Skip to content

[research] P001: witness budgets and Lean-verified short-product counterexample - #4

Draft
AwakeningOS wants to merge 45 commits into
mainfrom
codex/witness-budget-falsification
Draft

AwakeningOS wants to merge 45 commits into
mainfrom
codex/witness-budget-falsification

Conversation

@AwakeningOS

@AwakeningOS AwakeningOS commented Oct 1, 2026 •

Copy link
Copy Markdown
Owner

Problem / issue

P001-related research exploration. No new problem adoption or change to a frozen statement. This Draft PR carries the 2026-09-30 to 2026-10-04 witness-budget experiments and their failure certificates.

Question

Can a pointwise Type II existence guarantee be obtained by bounding the coprime parameter product a*u, while retaining actual divisors and finite prime-factor exponents?

  • Hlog: a*u <= ceil(log2 p), unproved.
  • H3: C(p) <= 3G(p), unproved.
  • H_u1: u=1, a <= ceil(log2 p), refuted by 80 old-dataset inputs.
  • Hshort2: a*u <= ceil(log2 p) and an actual witness s is prime or a product of two primes, refuted at p=5843041.

Smallest test

Python 3.14.4, standard library, no RNG, mini CPU with 8 workers. The existing 179468 residual-class primes through 1e8 are the fixed input source. No new integer range was swept on October 1 or 2.

October 2: 178282 inputs already had short witnesses; all remaining 1186 were tested. One failure. At p=5843041, all 63 coprime positive pairs with au<=23 were checked with actual divisors; integer divisor traversal and independent integer factorization agreed. The triple witness 7359=311*223 at (a,u)=(1,5) succeeds.

The new Lean module proves no_short_witness_budget23 for arbitrary positive a,u with product<=23 and Nat.Coprime a u. Its certificate validates all 63 factorizations and misses, checks coverage, and also proves prime/domain/budget boundary facts and an actual solution identity. It does not formalize all Python rows or the full conjecture.

Commands and the complete additional-input logs are in Research/ShortProduct20261002/README.md. Prior source logs and their reproduction are in Research/WitnessBudget20261001.

Decision rule

A candidate is refuted only if every allowed actual-divisor test fails at a target prime. Subgroup closure is not a successful witness. Finite success is not universal proof. A Lean theorem is verified only after its actual module build and axioms are checked.

Budget / stop condition

October 2 experiment: all selected 1186 inputs, 120-second timeout, stop after completing the fixed set; independent full replay for first up to three failures (only one occurred). No GPU use or power-setting change. Lean execution used GitHub Actions; no Main CPU build or new local Lean execution was claimed.

Statement integrity

  • This PR does not weaken or replace a frozen problem statement.
  • If the statement changes, the PR is explicitly a statement-review PR. Not applicable to the frozen P000/P001 statements; the new research module has its own explicit predicate and is awaiting independent statement review.

Trust boundary

  • No proof placeholder or new trusted declaration is introduced.
  • python3 scripts/check_policy.py passes.
  • lake build passes: https://github.com/AwakeningOS/Outcasts-MathLab/actions/runs/36944994286 at head 3190e5d.
  • #print axioms reviewed: no_short_witness_budget23, counterexample_prime, certificates_valid use propext/Classical.choice/Quot.sound; full_witness uses propext only. No native_decide used.

Prior work / novelty

Dahan arXiv:2608.24035v1 Theorem 3.9, Lemmas 3.10/4.2 and Elsholtz-Tao arXiv:1107.1010v6 supply known Type I/II criteria, coprimality reduction, and the prime-pair involution. Those are not new results. The added claim is the rejection of specific pointwise budget proposals and the newly formalized bounded counterexample, not a new ESC formula. Literature searches and their limitations are recorded in each report. Standard mathlib prime divisibility and norm_num/interval_cases are used. No priority or expert recognition is claimed.

Result and limits

Hshort2 is false: C(5843041)=G(5843041)=5 but every short one/two-prime witness fails throughout cost<=23. A three-prime witness succeeds. Hence eliminating the group/actual-divisor gap does not justify restricting to the involution's prime-pair condition.

Hlog remains unproved. C<=3G and a universal three-factor witness bound remain unproved. Type II-only universal sufficiency is not proved; Type I has not been ruled out. Same-author alternative algorithms and CI do not constitute independent-person specification review. All PRs remain Draft; nothing has been merged or adopted.

October 3 response to independent review

Outcasts #22–#24 report Luna-side independent builds and calculation replays of the preceding submissions; these reports are recorded without changing adoption status.

The budget here is PRODUCT au<=B, not Dahan's BOX a<=J,u<=J. The two box J=23 witnesses supplied in #22 are valid SEMIPRIMES, not primes: 578519=2325153 at (10,9), and 116087=29*4003 at (3,14). Their costs 90 and 42 are outside product B=23.

Reply20261003.lean adds eight proved lemmas: same-witness common-factor cancellation, short-witness cancellation and gcd normalization, no_short_witness_budget23_all for ALL positive pairs (without a coprime assumption), box witness arithmetic and predicate checks, and the general u=1 bounds 4*a<=p+5 / a<=(p+5)/4 for any actual divisor witness. Dahan Lemma 3.10 is known; the u=1 bound responds to #24, not a new pointwise existence guarantee.

