七篇独立论文,每篇 = 1 份英文说明 + 1 份中文说明 + 1 支 C 验证器(= 1 个 exe)。 Seven self-contained papers: one English note, one Chinese note, one C verifier (one exe) each.
git clone https://github.com/zjaizlh8/github_upload.git && cd github_upload
make run # 编 7 支验证器并逐支实跑 / build + run the seven verifiersWindows 无 make 时 / without make:
gcc -Wall -Wextra -O2 -D_GNU_SOURCE -o t bsd-conjecture/code/verify_bsd.c -lm && ./t不随仓发布二进制 / no binaries are shipped in the repo:上面两条命令就地编出 7 个 exe;
打 v* tag 时 CI 才在 windows-latest + MinGW 上现场编译、逐支实跑、生成 SHA256SUMS.txt 并发 Release
(见 .github/workflows/verify.yml)。当前仓内与 Release 列表均无预编译产物。
The two commands above build the seven executables in place; only a v* tag makes CI compile
them on windows-latest + MinGW, run each one, write SHA256SUMS.txt and publish a Release. Neither
the repository nor the Release list carries prebuilt binaries today.
| # | 猜想 / conjecture | 英文说明 / English | 中文说明 / Chinese | 验证器(C) / verifier | 行数 / 独立读数 / 断言级独立 / rows · indep · assertion-level |
|---|---|---|---|---|---|
| 1 | Birch–Swinnerton-Dyer | paper.md |
paper.zh.md |
verify_bsd.c |
60 · 56 · 46 |
| 2 | Hardy–Littlewood(含孪生专章 / with the twin chapter) | paper.md |
paper.zh.md |
verify_hl.c |
35 · 35 · 35 |
| 3 | Hodge | paper.md |
paper.zh.md |
verify_hodge.c |
18 · 18 · 16 |
| 4 | Navier–Stokes | paper.md |
paper.zh.md |
verify_ns.c |
18 · 18 · 18 |
| 5 | P vs NP | paper.md |
paper.zh.md |
verify_pvsnp.c |
27 · 27 · 14 |
| 6 | Riemann Hypothesis | paper.md |
paper.zh.md |
verify_rh.c |
22 · 21 · 20 |
| 7 | Yang–Mills | paper.md |
paper.zh.md |
verify_ym.c |
18 · 18 · 9 |
| 合计 / total | 198 · 193 · 158 |
★ 2026-10-28(§158):这张表原来对不上账,已按台账重算。 原表首数合计 194(声称 197)、括号内合计 148(正文声称 151)。 两处差都来自「把某几篇的重复去掉了、另几篇没去」:
bsd已剔 3 条重复却仍算了 1 条非判据;rh的行级重复 1 条(rh:103#P4≡rh:106,同数9.694463e−11)已剔、断言级重复 2 条(§169)也剔。 ★ 2026-10-29(§169):154 / 44作废 →158 / 40。 去重轴试了三种机械规则:同篇同dev(dev=0一族 78 行互相碰撞 ⟹ 量的是 0)、同篇(dev, criterion)逐字相同(MECH_DUPs=100,100 条全是NA(zero-dev)占位符相等)、只在有名行上跑后者(覆盖 92/198)。⟹ 这条轴只有一半能跑:可跑的那半给rh22 行 1 组 3 行同标签 ⟹ 重复 2、独立 20;判断的那半(换参数实例化)照实标「人工判断·不可复算」。根因与 §168 同一个:106 行点不出判据名,任何以名字为键的规则都会退化成数占位符。 ★ 2026-10-29(§170):另开一轴 ——「论文里的数有没有来源」,结果是【已经守住】。 机械判(audit/chk_num_source3.sh):七篇正文里 197 个「像读数」的数(≥5 位有效数字),141 个在audit/outs/<paper>.txt里直接命中、40 个是正文截位(输印 8 位、文写 6 位)、16 个全仓无来源 —— 而这 16 个里 6 个是闭式常数(11/12、√0.9、π/100、4π、√3ħc/1710、2π)、10 个正文自带「归档 / 外部锚 / SKILL 归档 / 库内记录」标注 ⟹ 无标注 0 个。