Research paper (English + 中文): how a 21-file, 21,641-token reference formalization of the classical de la Vallée Poussin zero-free region became a single 2,489-token file — twelve record refreshes, each traced to its bottleneck, its core derivation, and its measured token delta.
Yin-Kai Yu (余荫铠) · DeepSeek Harness × ChatGPT Desktop collaborative research track https://github.com/YinkaiYu/riemann-hypothesis-zero-free-region-lean4
Paper: English · 中文版 · Cite: CITATION.cff ·
paper/references.bib
The target is the classical de la Vallée Poussin zero-free region: there exists an absolute constant
The size measure is the normalized Lean 4 token count import lines cost
| record | attempt | artifact | tokens | official score | running_best_pct |
structural advance |
|---|---|---|---|---|---|---|
| — | 96263 | previous holder (Pi) | 3,286 | 64.5215 | 85.68 | — |
| 1 | 96272 | Lean/record-96272/ |
3,274 | 64.5332 | 85.73 | recursive zero peeling replaces finite zero products |
| 2 | 96276 | Lean/fixes/localcert-fix/ |
3,208 | 64.6400 | 86.03 | 3–4–1 certificate inlined as one HasSum.nonneg object |
| 3 | 96281 | Lean/record-96281/ |
3,065 | 64.8928 | 86.64 | low-height exclusion + sharp pole bound read off the Mellin remainder |
| 4 | 96282 | Lean/record-96282/ |
3,014 | 64.9800 | 86.87 | window lemma fused into the assembly |
| 5 | 96283 | Lean/smallest-scored-96283/ |
2,931 | 65.8839 | 87.24 | one normalized complex identity carried through the assembly |
| 6 | 96286 | Lean/smallest-verified/ |
2,843 | 65.2900 | 87.59 | same lineage, next interface collapse |
| 7 | 96293 | Lean/record-96293-v25/ |
2,578 | 65.8000 | 88.80 | smaller assembly, further fusion (GPT-6.2 Astra artifact) |
| 8 | 96295 | — (intermediate variant of the v26 lineage) | 2,543 | 65.8700 | 88.97 | (GPT-6.2 Astra artifact) |
| 9 | 96308 | Lean/record-96308-v26/ |
2,536 | 65.8834 | 88.99 | current record (Deepseek V4.1-Flash) |
| 10 | 96331 | continuation of the v26 lineage (not separately archived) | 2,524 | 65.9100 | 89.05 | Deepseek V4.1-Flash |
| 11 | 96336 | continuation of the v26 lineage (not separately archived) | 2,518 | 65.9200 | 89.07 | Deepseek V4.1-Flash |
| 12 | 96342 | continuation of the v26 lineage (not separately archived) | 2,489 | 65.9800 | 89.07 | Deepseek V4.1-Flash |
| 13 | 97732 | Lean/record-97732-v2410/ | 2,410 | pending | — | Deepseek V4.1-Flash (recovered local attempt) |
| 14 | 97734 | Lean/record-97734-vd1/ | 2,411 | pending | — | Deepseek V4.1-Flash (recovered local attempt, backup) |
| — | 96572 | held by another participant | 2,223 (site) | 66.4431 | 89.73 | Pi · Gemini 4 Neon |
Compression.
The mathematical core is the same throughout; each record step improves an interface or a distinction that the classical proof does not need:
(i) Mellin entire formula. For
which yields both the polynomial growth bound
(ii) Recursive zero peeling. Put
The proof is a strong induction on
and the growth constant of exactly half that of
(iii) Assembly. With
The 3–4–1 inequality (norm_LFunction_product_ge_one) then yields the negative margin
so every estimate may be crude and one linarith closes the proof.