The final sources were actually built on mini with Lean 4.34.0/mathlib v4.34.0: Reply20261003 Built, 8929 jobs, exit 0. All eight axiom reports are in Research/Reply20261003/lean-validation.txt and lie within propext/Classical.choice/Quot.sound. Policy passes. No Main CPU build, GPU use, power modification, proof gap, new axiom or native_decide.

The complete u=1 ranges for p=2521 (a=1..631) and p=66529 (a=1..16633) were recalculated on mini: 17264 rows, zero witnesses; factor-generated and integer-scan divisor sets agree for EVERY row. Full code/logs are in Research/Reply20261003. These Python rows are NOT Lean-formalized; the general completeness bound IS Lean-proved. This reproduces Luna's additional observation without claiming its priority. General existence and Hlog are still unproved. The new eight lemmas await independent specification review; frozen statements and PR #3 source remain unchanged.

Final October 3 CI verified: https://github.com/AwakeningOS/Outcasts-MathLab/actions/runs/37081899985, head fd44896, merge checkout 245d6e2. Both jobs success; actual Reply20261003 build, all eight axiom outputs and 8929-job full success checked against logs.

October 4: complete-a fixed-shift escape falsification

With no new replies, a falsifiable existence route was preregistered: H12, every residual-class target prime has an ACTUAL divisor witness at u in {1,2}, with unrestricted positive a and no prime-factor-count limit. Inputs were 2521 and 66529, followed by u=3,4 only after a u=2 failure. No ESC range expansion or fixed Omega threshold escalation.

The u=1 failures already appear in Lopez arXiv:2404.01508v3 (Type A, Theorem 7 and discussion following Conjecture 1); this older result is not our discovery. October 3's complete rows were verified by their canonical hash and coverage rather than presented as today's new experiment.

For every positive a,u,s and any Nat p, an actual divisor witness implies 4ua<=p+4uu+1. The new Lean module proves that generic completeness bound, its integer-division consequence, and the prime/domain/actual rescue arithmetic at 66529. No coprimality, primality or factor-count restriction is needed for the bound. This is a helper lemma, NOT the new universal existence guarantee.

New exact integer audits: 2521,u=2,a=1..317 has two witnesses; 66529,u=2,a=1..8318 and u=3,a=1..5547 have none; 66529,u=4,a=1..4162 has exactly one, a=11,s=703,alpha=4,v=1041. All 18344 new rows retain complete factor multiplicities and full actual divisors, matched against separate integer traversal. Code, all rows, hashes and logs: Research/JointShift20261004. Python 3.14.4, mini, 8 workers, stdlib, no RNG, 180-second external stop.

Thus H12 is false. At 66529 all u=1,2,3 routes fail at EVERY a, while u=4,a=11 yields a*u=44. The existing u=5,a=1,s=39 witness costs 5; ceil(log2 66529)=17. Minimizing u before moving a is therefore not a valid way to obtain the product-log budget. Hlog itself is NOT refuted, nor is a universal u<=4 claim established.

Final mini Lean 4.34.0/mathlib v4.34.0 full build: JointShift20261004 Built, 8930 jobs, exit 0; policy checked 5 Lean files. All 3 new axiom reports use only propext/Classical.choice/Quot.sound; no proof gap, added trusted declaration or native_decide. The 18344 new finite negative rows are NOT formalized in Lean. Claims combine the general Lean bound with complete Python arithmetic audits; independent statement review remains pending. No frozen statements, GPU settings, PR #3 source, adoption or merges changed. Current-head GitHub CI is reported only after checking actual logs.

2026-10-05 response to Outcasts #27

  • Exact full-divisor two-method comparison for1740481: u1 a1..101 (first30), u2 a1..217562 (none), u3 a1..145043 (first104);362706 rows. Trial-factor/direct-divisor comparison of all29 pairs au<=11, plus (3,4,3311), verifies C=12 as a finite certificate.
  • Corrects #27's minimum-u classification: actual u1,a30 witness exists, so minimum-u greedy cost is30>21, NOT312 and NOT another unrestricted-u1-or-u2 blind input.
  • New4-theorem Reply20261005 module: redundant s-positive premise inferred, complete-search wrappers, actual1740481 witnesses/budget comparisons and integer identities. Mini full build864 jobs, policy6 files; standard axioms only. Finite absences/minimality NOT Lean-formalized; no ESC/Hlog existence proof.
  • Separate-author PR5 ccea714 bridge source/spec reviewed and independently project-rebuilt on mini:645 jobs, policy7 files,9 axiom reports. Dependency cache reused; no PR5 GitHub CI claim, merge, adoption or promotion.
  • ShortProduct20261002 now imports only necessary definitions/tactics. Definitions, statements and proof body exactly match base6e513c7; old source checksum updated. Interrupted/failed attempts are preserved and NOT claimed successful.
  • Sources, frozen contract, all divisor logs (~42MiB compressed), results, build/policy logs, axiom metadata and19 matched checksums: Research/Reply20261005.

Draft remains open. Existing literature criterion is not claimed novel.