这条轴不需要修,记它是为了以后能回答「凭什么说论文里的数都可复现」。 现按同一口径逐篇重算:独立读数 = 行数 − 行级重复 − 非判据;bsd 60−3−1=56、rh 22−1=21。三个数合计 = 198 · 193 · 158。 ★ 2026-10-28(§165):pvsnp26 → 27。 新判项P6b= 秩律延拓到n=9,10(秩 65、82)且跨字流式实现与单字全消元在重叠域n=2..8逐 n 相等 —— 这是新增的一条判据(不是换参数实例),所以行数/独立读数/断言级三个数同时 +1(源:p-vs-np/code/verify_pvsnp.c的P6b行,实跑pvsnp27 PASS)。 ★ 2026-10-29(§167):rh22 · 17 · 11 → 22 · 21 · 20。 先算后用:rh的 7 行原先被记成「与rh:106同一读数」(P5\|chi\|3 行 + 段级RESULT maxdev4 行),逐行对台账dev核过后发现没有一个等于rh:106的9.694463e−11(2.220000e−16/0.000000e+00×2 /1.998e−15/3.609e−14/1.776e−15/2.842e−14)⟹ 它们不是rh:106的副本。但rh:110#P9与rh:118#P10同为dev=0且criterion逐字相同(P3/P9/P10 (zero-dev))⟹ 这一对确实是重复。⟹ 断言级按「同篇内(dev, criterion)逐字相同只算一次」这一可跑规则重算:rh22 行里仅 1 组 3 行同标签 ⟹ 重复 2、独立 20(verify_audit.c的INDEP[5]11 → 20,实跑audit/outs/audit.txt第 57 行印「rh总账 22 → 独立 20(重复 2)」)。§167 初版此处的1.625e−13是我凭印象补的数,台账与该篇输出里都不存在,已作废(见AUDIT_20261027_非被迫结构.md§167.2 的更正段)。
合计 198 行逐条读数,全部为有限格上的可复算断言;make run 七支全 rc=0。
Total 198 ledger rows of machine-checked readings on finite grids; make run returns rc=0 for all seven.
★ 第五个正交口径:
named(能不能逐条点到) —— 2026-10-29 实测(源:AUDIT_20261027_非被迫结构.md§171,机械判grep -cE "^ *(\[PASS\]|PASS )"对RESULT PASS=):bsd 60/60 · hl 35/35 · hodge 18/18 · ns 18/18 · pvsnp 27/27 · rh 22/22 · ym 18/18⟹ 可命名 198/198,缺口 0。 上午测得的rh 10/22、ym 9/18(当时记「177/198 = 89.4%」)已随两处修复消失:rh的 part1 只累加计数器不印逐条行(verify_rh.c的seg0_main现每条后接crit(name, v),12 条全印PASS <名> dev= tol=);ym的S0_judge把判决词印在行尾(现前置,与其余六篇同格式)。⟹ 这条轴现在是【一条正则 +gap 必须为 0】,能翻红。 ⟹ 「198 PASS」与「198 条都点得出名字」现在是同一个数。 缺口原来全在两篇的 part1:verify_rh.c的seg0_main与verify_ym.c的S0_judge。两处已按 §171 修完(rh加crit()逐条打印、ym把判决词前置),实跑named = RESULT逐篇全等。 Fifth orthogonal axis:named(can each reading be pointed at?) — measured 2026-10-29 (source:AUDIT_20261027_非被迫结构.md§171):bsd 60 · hl 35 · hodge 18 · ns 18 · pvsnp 27 · rh 22 · ym 18⟹ 198/198 nameable, gap 0. The morning reading (rh 10/22,ym 9/18, "177/198 = 89.4%") is superseded by the two fixes:rh's part-1 now prints one judged line per criterion viacrit(name, v), andym'sS0_judgenow prints the verdict token first. One regex plusgap must be 0makes the axis falsifiable.
★ 三个不同粒度的「去重」数必须分开写(G29),混用就是虚账:
口径 总 独立 重复 重复的定义 行级字面重复(同篇同 dev同tol,dev≠0)198 193 4 同一残差被印两遍( bsd3 对:1.578e−11/4.316e−10/8.230e−11·rh:103#P4≡rh:106的9.694463e−11);另有 1 行是RESULT maxdev不是判据 ⟹ 193 + 4 + 1 = 198(源:matrix/forcedness_ledger.csv按paper|dev分组实扫)断言级去重(同一律换参数实例化) 198 158 40 规则分两半:可跑的那半 = 同篇内 (dev, criterion)逐字相同只算一次(awk一行,实扫只有 1 组:rh的 3 行同标签 ⟹ 2 条重复);判断的那半 = 同一律换参数(pvsnprank 律 ×7、可解性 ×7、仿射秩 ×3 ⟹ 27 条只有 14 条独立;ym两段同名 P1..P9 ⟹ 18 条只有 9 条独立;hodge段二 P2 与 P6;bsd的 11 条逐字重复)—— 这半是人工判断,不可复算,照实标注(§169)对象已标 198 76 121 未标 121 = 106 条 criterion=NA(zero-dev)+ 15 条地址落空 ⟹ 记了数,没记这个数在量什么(§158)地址有效( line指向论文的非空行)198 112 68 落空 按篇去重后 180 个不同地址, sed -n "L p"取不到非空行的有 68 个(37.8%)⟹line是写作时的指针,不是读数;它随论文编辑漂移(§172)⟹ 任何一处只写一个数都算虚账。 对外可引的最硬说法是:198 行 · 193 条互不重复的读数 · 158 条互不重复的断言(其中可跑规则能复算的只有
rh那 2 条重复)· 其中 76 条对象已标。 另记:钝判据 32 条 ——NA(manual-read):这是人工读源码逐条转录的数,台账层不可复算(源:audit/verify_audit.cL6 已照此标注;台账层能复算的只有;same ∧ dev=0的 6 行与;same ∧ dev≠0的 18 行,差额 24 条必须到源码层解释 —— 源:matrix/tol_stat_out.txtL3b 实跑)。第二轮已把其中 9 条改成可翻红;第三轮(§174)另出一类、不并入这 32 条——ratio=dev/tol<1e−3的 5 行(全在rh,tol统一1e−08对dev≈1e−16~1e−14等于松10⁴–10⁸倍),那 5 行原被手填成1e−12/1e−15而伪装成硬判据,判据是tol_stat的 P2(源:matrix/tol_stat_out.txtL2/L3 实跑,md5921551b4497266cb0ba2a0ab67d79829);它是「阈值松」这一轴,与这 32 条的「人工转录」轴不可相加。余数照实留档,不删不改总数。详见AUDIT_20261027_非被迫结构.md§162 与 §174。audit/verify_audit.c的 L5 层逐篇核这两个数(重复处数必须已声明、独立条数 ≤ 总账),L6 层是钝判据台账。 ★ 2026-10-29(§175):台账六条轴合成一条命令 ——bash audit/verify_ledger.sh(X1–X6,全绿时退出码 0)。 它读下面这个LEDGER-AXES块与台账现算逐位比:声明可以假,但声明与台账不许对不上。三处齿测(dev改一位 / README 声明改一位 /class标签改错)分别打出FAIL=4 / 1 / 2。 ★ 2026-10-29(§176):论文正文里【抄的那段实跑输出】也进机械核 ——bash audit/chk_paper_quote.sh(围栏内[PASS]/[sanity]/RESULT/EXIT/k=/res=行,先按标签切段再比;#注释里的RESULT PASS=n与N · M 小计另走一桶)。现况 段 40 · 逐字 33 · 数字序列同 7 · 漂移 0。抓到并已修 4 处陈旧计数(bsd五段小计12…与RESULT PASS=57→15…与60;bsd复现块PASS=12/k=12/domain=12→15/15/15;p-vs-np两段17 · 9/PASS=26→18 · 9/27)+ 验证器里硬编码的「17 项」→ 活读 +[sanity]行被截短(丢掉dev=0.000e+00)。齿测:把注记改回PASS=26⟹DRIFT-C、EXIT=1。详见AUDIT_20261027_非被迫结构.md§176。
★ 2026-10-29(§177):锚换成【验证器输出的物理行号】,并把三条「没有结局」的轴建成表 ——
bash audit/verify_anchor.sh(A1–A6,已由verify_ledger.sh末尾转调、FAIL并入其退出码)。 三张表:audit/census2.csv(paper,seq,src,name,dev,tol,src=audit/outs/<篇>.txt的物理行号,二进制不变它就不动)·matrix/ledger_anchor.csv(ckind = named 77 / addr_fail 15 / no_object 106、status = forced 61 / ambiguous 137)· 组表(按(篇,块,标签,骨架)归组 = 183 组 / 198 行 ⟹ 参数点重复 15 行)。 最硬的一条是A2自证:每行按自己的src重读输出,必须仍是判决行、且含本行的dev字面与名字前 6 字 —— 表能证明自己。齿测:src改一位 ⟹A2红;锚表强制行seq改一位 ⟹A6红。 同时修掉一个真错:name轴原记92,106,其92 = 76(有名)+15(地址落空注记)+1(非判据)—— 把 15 条「地址落空注记」算成了名字 ⟹ 现记77,15,106(77 = 76 + 1非判据也有名)。