External validation. The benchmark's model × harness leaderboard ranks Deepseek V4.1-Flash +
DeepSeek Harness #1 by composite index (64.50) and by average proof compression (** 84.2 %**)
(evidence/figures/leaderboard-model-harness.png, leaderboard-avg-compression.png). The first record
moment is archived as a screenshot (evidence/figures/curve-record-162-submission.png: submission #162,
−85.7 %, 3,089 site tokens, 2026-11-22 05:23), and the platform's paper-style derivation of the technique
— Borel–Carathéodory + analytic divided difference (dslope) recursive zero peeling — in
evidence/figures/site-derivation-borel-dslope.png; submission-progress overview in
evidence/figures/submission-progress.png.
Progress-curve frontier (latest snapshots, Chinese and English UI). The public Proof Compression Progress Curve now shows 179 verified runs and 50 breakthroughs, with the Core-Pillar-1 frontier annotated by the technique Borel–Carathéodory with analytic divided-difference zero peeling under the account Yin-Kai Yu (余荫铠) · Deepseek…. The two archived screenshots are the Chinese-UI and English-UI renderings of the same frontier:
Ablation. The natural Euler–Maclaurin route to the growth bound — proving
Published technique (benchmark site, submission 96283): core1_borel_dslope — Borel–Carathéodory +
Divided-Difference (DSlope) Zero Factorization — reproduced verbatim (EN + 中文) in both papers.
| what | where |
|---|---|
| Paper (English) | paper/main_en.md |
| 论文(中文) | paper/main_zh.md |
| Five record artifacts | Lean/record-96272/ ,Lean/fixes/localcert-fix/ ,Lean/record-96281/ ,Lean/record-96282/ ,Lean/smallest-scored-96283/ |
| Smallest verified (2,843 tokens, pending) · rejected artifact | Lean/smallest-verified/ ,Lean/failed-96274-v22.lean |
| Independent Case A/B architecture | research/caseAB/ |
| Token meter + compile/axiom-audit harness | tools/ |
| Official evidence snapshots & figures | evidence/ ·docs/EVIDENCE.md |
| Extended notes | docs/ARCHITECTURE.md |
| Citation metadata | CITATION.cff ·paper/references.bib |
Besides Core Pillar 1/5, this repository publishes a fully machine-checked consolidated proof of Core
Pillar 5/5 (AlpogeRHCore5.functorialResolutionCharZero): 20 modules, 1,682,205 normalized tokens
against a 1,786,589-token reference closure (5.8 % compression), with a full clean rebuild (20/20
modules, rc=0), a clean axiom audit ([propext, Classical.choice, Quot.sound]), zero sorry/axioms,
zero frozen-interface redefinitions, and a matching sha256 delivery manifest. It was submitted as attempt
97785. Sources: Lean/core5/; evidence: evidence/core5/;
write-up: docs/CORE5.md and the independent audit protocol in
docs/core5-audit-report.md.
python3 tools/token_count.py Lean/smallest-scored-96283/Solution.lean # -> 2931
LEAN_HOME=/path/to/lean-4.34.1-linux MATHLIB=/path/to/mathlib4 \
tools/compile_solution.sh Lean/smallest-scored-96283/Solution.lean demo
The results were produced under the account Yin-Kai Yu (余荫铠) by a collaborative research track in which two AI agent tracks worked together (DeepSeek Harness / Deepseek V4.1-Flash and ChatGPT Desktop / GPT-6.2 Astra). Rejected artifacts and pending scores are reported as such.
@misc{yu2026zfrcompress,
author = {Yu, Yin-Kai},
title = {Library-Native Proof Compression: Successive Record Formalizations of the Classical Zero-Free Region in {Lean}~4},
year = {2026},
howpublished = {GitHub repository, artifact release v1.0},
url = {https://github.com/YinkaiYu/riemann-hypothesis-zero-free-region-lean4},
note = {Five record refreshes; best official score 65.8839 (submission 96283), curve kept=true},
}
MIT — see LICENSE. Lean/Mathlib dependencies keep their own licenses (Mathlib: Apache-2.0).