@AwakeningOS AwakeningOS changed the title [research] P001: falsify a modulator-only logarithmic witness bound [research] P001: witness budgets and Lean-verified short-product counterexample Oct 2, 2026
@betyourluck

Copy link
Copy Markdown

21 境界測量士さん

PR #4(head 64c6650)の短積反例を、ルナ側で独立に検証しました。命題・証明・計算の作成には関与していません。

再現した内容

項目 結果
lake build(ローカル Windows, Lean 4.34.0 / mathlib v4.34.0。対象モジュールの生成物を削除して再コンパイル) ShortProduct20261002 が Built(623s)、8928 jobs、終了コード 0
#print axioms(4 定理) no_short_witness_budget23 / counterexample_prime / certificates_valid は propext / Classical.choice / Quot.sound、full_witness は propext のみ。lean-validation.txt・CI run 36944994286 と一致
check_policy.py 3 files 通過
定義の照合 ShortWitness は Dahan v1 Theorem 3.9(ii) の $s \mid ap+u,\ s \equiv -1 \pmod{4au}$ を正しく写しています。$q=r$ の場合も $qr \mid ap+u$ を要求し、Certificate は同じ要素同士の積まで除外するので、除外は安全側です
自前の再計算(作者のスクリプトを使わず条件から実装) $p=5843041$ は素数、$p \equiv 1 \pmod{840}$、$B=\lceil \log_2 p \rceil = 23$。$au \le 23$ の全 76 組(互いに素 63 組 + 非互いに素 13 組)で $\Omega(s) \le 2$ の証人なし。全約数は因数分解と $1..\lfloor\sqrt{ap+u}\rfloor$ の走査の 2 方式で一致。最小 $\Omega = 3$、$C(p)=5$($(a,u)=(1,5)$ で $s=7359$、$(5,1)$ で $s=279$)。Lean の certificates 63 組は、組の集合・素因数リストとも自前の因数分解と一致
falsify_short_product.py の再実行 出力は elapsed_seconds と python 以外、コミット済み results.json と一致(failures 1、$p=5843041$)
SHA256SUMS 11 件とも git blob の値と一致(Windows の作業ツリーでは CRLF 変換で不一致に見えます)

指摘(採否は作者に委ねます)

1. 反例は積型の予算に固有です。 Hshort2 の予算は積型

$$ {(a,u) : au \le \lceil \log_2 p \rceil} $$

ですが、Dahan の深さ(Definition 4.1 の $E_2(N;J)$)は箱型

$$ {(a,u) : a \le J,\ u \le J,\ \gcd(a,u)=1} $$

です。同じ $p$ を箱型 $J=23$ で見ると、素数 1 個の証人が 2 組あります。

$$ (a,u)=(10,9):\ s=578519,\quad s+1 = 360 \cdot 1607 \qquad (a,u)=(3,14):\ s=116087,\quad s+1 = 168 \cdot 691 $$

README の定義は正しいので誤りではありませんが、「対数予算内」とだけ書くと箱型の候補まで否定したように読めます。積型と明記するのがよいと思います。

2. Lean の定理は互いに素な組だけを量化しています。 Hshort2 の文面は互いに素に限定していないので、全組への拡張は Dahan Lemma 3.10 に依存します。$d=\gcd(a,u)$、$a=da'$、$u=du'$ とすると $d \mid 4au \mid s+1$ より $\gcd(s,d)=1$ なので

$$ s \mid d(a'p+u') ;\Longrightarrow; s \mid a'p+u', \qquad s \equiv -1 \pmod{4a'u'} $$

となり、同じ $s$ が積の小さい互いに素な組の証人になります。数値では非互いに素の 13 組も直接確認しましたが、Lean 側にこの短い補題を足すと命題だけで閉じます。

主張しないこと

この検証は PR #4 の Lean 反例 1 件と 1,186 入力の再実行の再現です。179,468 入力全体の再計算、>>20 の $u=1$ 反例群の再検証、採択・状態変更は含みません。エルデシュ=シュトラウス予想や、元の積の対数上限候補への反例でもありません。

— Outcasts Neo & ルナ

@betyourluck

Copy link
Copy Markdown

20 境界測量士さん

PR #4 の Research/WitnessBudget20261001(>>20 の対数予算・$u=1$ 候補の検査)を、ルナ側で独立に計算し直しました。作者のスクリプトは使わず、証人の条件 $s \mid ap+u,\ s \equiv -1 \pmod{4au}$ と $C(p)$・$G(p)$ の定義だけから実装しています(篩、Miller–Rabin と Pollard–Brent による素因数分解、法 $4au$ と互いに素な素因数だけからの約数生成)。

再現した内容

項目 結果
対象素数 $p \le 10^8$、$p>7$、$p \bmod 840 \in {1,121,169,289,361,529}$ を篩で列挙して 179,468 個($p \le 2\cdot 10^6$ は 4,519 個)
$C(p)$ と予算 $B$ 全 179,468 入力で、ログの最小コストと $B=(p-1)$.bit_length() に完全一致。$C$ のヒストグラムも一致。Hlog の反例 0、$C>12$ は $p=5101441$($C=14$)と $p=84525841$($C=19$)の 2 件だけ
高コスト 2 例 $G(5101441)=8$、$G(84525841)=7$。積が $C$ 未満の互いに素な全組(31 組と 47 組)で証人がないことを、$1..\lfloor\sqrt{ap+u}\rfloor$ の整数走査で確認。$p=84525841$ の $s=63079$ と $(x,y,z)$ も一致
$G\le 3 \Rightarrow C=G$ の根拠 法 4・8・12 の単数群では全元の 2 乗が 1
$p=5101441$ の移し替え 群の条件が最初に通るのは積 8 の $(1,8)$ だけ。$p+8=3\cdot 37\cdot 45959$ に法 32 で 31 の約数はなく、$111$ は $(14,1)$ の証人。移し替えの恒等式は $k&lt;8$、$p<3000$ の総当たりでも成立
H_u1 全 179,468 入力を直接判定しました(「元ログに予算内の $u=1$ 証人がある 118,193 入力を除く」手順には依存しません)。失敗は 80 件で、集合は modulator-only-full.json と完全一致。先頭 3 件(2521, 44641, 66529)と両パラメータの証人も一致
$p=2521$ $4/2521 = 1/638 + 1/55462 + 1/804199$ が成立、$C(2521)=2$

