as proof, sampled endpoints as interval
coverage, opaque solver output, theorem-strength assumptions, or a Lean file
merely because it compiles. Exact counterexamples to failed helper claims are
preserved in docs/failed-lemmas.md.
See AGENTS.md for the complete research protocol and
CONTRIBUTING.md for the standard expected of changes.
This repository is an audit, synthesis, and reproducibility package. It builds
on the original Erdős–Sárközy problem; the sufficiently-large-(N) work of
Mehtaab Sawhney; Nat Sothanaphan's explicit threshold; the pinned ART-005
all-$N$ certificate repository; the pinned ART-006 Lean development; and
Denis Hanson's prime-counting estimate. Exact URLs, revisions, retrieved-file
hashes, statements, and known defects are in
docs/source-audit.md.
The research, code, proof drafting, and adversarial review were carried out in
a multi-agent OpenAI Codex workflow under Danny Ward's direction. “Independent”
in this repository means that separately assigned implementations or review
lanes did not certify their own work; it does not mean external scholarly
peer review. This public proof package remains unrefereed, and the public
Erdős Problems tracker had not incorporated it as a resolution when rechecked
on 13 August 2026. See PROVENANCE.md for the authorship and
evidence boundary.
No third-party source repository or PDF is silently vendored here. The
ignored external/ and sources/cache/ trees are reconstructed from immutable
identities in REPRODUCE.md.
Citation metadata is provided in CITATION.cff. Original code,
Lean files, and machine-readable verification material are under Apache-2.0;
original proof and documentation prose are under CC BY 4.0. Third-party works
retain their own terms. See LICENSE.md and
THIRD_PARTY_NOTICES.md.