Pn不是标识符:同篇同一Pn落在 ≥2 条判决行且骨架不同的有 46 组;rh最重(part1 的crit块P1..P12与 part2 的S1_judge块P1..P10同号不同物)⟹ 已把 part2 十条改印S1..S10(只改标签、阈值与计数一字未动,实跑仍22 PASS / 0 FAIL),rh撞号归 0;其余六篇照实报读数、不假装已修。详见AUDIT_20261027_非被迫结构.md§177。 ★ 2026-10-29(§178):拿画像去问「还有没有新被迫结构」,抽出一条、作废两条、更正一句 ——D:/dpgo/cx/resid_type.c(五道门禁全exit=0、-Wall -Wextra -O2零告警、RESULT PASS=6 FAIL=0 maxdev=2.223404e-07)。 抽出的那条 = 残差二型R/A:dev写成k·2^-52(k = m/2^j,m ≤ 64,j ≤ 8),恰好落在格点上的 11 / 60 = 18.33%;同 binade 同窗(RELW=2e-6)背景期望 0.008202 行、sigma=0.090565⟹ 偏离 ≈121σ(不是噪声)、富余 1341×(旧写法 0.000192 行 / 57292× 是「背景条件化在观测上」的算法错,已修,见 §182);跨bsd/ns/rh/ym四篇;归心R(R(F(dev)))−dev最大相对残差 2.223404e-07 ≤ 1e-6(界由printf %.6e截位给,不是拍的);两齿(一行推出格点 ⟹ R−1;一行 A 型置成8·2^-52⟹ R+1)都咬住 ⟹ 不是「凡是双精度都算 R」的恒真分类。层位照实:读数级结构 —— 它说这 11 个残差是同一二进制对象在不同 binade 的副本,不说为何是这 11 条。 作废两条:dev/ulp比值(对 2 的幂恒不变 ⟹ 重言式,判别力 0)· part 级零率(part 大小不均 ⟹ 量的是分组不是规律)。 更正一句:§177 的「dev≠0⟹ 有唯一键」不成立 —— 唯一可定位的 61 行 = 54(dev≠0)+ 7(dev=0但同篇tol不同),反向还有 6 行dev≠0不唯一;正解是(dev,tol)对互异 ⟹ 键,dev单独从不承重。 一条读数、不设判据:tol的十进制指数减dev的十进制指数,49/60 = 81.7% 落在 +1…+3(容器 = 输出侧census2,两侧量级算子统一用floor;台账侧只有 50 行有数值tol,同口径为 42/50 = 84.0%;例外集中在rh的crit()统一1e-08)⟹ 容差列在八成以上行上是dev的影子;且tol有数值 ⟺ 该行在对象面(对象面 187/187 有值;FLOW/TREND10 行全部撤回NA)⟹ 它更像【面】的标志位而不是独立测量,不携带独立信息。没有机制给出「该 +1 还是 +3」,故只作读数(与 §174.5「5 条钝判据全在rh」是同一件事的两面)。详见AUDIT_20261027_非被迫结构.md§178。
★ 2026-10-29(§179):把「容差拍脑袋」做成能翻红的机械判据 ——
awk -F',' -v MODE=report -f matrix/tol_mech.awk matrix/forcedness_ledger.csv(三条规则:R1 反抄(tol/dev ∈ (1.9,2.1)⟹ 界取自本次读数 ⟹ 自比 ⟹ 判TREND)· R2 机制界(tol/u ∈ [0.5,100]或criterion写1/N/h^p/收敛)· R3 其余(圆整界只报不判,它是松界、能 FAIL)· 另加 R0 豁免:dev=0的整数域行不判tol)。 实测TOTAL=198 行 · INDEP=183 组 · 重复=15(源:awk -F',' -v MODE=report -f matrix/tol_mech.awk matrix/forcedness_ledger.csv的输出行TOTAL=198 · R1_反抄=1 · OK_MACH=13 · OK_ORDER=3 · OK_ZERO=138 · NO_MECH=44):R1 命中 0(首跑时ym:84的tol/dev = 2.00097命中 ⟹ 该行已降TREND、tol清成NA;判据修好了它的第一个病人,读数随之归零)·OK_ZERO 138·OK_MACH 13·OK_ORDER 3·NO_MECH 44(松界,照实记未导出机制)。
★ 2026-10-29(§180):审计档被整档
write覆盖后已恢复 —— 恢复率 97.1%(9756/10044 行,166 个二级标题)。 恢复法 = 把会话转录(zstd 多帧)里对本文档的五类事件按时间回放(整档设值/唯一锚替换/heredoc 追加/printf追加/shell 原地替换),再按节号插回失败的替换、去重、按节标题反查补齐(D:/dpgo/cx/_recover/四个脚本)。承重影响 = 0:198 条判据与六轴数字全在台账 CSV 与七个脚本里,脚本一个字节没动,恢复后全链重跑全绿。缺节照实标(§13 §57 §75 §79 §96 §107 §115 §116 §122无人引用;§16 §74 §93 §117 §162引用悬空 —— 凭记忆重写等于伪造审计记录,故不补)。教训两条:长档禁write(只走追加与单点替换);可恢复性来自事件流,不是备份。
★ 2026-10-29(§181 → 2026-10-30 §184 更正):结构数据画像 v2 ——