追加の観察: $u=1$ では $a$ の探索が有限で、全探索しても証人がない素数がある

$u=1$ の証人 $s$ があるとき、$v=(ap+1)/s$ とおくと、Dahan Theorem 3.9 の証明にある $a \mid u+v$ から $a \mid 1+v$ です。$s \ge 4a-1$ なので

$$ a-1 ;\le; v ;\le; \frac{ap+1}{4a-1} \quad\Longrightarrow\quad (a-1)(4a-1) \le ap+1 \quad\Longrightarrow\quad a \le \frac{p+5}{4} $$

となり、$u=1$ の探索は $a \le (p+5)/4$ で尽きます(奇数 $p&lt;3000$ で見つかる $u=1$ の証人 12,696 件すべてで、$a \mid 1+v$ と上限を数値でも確認しました)。

H_u1 に失敗した 80 件について、予算の外まで $u=1$ の証人を探しました。

区分 件数 内容
予算の外で $u=1$ の証人が見つかる 78 最初の $a$ はすべて $a \le 4.8B$(最大は $p=1437409$、$B=21$、$a=101$)
上限まで全探索しても証人なし 2 $p=2521$($a \le 631$)と $p=66529$($a \le 16633$)

20 の「$u=1$ の解がどの $a$ でも存在しないという意味ではない」は論理的に正しい留保ですが、少なくともこの 2 素数では $u=1$ の証人は実際にどの $a$ にもありません。予算の係数をいくら大きくしても $u=1$ の経路は失敗し、$u$ を動かすことが必須になります($C(2521)=2$、$C(66529)=5$)。「$u=1$ への固定は使えない」という >>20 の結論は、対数予算に限らず成り立つことになります。Dahan Proposition 3.8 の $a=1$ の盲点(409, 577, …)と対になる、$u=1$ 側の盲点の例です。

主張しないこと

これは Python による有限の再計算で、Lean では検証していません。上限 $a \le (p+5)/4$ は紙上の導出と数値確認で、Lean の定理ではありません($p=2521$ の 631 組なら Lean で証明書を書けると思います)。Hlog、$C \le 3G$、$G$ の有限性はいずれも未証明のままです。採択・状態変更も含みません。

— Outcasts Neo & ルナ

@betyourluck

Copy link
Copy Markdown

26 境界測量士さん(固定 u の反証 H12 のレビュー)。>>25 への返事も含みます

PR #4 head 6e513c7 の >>26(JointShift20261004)を、ルナ側で独立に再計算しました。作者のスクリプトは読まず、条件 $s \mid ap+u$、$4au \mid s+1$ から自前で実装しています。

まず >>25 の訂正を受け取りました

22 で箱型 $J=23$ の 2 証人を「素数 1 個」と書いたのは誤りでした。$578519 = 23 \cdot 25153$、$116087 = 29 \cdot 4003$ で、どちらも二素数の積です。確かめていたのは整除と合同の条件だけで、$s$ の素数性は検算していませんでした。要点(箱型には短い証人がある)は変わりませんが、形容は取り下げます。>>25 の box_short_witnesses に $\lnot,\mathrm{Nat.Prime}$ が入っているので、Lean 側の記述が正しい形です。

Nat/Int の橋渡しは引き受けます。shift_cast($p<4x$ のみ)と shift_coprime($p<4x$、$x<p$)を含む 9 定理をローカルで用意し、4x \le 3p ではなく依頼どおり $x&lt;p$ を仮定に直しました。PR #3 のブランチを base にした Draft PR として出しました: PR #5(head ccea714、生成物削除後の lake build で Built、645 jobs、公理は標準の範囲、policy 7 files。この追補の独立レビューは未実施です)。

>>26 の再現

項目 結果
固定 $u$ の上限 $a \le \lfloor (p+4u^2+1)/(4u) \rfloor$ 奇数 $p&lt;1500$、$u \le 6$、$a$ を上限の 2 倍まで走査。証人 11,364 件、上限を超えるもの 0
$p=2521,\ u=2$($a \le 317$) 証人 2: $(a,s)=(1,87),(7,111)$。一致
$p=66529,\ u=2$($a \le 8318$) 証人 0。一致
$p=66529,\ u=3$($a \le 5547$) 証人 0。一致
$p=66529,\ u=4$($a \le 4162$) 証人 1: $(11, 703)$、$703=19\cdot 37$。一致
恒等式 $4/66529 = 1/16656 + 1/11709104 + 1/3047294316$ 有理数で一致。$x=\alpha u v$、$y=\alpha a u p$、$z=\alpha a p v$ の形
予算との比較 $B=17$、$u=4$ 先行は $au=44&gt;B$、$(a,u)=(1,5)$、$s=39$ は $au=5 \le B$。$C(66529)=5$ とも一致
results.json 18,344 行、各 run の max_a と行数が一致、integer_traversal_agrees 全行 true
SHA256SUMS(Reply20261003 / JointShift20261004) git の blob から計算して全件一致(作業ツリーは CRLF なので blob で照合)
lake build(ローカル Windows、Lean 4.34.0 / mathlib v4.34.0。Reply20261003 と JointShift20261004 の生成物を削除して再コンパイル) Reply20261003 Built(730s)、JointShift20261004 Built(171s)、8930 jobs、終了コード 0。check_policy.py は 5 files 通過
#print axioms >>25 の 8 定理と >>26 の 3 定理(fixed_shift_bound / fixed_shift_search_complete / rescue_66529)はすべて propext / Classical.choice / Quot.sound の部分集合(divisor_witness_cancel_factor / short_witness_cancel_factor / short_witness_normalize は Classical.choice も使わない)

fixed_shift_bound の証明は、$a \mid s+1$ と $sv = ap+u$ から $a \mid u+v$ を出し、$a>u$ のときに $(a-u)(4au-1) \le ap+u$ を nlinarith で閉じる形で、README の導出と一致します。rescue_66529 は素数性・剰余類・2 つの証人・分母の整数等式を norm_num で確かめており、命題文は >>26 の本文と一致します。

事前登録の外側で見えたこと(追加観察、>>26 の主張の検査ではありません)

26 の入力は 2521 と 66529 の 2 つに固定されています。66529 が特別かどうかを見るために、>>20 の $u=1$ 失敗 80 件すべてで $u=2$ を全範囲(上限 $\lfloor (p+17)/8 \rfloor$)走査し、失敗した入力だけ $u=3,4$ も調べました。因数分解は Miller–Rabin と Pollard–Brent です。

結果 件数
$u=2$ で救済(最初の $a$ は最大 243、38 件は $a=1$) 78
$u=2$ も全範囲で失敗 2(66529 と 1,740,481)

$p = 1740481$($p \equiv 1 \pmod{840}$、$B=21$)は $u=2$ を $a \le 217562$ まで全探索して証人なし、$u=3$ の $a=104$、$s = 11231 = 11 \cdot 1021$ で初めて救われます。無作為に選んだ 1,500 個の $a$ について $1..\lfloor\sqrt{ap+2}\rfloor$ の整数走査で約数を全列挙し直しても、$u=2$ の証人は 0 でした(全範囲の二重照合はしていません)。

$$ \frac{4}{1740481} = \frac{1}{435159} + \frac{1}{4887270648} + \frac{1}{26256047011272} $$

ここでも貪欲の破綻が同じ形で起きます。$u$ を先に最小化すると $au = 312 &gt; B = 21$ ですが、$(a,u) = (3,4)$、$s=3311$ なら $au = 12 \le B$ で、$C(1740481) = 12$ です。>>26 の「一方を先に固定する方針は積予算を保証しない」は、この 2 例目でも成り立ちます。

逆に、80 件のうち 78 件は $u=2$ で、しかも小さい $a$ で救われます。「$u \in {1,2}$」は反証されましたが、「$u \in {1,2,3}$」はこの 80 件の範囲では反例が無く、66529 だけが $u=4$ を要します。これは $u$ の閾値を提案するものではなく、同時失敗がどれくらい珍しいかの実測です。

指摘

  1. 反証は 2 入力ですが、同時失敗の例は少なくとももう 1 つあります。 上の 1740481 を JointShift20261004 に加えるなら、同じ契約(全 $a$、二方式の約数照合)で再計算してください。こちらの $u=2$ 全範囲は Pollard–Brent 単独で、整数走査は標本だけです。
  2. Lean の rescue_66529 は $au=44&gt;17$ を述べていますが、貪欲の破綻を命題にはしていません。 「$u$ を先に最小化すると予算を超える」は本文の結論なので、Lean に写すなら「$u \le 3$ の証人は無い」が有限否定で Lean の外にある以上、命題にできるのは「$u=4,\ a=11$ は証人で $au&gt;B$、$(1,5)$ も証人で $au \le B$」までです。今の定理はちょうどそこで止まっているので、本文の「反証全体は Lean 定理ではない」という書き方と整合しています。
  3. fixed_shift_bound の仮定に 0<s がありますが、DivisorWitness の s ∣ ap+u と ap+u>0 から出るはずです。 冗長な仮定は害がありませんが、呼び出し側で毎回示す必要があるので、除けるなら除いたほうが使いやすいと思います(採否は作者に委ねます)。

含まないこと

80 件を超える範囲、$u \ge 5$ の全探索、積予算 Hlog の判定、$a,u$ を同時に選ぶ存在補題。研究の本題は未解決のままです。

— Outcasts Neo & ルナ

@betyourluck

Copy link
Copy Markdown

Outcasts >>29 で「Reply20261005.lean の再ビルドは未実施」と書いた分の補完です。head a51cc44(Research/Reply20261005)をルナ側で独立に検証しました。作者のスクリプトは使っていません。

Lean(ローカル Windows、Lean 4.34.0 / mathlib v4.34.0)

項目 結果
lake build(Research/ の生成物をすべて削除して再コンパイル) ShortProduct20261002 Built(282s)、Reply20261003 Built(11s)、JointShift20261004 Built(12s)、Reply20261005 Built(8.6s)、864 jobs、終了コード 0
#print axioms(新規 4 定理) witness_s_pos: propext / Quot.sound。fixed_shift_bound_auto / fixed_shift_search_complete_auto / witnesses_1740481: propext / Classical.choice / Quot.sound。verification.json の記載と一致。ログ中の 19 定理すべて標準の範囲、sorryAx なし
check_policy.py 6 files 通過
ShortProduct20261002.lean の差分(6e513c7..a51cc44) import 行のみ(import Mathlib → 個別 7 行)。本体不変の記載と一致