bash audit/build_profile.sh(只读)。把「输出侧 198 条」与「声明侧 198 行」保序对齐:pair 50(名字可传)·lcs 139(只能靠行序,名字不可传)·none 9(源:bash audit/verify_ledger.sh的判项行[PASS] X7a … (pair,lcs,none = 50,139,9))。§184.2 更正:原记49/145/4与「相位七篇全 0」都是旧匹配键的产物 ——norm()曾把「读出来的零」与「由PASS推出的零」归一成同一个Z,段界消失后 LCS 在段内自配对;分Z0/ZI后相位真实为pvsnp 9·rh 10(源:verify_ledger.sh的[PASS] X7f 相位声明 = 现算(2,10)),对齐行dev不一致 30 → 0(源:node audit/join_audit.mjs的「dev 不同」计数行)。地址状态valid 112 · blank 83 · oob 3(源:bash audit/build_profile.sh末行[addr] 台账地址状态:blank=83 oob=3 valid=112;修掉一个真错:line列带#标签时整列转数得 0,曾把 49 行误判oob);criterion_final三分 台账名 86 ·paper:NNN55 ·NA(addr_dead)57 ⟹ 可命名 141/198(71.2%)(源:verify_ledger.sh的[PASS] X7b 判据文本三分合计 = 198(sigma/paper/NA_dead = 86,55,57);与 §173 的92/106是两个口径,不可混写)。§172 三选一定案 = B:地址轴只作读数不作判据(line会随排版漂移;承重的是输出侧物理行号——census2的src列,A2每行按自己的src重读自证)⟹ 名字只从输出侧往台账侧传,传不过去的 57 行标NA(addr_dead),不按行号猜名。 把能导出机制的bsd三条改成机制界:Ω 一致1e-8 → 1e-10(Simpson 累加界ε·NPTS·max|f|·4h/3)· Tamagawa1e-9 → 1.53e-10(锚截断 + Ω 累加,两源独立相加)· 倒数账相消1e-9 → 8.0e-10(16·½·10⁻¹⁰)。 最硬的一条是预测而非拟合:16·(L真 − LEXT) = 4.315e-10而实测4.316e-10(源:audit/outs/bsd.txt的倒数账两项相消行)⟹ 四位吻合 ⟹ 残差 100% 来自 10 位外部锚的截断(锚不许由恒等式反推)。 七行FLOW与bsd:50/51(「Abel 和趋零」)判为过程面 ⟹tol清NA、降TREND(结构面由 part3 整数域判据承担)。 齿audit/tooth_tol_mech.c(PASS=6 FAIL=0,源:audit/outs/tooth_tol_mech.txt,已挂run_chain.sh):Δ=+5e-11必须超界(咬得住)·Δ=+2e-11必须仍在界内(不过紧)·sdev(Δ)−sdev(0)=16Δ(残差6.6e-17)· 自比界 4 次扰动 0 次拒绝 ⟹ 判别力 0,机制界拒绝 1 次 ⟹ 两界不同形。 随台账同步:CLSCSV五档扩六档(加TREND,否则「没有这一格」会被算成「这一格是手填的」)·X2增tol=NA分支(只校dev)·profile_md5 → 8c563d0cf22f。复跑:run_chain.sh⟹ALL_FAIL_ZERO=1;verify_ledger.sh⟹FAIL=0 / EXIT=0。详见matrix/_tol_mech_20261029.md。
★ 2026-10-30(§184):判的界 ≠ 印的界 —— 三处结构缺陷,各自修到能翻红。 ① 中间产物陈旧:
census2由上一轮手工命令生成,bsd三条判据的tol早已收成机制界而它仍记旧值,X2只自证dev不证tol⟹ 背离藏了一整轮。修法三层:build_profile.sh第 0 步每次从当前输出再生成(并补回表头)+ 判据 X9(audit/x9_tol.awk:台账tol与输出tol在两侧都是数值的行上必须相等)+ 齿audit/tooth_x9.sh。现况BAD 0 TOT 161 PROC 0(源:bash audit/verify_ledger.sh的判项行[PASS] X9 台账 tol = 输出 tol(两侧都是数值的 161 行,对象面 0 处背离);齿RESULT PASS=4 FAIL=0)。X9 的判别力超出 X2:把一行tol改成"本篇输出里真有的另一个数",X2 会放过、X9 必然咬住。 ② 两种零同名:norm()把「读出的零」与「推断的零」归一成Z⟹ 段界消失、LCS 段内自配对,造出「相位全 0」的假象。分Z0/ZI后 join 变50/139/9、相位真实为pvsnp 9/rh 10、对齐行dev不一致30 → 0⟹ §181 的「无整体错位」作废(源:verify_ledger.sh的[PASS] X7a …与[PASS] X7f 相位声明 = 现算(2,10))。 ③ 判的界被印成另一个界:verify_rh.c的crit()原统一印1e-08⟹ 改为逐条给界(P6 1e-14,P7/P11/P12违例计数0);verify_ns.c的P4由1e-12改0.0(实测恰 0)。台账随计算对齐 ⟹ 两侧tol不一致21 → 10,余下 10 行全是过程面(X8 的设计)。