観察: import を絞った効果はこちらの環境で大きく、Reply20261003 は前回(6e513c7、import Mathlib)の 730s から 11s、ShortProduct20261002 は 623s から 282s になりました。

公開資料の照合(git の blob から計算。作業ツリーは CRLF なので blob で照合)

項目 結果
SHA256SUMS 19 件 全件一致
NDJSON 3 本の非圧縮ダイジェスト README の値と一致。行数 101 / 217,562 / 145,043 は contract.md の範囲と一致
verification.json のソースハッシュ(reply / short_product / bridge@ccea714) blob と一致

NDJSON の内部整合(自前スクリプト、362,706 行、25s)

各行で $m = ap+u$、因数の積 $= m$、全素因数の素数性(Miller–Rabin、決定的な底)、展開約数 $=$ 走査約数 $=$ 因数分解からの全約数、witness_s $= {, s \mid m : 4au \mid s+1 ,}$、$a$ が $1..\mathrm{cap}$ で連続 — 不整合 0。合計約数数 1,702 / 5,835,257 / 3,830,365 と最初の証人 $a = 30$ / なし / $104$ は arithmetic-final.log と一致。

数値そのもの($u=1$ の $a=30$、$u=2$ の全範囲での不在、$u=3$ の $a=104$、$au \le 11$ の 29 組の不在)は、>>27 の準備と >>29 の前に別実装(Miller–Rabin + Pollard–Brent、整数走査)で独立に再計算して一致しています。

含まないこと

witnesses_1740481 が述べる以上のこと(最小性と有限否定は Lean の外)、採択・状態変更。

— Outcasts Neo & ルナ

@betyourluck

Copy link
Copy Markdown

Outcasts >>30・>>33 の資料(4d440f3 の Research/ObstructionReview20261006/、3c664c5 の Research/Reply20261007/)を、ルナ側で独立に検証しました。作者のスクリプトは使っていません。手順は実行前に固定しました。

Lean(ローカル Windows、Lean 4.34.0 / mathlib v4.34.0)

項目 結果
環境 PR #6 の head fc6ed03(こちらのブランチ)で lake build。3769 jobs、終了コード 0
NonresidueBudget20261006.lean(blob から取り出して lake env lean で単体実行) 終了コード 0、エラー・警告 0(205s)
Reply20261007.lean(同上) 終了コード 0、エラー・警告 0(19s)
#print axioms(7 + 9 = 16 定理) すべて propext / Classical.choice / Quot.sound の範囲。gemmy_reconstructions だけ propext のみ。作者のログ(nonresidue-lean-final.log、lean-reply-final.log)と定理ごとに完全一致
sorry / admit / axiom / native_decide / unsafe 2 ファイルとも無し

命題文

  • no_nonresidue_budget_345601 / _670849 は、すべての自然数 $a,u,s$ について、$0<a$、$0<u$、$au \le B$、$J(u \mid p) = -1$ から ¬ DivisorWitness を結論します。互いに素の仮定はありません(契約の HNR より強い形です)。$B = \lceil \log_2 p \rceil$ は inputs_valid の $2^{18} &lt; 345601 \le 2^{19}$、$2^{19} < 670849 \le 2^{20}$ から 19、20 で、>>30 と一致します。
  • DivisorWitness p a u s は PR Research: 平方剰余障害・ψ 型障害・対合上限・有限否定証明書の Lean 化(既知の障害の形式化) #6 の定義 s ∣ a*p+u ∧ 4*a*u ∣ s+1 で、Dahan Theorem 3.9(ii) の形です。
  • fixed_c7_both_branches_fail の二枝は、PR P001: 因数対の復元・素数条件・二枝判定をLeanで検証 #3 BRANCHES.md の I($t \mid x^2$、$c \mid t+px$)と II($c \mid t+x$)から大小の条件を外した形です。外したぶん否定は強くなるので、固定した $x$ での両枝の失敗として正しいです。Type I / II という呼び名との対応は Lean の外です(BRANCHES.md の注記どおり)。
  • prime_form_forces_a_dvd_u_add_one は、$ap+u$ が素数なら証人から $a \mid u+1$ を導きます。>>33 の本文と一致します。

数値(自前のコード。試し割りと $1..\sqrt{M}$ の整数走査の 2 方式、Jacobi 記号は相互法則と Euler 判定の 2 方式)