(源:audit/outs/rh.txt的PASS P6行dev=3.552714e-15 tol=1.000000e-14·audit/outs/ns.txt的P4 locking行dev=0.000000e+00 tol=0.000000e+00) 零的三分:dev=0拆成 几何零 112 / 计数零 2 / 隐含零 24,加非零 60 = 198(源:node D:/lw/github_upload/matrix/zero_kind.mjs的分篇输出行)。 画像 v3:audit/profile_v3.awk把三条划分并进一张表 ——face 187/10/1·zk 112/2/24/60·ra R 11/A 138/X 49,每条都加回 198(源:verify_ledger.sh的[PASS] X10a 画像 v3 三条划分各自加回 198(face=187,10,1 · ra=11,138,49));X10a/X10b/X10c 判这三条与 README 声明逐位一致,其中 X10b 是双路交叉(画像的R 11== C 探针resid_type自印的11 / 60 = 18.33%,源:D:/dpgo/cx/resid_type_out.txt:12)。ra的容差必须由印位数给(固定1e-3会把 23 行误判成R;改成5e-7·m后 = 11)。 全链:七支198+audit 25+tooth 6+tooth8 4+tooth9 4= 237 条 PASS /FAIL=0/EXIT=0、ALL_FAIL_ZERO=1(源:bash audit/run_chain.sh末九行)。 ★ 2026-10-30(§185)遗留登记:能修的四类已修(产物陈旧 / 界与代码背离 / 两种零同名 / 三条划分无判据);不能修的三类登记成读数——缺 11 节原文(§13 §16 §57 §74 §75 §79 §93 §96 §107 §122 §162,照实标缺、不凭记忆补写)· 腐地址 86 = 83 空 + 3 越界(定案 B:line不承重,源:build_profile.sh末行[addr] 台账地址状态:blank=83 oob=3 valid=112)· 无名判据 57(86 + 55 + 57 = 198,可命名 71.2%)。X型 49 行(dev≠0而不可写成k·2^-52)是画像里最大的未解释残余:R1 无机制 ⟹ 不称律,R2 未算背景 ⟹ 不称噪声,只能标「未解释」。详见AUDIT_20261027_非被迫结构.md§184 / §185。
★ 2026-10-30(§186)独立复算复核:另一条实现
D:/dpgo/cx/_recover/verify_claims.mjs(不复用仓库任何 awk,声明一律现读LEDGER-AXES)逐轴复算 ⟹ 9 条 MATCH(198·class 138,43,6,7,1,3·invariance 153,23,4,2,15,1·dev 83,31,60,24·dev=0 114· 地址112,86(2026-10-30 中英改文后现算111,87,§187)·join 50,139,9·criterion 86,55,57·R 11),3 条REF(背景期望/sigma/富余:粗模型按 ulp 密度、探针按同窗RELW,不是同一对象,不硬凑相等),1 条READ(tol−dev指数差42/50,按 G54-R3 只作读数、不设判据)。复核中查出的三处假 MISMATCH 全在复核器自己(addr_state无表头却slice(1)· 判零按字面而非数值 · 格点容差1e-9未由印位数给)。详见AUDIT_20261027_非被迫结构.md§186。 ★ 2026-10-30(§187)中英 + 代码三方一致性:新增audit/check_lang_parity.mjs(L1–L6:标题层级 / 常数取值集合 / 逐节块公式与列表项 / 逐节散文量带 / 逐节常数集合 /RESULT PASS序列)与audit/tooth_lang_parity.mjs(T1–T7 七齿,各用独立副本并发)。改掉 7 处真不一致(ns 两处「发散被下一段收回」+ 英文缺Gs=0.125664· bsd 英文缺一整行读数 · rh 中文一处拼接残行与一处结论截断 · hodge 中英各缺三个最小值读数)。终态:不一致 0 · L1–L6 全 MATCH · 齿 7/7 咬住 · chk_paper_quote 漂移 0,run_chain ALL_FAIL_ZERO=1。改文的副作用照实记:台账地址读数112,86 → 111,87、画像指纹641a703e9069 → 1397c1827edf(§172 定案 B:指针不承重)。详见AUDIT_20261027_非被迫结构.md§187。
按"为什么能过"分五档(与去重是两个正交口径):
档 数 判据(凭什么能过) FORCED138 dev位相同 = 0 —— 结构/整数恒等,无容差可藏。但138要拆(§173):114行实跑真打出0.000000e+00,另24行之判决行里【没有dev=token】,零是PASS推的 ⟹ 台账已写NA(implicit)(源:matrix/tol_stat_out.txt的[L1]第四行D 隐含零 dev=NA(implicit) 24 行,md5921551b4497266cb0ba2a0ab67d79829,实跑RESULT PASS=6 FAIL=0)FORCED_eps43 dev > 0,且能指出来源:① 累加舍入(放大10⁰–10²:hodge14× ·ym17× ·ns84×);② 具名放大源(放大10⁶–10⁷:rh的Γ7.0e6× ·bsd的Γ/ζ1.9e6× ·hl的升阶乘级数 1.4e6×)FORCED_trunc7 dev > 0,且h与阶都可指(ym的1/N一阶 ·ns的h²二阶 ·hl:15的3/log t·hl:20的 Mertens 截断)FLOW7 流: dev是被界量或被界读数,随参数移动 —— 流本来就在动,不是"无机制"(含原TREND3)DUP3 重复行:被另一条判据逻辑蕴含( bsd两条 rank 由"分离度"蕴含;rh:103#P4与rh:106同数同对象),dev是被界的量本身 ⟹ 不计作独立判据
$$198 = 138 + 43 + 7 + 7 + 3$$ ★ 2026-10-29(§167):
BOUND档已清空、SEG档已并回判据。 原BOUND4 条与SEG4 条经复核都读出了可判的数(不是"只有界没有值")⟹ 按dev真相重分派:rh的 4 条段级RESULT maxdev里 3 条归FORCED_eps、1 条归DUP;原BOUND4 条归FLOW/FORCED_eps。⟹ 一张表上的"档"不是标签,是dev的读数;读数为 0 就必须进FORCED。★ 2026-10-29(§173):
dev列已按【实跑原值】重录,台账与实跑逐类相等。 三件:①rh8 行 +ym4 行原按 4 位有效数字转抄(1.562000e−09),而实跑印 7 位(1.562375e−09)⟹ 已换回输出原串(12处替换、48处本已一致、0处无对应,源:audit/fix_dev.awk实跑 stderr 计数行);②pvsnp 18 + hl 6 = 24行的判决行里没有dev=token,零是PASS推的 ⟹ 台账改记NA(implicit);③ 由此台账与实跑逐类相等:114印出零 +60非零 +24隐含 =198(源:matrix/tol_stat_out.txt的[L1]四行,md5921551b4497266cb0ba2a0ab67d79829;audit/reconcile.sh实跑七篇逐篇 0 条无支撑)。可迁移一句:台账里的数必须【逐字】能在输出里 grep 到,否则它是【推断】不是【读数】。两条已钉死的读法纪律:①
dev列能判档,tol列不能(拆分真值:四类表83 / 31 / 60 / 24——83行dev=0∧tol=0(阈值就是 0)、31行dev=0∧tol>0(实测精确,tol只是余量、不参与判定)、60行dev≠0(tol真参与判定,其ratio=dev/tol55/60落在[1e−3, 1):既无「松 1000 倍以上」也无「贴脸 ≥0.99」;另5行低到1e−9–1e−6,全部在rh—— 因为rh的crit()用统一阈值1e−08,对dev≈1e−15等于松了10⁷倍 ⟹ 这5条是【可复算的钝判据】)、24行dev=NA(implicit)(判决行无dev=token,零是推的 ⟹ 不参与任何判定;§173 之前写作107/31/60,其107=83+24)⟹ 一列装两种语义(138行阈值 /31行余量),但数值本身健康(源:matrix/tol_stat_out.txt的[L1]四行,198 行台账实跑,md5921551b4497266cb0ba2a0ab67d79829),而判定只认dev;源:D:/dpgo/cx/tol_stat.c实跑PASS=6 FAIL=0,输出matrix/tol_stat_out.txt;详见AUDIT_20261027_非被迫结构.md§161/§173);② "小" ≠ "浮点" ——dev小到1e−6也可能只是截断,必须指得出放大机制或h。③ 衍生列必须重算:第 6 列ratio是dev/tol的函式,不是独立读数 ——§174改过 7 行tol之后它立刻陈旧(rh 71#P1记0.0020,而dev/tol = 1.998e−07,差10⁴),另有 7 行按 2 位有效数字手抄(bsd 42记0.0016,真值1.578e−03)⟹ 已全列重算(audit/fix_ratio.awk:NA(implicit)记NA(implicit)、tol=0记0.000000e+00、其余%.6e(dev/tol)),复检91逐位一致 +83阈值零 +24隐含 =198,不一致0(源:matrix/forcedness_ledger.csv实跑awk复检,md5ce11ffff987bd275605291b59f401233);同一轮还把第 4 列剩下 10 行手写的0换回各自行原串(0.000e+00/0.000000e+00)⟹ 现在零也逐字可查。另有第三个正交口径:
invariance(切片无关性) —— 台账第 8 列:
值 条数 含义 theorem23 陈述对全部切片成立(有证明来源 + 对象关系),本次只抽了几个点 sample153 只在该切片读过,跨切片未知 WALL1 墙的坐标: hodge:23= 霍奇猜想本身 —— 不是"未完成",是这条路唯一入口的标记na15 行号指向空行/标题/表格/方法论 ⟹ 地址落空(见下方更正:落空 ≠ 无判据) cited4 该数是从别处引用来的,不在这里实读 ⟹ 既不是 theorem也不是sampledup2 重复行