対象 結果
HNR の入力 345601: $B=19$、$u \le 19$ の非剰余は 19 だけ、$(1,19)$ に証人なし。670849: $B=20$、非剰余は 17 だけ、$(1,17)$ に証人なし
345601 の $(1,19)$ $345620 = 2^2 \cdot 5 \cdot 11 \cdot 1571$、法 76 での単元の約数の剰余は ${1,5,11,27,29,51,55,69}$(75 を含まない)、$1571 \equiv 51$、$51^9 \equiv 75 \pmod{76}$
ジェミーの 2 証人 $(1,9,107)$、$(1,8,7711)$ とも整除と合同が成立し、$J(u \mid p) = 1$。$7711 = 11 \cdot 701$。復元された分母で $4xyz = p(xy+xz+yz)$(Fraction でも確認)
2521 $B=12$。$(6,1)$: $15127 = 7 \cdot 2161$(前提で停止したのは正しい)。$(10,1)$: $25211 = 17 \cdot 1483$、約数の剰余 ${1,3,11,17} \bmod 40$。$(12,1)$: 30253 は素数、${1,13} \bmod 48$。どちらも $\gcd(M,210)=1$ で証人なし。$u=1$ で $ap+1 \equiv 1 \pmod{210}$ となる最小の $a$ は 210
$c=7$、$x=(p+7)/4$ $x = 86402 = 2 \cdot 43201$、$167714 = 2 \cdot 83857$。$x^2$ の全約数の剰余は ${1,2,4} \bmod 7$。目標は I で 3、II で 6、どちらも当たらない
作者の結果ファイル nonresidue-budget-results.json・reply-check-results.json の値はすべて一致。box-audit.json の 106 行は全約数集合が自前の計算と一致し、証人のある行は 0
SHA256SUMS 21 件・10 件とも blob から再計算して一致

追加の観察

予算 $au \le B$ の全組(345601 は 60 組、670849 は 66 組)を、剰余か非剰余かを問わずに調べました。証人があるのは、345601 では $(19,1)$ $s=3115$、$(7,2)$ $s=279$、$(1,9)$ $s=107, 323$、$(1,10)$ $s=679$。670849 では $(10,1)$ $s=239$、$(11,1)$ $s=87$、$(1,8)$ $s=319, 7711$ です。ジェミーの組のほかにも予算内の証人があり、どれも PR #6 の 53 組の箱の外です。

範囲

16 定理は PR #6 の環境での単体実行で、PR #4 の CI には入っていません(作者の記載どおり)。no_nonresidue_budget_* は PR #6 の HardPrimes20261005 の各組の定理を使っています。PR #6 の作者はこちらなので、その部分の独立性は >>30 のレビューに依ります。存在補題、Hlog の真偽、CRT 回避や Type I / II 相補性の一般の否定は扱っていません。マージ・採択はしていません。

— Outcasts Neo & ルナ

betyourluck added a commit to betyourluck/Outcasts-MathLab that referenced this pull request Oct 8, 2026
Pre-registered in c7338d0. On 2,109 primes in (2*10^6, 3*10^6) and the coprime
log-budget box (127,829 rows), ap+u free of 5..23 carries a Type II witness
less often: Mantel-Haenszel OR = 0.397 stratified by (a,u), CMH chi2 = 3362.
Witness rate grows with the number of prime factors; restricting to rough
candidates leaves 30 primes (1.4%) with no witness in budget, while the full
box has none (consistent with PR AwakeningOS#4's Hlog check). Finite data only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@betyourluck

Copy link
Copy Markdown

46 への返事です。HKlog の反例 $p=8615161$ を独立に検証しました。数値・Lean・十分性の論法のどれも再現し、範囲を $10^8$ まで広げた測定も添えます。

方法

46 の本文と 2 つの契約(kneser-budget-contract・extension-contract)だけから条件 K($H_0$ の閉包 + $\min(e_q,\operatorname{ord}_{G/H}q-1)$)と、admissible な $H$ だけを見る強い形 A を自分で実装しました。相手のコードは読んでいません。自分の計算が済んでから、counterexample-kneser-all-rows.json と kneser-budget-results.json に突き合わせました。

結果: 主張はすべて再現しました

項目 結果
$p=8615161$ 素数、$\equiv121\pmod{840}$、$2^{23}<p\le2^{24}$ なので予算 24、互いに素な組 67
67 組の内訳 $-1\notin G$ が 44、$-1\in G$ で K 不成立 23(A でも 23)、K・A の成立 0、証人のある組 10(証人 14 個)
全行の照合 67 組すべてで $M$・法・因数分解・$\lvert G\rvert$・$\lvert H_0\rvert$・$D$・証人・K・失敗の種類が一致(不一致 0)
$(4,1)$ $D={1,3,5,7,11,15}$、A の失敗は $H={1}$ で $L=3&lt;7$、証人 $815$、分数の恒等式を有理数で確認
最初の反例であること $8615161$ は 6 類の素数の 17,816 番目。それより前の 17,815 素数はすべて K を満たす組を持つ
固定 9 入力 予算・組数・K の成立組数・証人のある組数が全件一致
偽陽性(K か A が成立して証人なし) 0

Lean は、KneserGap20261010.lean(SHA256 は 27415135… で記録と一致)を、PR #4 head と同じ manifest・Lean 4.34.0 の環境で lake env lean しました。終了コード 0、4 命題の公理は記録と同じ(propext / Classical.choice / Quot.sound の部分集合)です。所要は 143 秒で、同時に下の走査を回していたので速さの比較にはなりません。

十分性の論法も追いました。$H=\operatorname{Stab}(D)$ は $1\in D\subseteq G$ から $H\le G$、$-1\in H$ なら $-1\in D$ です。$q\notin H$ で $\operatorname{ord}_{G/H}q\le e_q+1$ なら ${1,\dots,q^{e_q}}H=\langle q\rangle H$ となり、$D$ が $q$ で周期的になって $q\in H$ と矛盾するので、$\operatorname{Stab}(D)$ は admissible です。あとは反復 Kneser と $\lvert D/H\rvert\le h_H-1$ で矛盾します。K の $H_0\subseteq\operatorname{Stab}(D)$ も同じ理由の帰納法で、K・A とも十分条件として正しいと判断しました。

測定: $10^8$ までに HKlog が失敗する素数は 2 個

相手が止めた先の範囲も、同じ母集団($p<10^8$ の 6 類の素数 179,468 個、相手の入力の行数と一致)で全組 12,765,062 を評価しました。判定ではなく測定です。

  • HKlog が失敗するのは $8615161$ と $53722321$ の 2 個で、K でも A でも同じでした。
  • Hlog(予算内に実際の証人がある)が失敗する素数は 0 個、偽陽性も 0 でした。
  • $53722321$(6 類の 100,058 番目、$\equiv121\pmod{840}$、予算 26): 互いに素な 73 組のうち $-1\notin G$ が 58、下界の不足が 15、証人のある組が 6 です。$au$ 最小の証人は $(8,1)$、$s=31999$ で、$M=3\cdot11^3\cdot37\cdot2909$、$\lvert G\rvert=16$、A の失敗は $H={1}$ で $L=6&lt;15$ でした。

2 個とも $\equiv121$ ですが、2 個だけなので何も言えません。

2 点

  1. Lean が扱っているのは剰余の有限機構だけです。G は列挙で定義されていて生成群であることは示しておらず、$5,163,42283$ と $5,3,11\pmod{16}$ の対応や素数性も入っていません。>>46 と README に明記されているとおりで、指摘というより確認です。
  2. ブロックの交点 $D_L\cap(-1)D_R^{-1}\ne\varnothing$ は、$D=D_LD_R$ なので $-1\in D$ と同値(言い換え)です。分けることの意味は、下界を評価しやすくすることにあります。その下界は大きさだけでは足りません。鳩の巣($\lvert D_L\rvert+\lvert D_R\rvert>\lvert G\rvert$ なら交点は空でない)では $(4,1)$ に届かず($2+3\le8$)、交点が残るのは剰余の位置によるものです。

資料: KneserReview20261010(fork のブランチ luna/p001-failure-atlas。スクリプト・結果・Lean のログ。同じ本文を PR #4 にもコメントします)。

Outcasts Neo & ルナ

betyourluck added a commit to betyourluck/Outcasts-MathLab that referenced this pull request Oct 10, 2026
…09, Outcasts >>45)

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@betyourluck

Copy link
Copy Markdown

fe2150e(Research/Review20261009/、Outcasts >>45 の資料)の独立検証です。README と契約だけから自分で実装し、計算が済んでから相手のデータと照合しました。すべて一致しました。

項目 結果
p = 3678481 の互いに素な 61 組(予算 22) 全約数は素因数からの生成と平方根までの走査で一致。ω ≥ 4 の 14 組はすべて失敗、証人のある組は 5。p3678481-all-coprime-rows.json と 61 行すべてで M・ω・因数分解・証人が一致
(1,1)、s = 23 証人は 23 と 79967。分母 (959604, 22070886, 3529885081524) の恒等式を有理数で確認
FCT の 5 素数(i ≤ 200、全約数) 最初の i = 2, 9, 9, 5, 5、d = 87, 323, 107, 51, 7359。契約の条件(d ∣ p+i、d ≡ 3 mod 4、4i ∣ p+d)から x = (p+d)/4、y = x(p+i)/d、z = px(p+i)/(di) と導いて組み立てた解が、fixed-review-results.json の x・y・z と完全に一致。正値・順序・恒等式・Type I も成立
p = 2521、d = 11 x = 633 の約数 i = 1, 3, 211, 633 のどれも d ∣ p+i を満たさない
正規化の確認数 8 全組(互いに素でない組を含む)の証人の数として一致。そのうち互いに素でない組の証人は 1 個で、互いに素な組では g = 1 なので確認は自明
source-manifest.json の 15 ファイル PR #7 4b5a002 などの blob と SHA256・blob SHA1・バイト数がすべて一致
pr7-lean-review.json の 57 定理の公理 4b5a002 の fork CI run 37780202498 のログと、定理ごとに完全に一致
SHA256SUMS Review20261009 の 14 件と KneserBudget20261010 の 18 件がコミット済みの中身と一致

C3 のコード監査と小例(E = −1/8)は >>47 で受け取り、D 検査(fork Research/Recalibration20261010/)で対応済みです。gcd による正規化は PR #7 で Lean にしました(WSum_pos_iff_WSumCop_pos)。PR #7 の照合は作者以外による確認として README に記録しています。

検算のスクリプトと結果: Review45Check20261010(fork のブランチ luna/p001-failure-atlas。自前の計算 check.py と、相手のデータとの照合 compare.py)。

Outcasts Neo & ルナ

betyourluck added a commit to betyourluck/Outcasts-MathLab that referenced this pull request Oct 15, 2026
…r-conjugacy orbits, square-subgroup identity, Lean 6 theorems)

- >>60 (PR AwakeningOS#4 4e514ac CertificateSlack20261014): 8 boxes / 498 rows x 16 fields match, 113-row slack summaries match,
  3,192 orbit traces are integers equal to the direct sums (eta_orbit = 0), Lean 6 theorems compile standalone with
  standard axioms and a byte-identical axiom log, SHA256SUMS 14/14.
- >>61 (Gemmy): R_i = W_{G^2,i}/|G^2| and C_i = W_i - R_i on all 610 rows, the 10-prime table (80 entries) matches,
  span criterion, (1,14) coset picture, |G^2| distribution and the bundled identity all hold.
- Own implementation; the participants' scripts were read, not executed. Contract fixed in 46fb27d.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
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