$$198 = 23 + 153 + 4 + 1 + 15 + 2 \quad(\text{theorem} + \text{sample} + \text{cited} + \text{WALL} + \text{na} + \text{dup})$$ ★ 2026-10-29(§167):本表原来少报
cited4 条、多报seg4 条(seg已并回判据、CSV 实扫为 0)⟹ 表面合计 198、实际两处错位。现按 CSV 六档重写(源:awk -F, '{v[$8]++}' matrix/forcedness_ledger.csv)。★★ 2026-10-28 更正(§158):
na的判法不成立。 本轮实测line列不能复建: 它既不指向paper.md的行,也不指向verify_*.exe的输出行,且无统一偏移。 ⟹ 「行号指向空行」是【这一列坏了】的性质,不是【那条判据不存在】的证据。 这 15 行的真状态是 对象未知,不是「不是判据」。⟹ 把五档按对象是否可定名重新切一刀(互斥·合计 198):
桶 条数 说明 E 对象已标 76 72 条真独名 + 4 条 dup(dup 也带名 ⟹ 计入 E,不另立桶) D 对象未标· NA(zero-dev)106 只记了「dev=0」,没记在量什么 A 对象未标·地址落空 15 line坏,dev=0又不能定名。注意口径:15 是「invariance=na这个子集里地址坏的」,全 198 行口径下地址落空是 68 个(§172),两个数都对但不可混写(G29)B 确认非判据 1 bsd:149= 段RESULT maxdev,不是判据合计 198 76 + 106 + 15 + 1 = 198 ✓ ★ 2026-10-29 更正(§167):原表把「C 重复 8」另立一桶,而「E 76 = 68 + 8 dup」里已含这 8 条 ⟹ 五行相加 206 ≠ 198(虚账)。现按「dup 归 E、只作标记」重写;且 dup 数按 §167 的实扫从 8 降到 4(
rh的两组同数读数经重跑已各自独立)。
$$198 = \underbrace{193}_{\text{独立读数}} + \underbrace{4}_{\text{重复}} + \underbrace{1}_{\text{确认非判据}}$$ ⟹ 本仓的准确说法是「193 条独立读数 + 4 条重复 + 1 条非判据」,且其中只有 76 条【对象已标】。 不写「178 条判据」 —— 那个数把 15 条地址落空的行当成了「不存在」,而证据不成立。
为何
dev救不了这 121 条(本轮实测·被迫负结果):$$\boxed{\texttt{dev}\ne 0 \Longrightarrow \texttt{dev} \text{ 是唯一键(24/24 唯一命中)} ;\Big|; \texttt{dev}=0 \Longrightarrow \texttt{dev} \text{ 不是键(106/106 多义)}}$$ 零残差是所有恒等判据的共同读数,它不含对象信息 ⟹ 「数」不能反查「对象」。invariance的判定方法见audit/SUBMIT_定理判定方法.md。五档的判法都写在
audit/GATE_RULES_STRUCTURE.md与AUDIT_20261027_非被迫结构.md§28–§63;逐条台账见matrix/forcedness_ledger.csv。TREND只有 3 条,且每条都在审计档里写了"为什么必须留"——不是欠查,是无法用阶来度量。By "why it passes", six classes (orthogonal to the de-duplication count):
Class n Criterion (why it passes) FORCED138 devis bit-identical0— structural / integer identity, no room for tolerance. But138splits (§173):114rows really print0.000000e+00; the other24rows have a verdict line with nodev=token at all, so their zero is inferred fromPASSand the ledger now recordsNA(implicit)(source:matrix/tol_stat_out.txt[L1]line 4 —D 隐含零 dev=NA(implicit) 24 行, md5921551b4497266cb0ba2a0ab67d79829, runRESULT PASS=6 FAIL=0)FORCED_eps43 dev > 0with a named source: ① accumulation rounding (amplification10⁰–10²:hodge14× ·ym17× ·ns84×); ② named amplification (10⁶–10⁷:rh'sΓ7.0e6× ·bsd'sΓ/ζ1.9e6× ·hl's rising-factorial series 1.4e6×)FORCED_trunc7 dev > 0with bothhand the order identifiable (ym's1/Nfirst order ·ns'sh²second order ·hl:15's3/log t·hl:20's Mertens truncation)FLOW7 flow: devis the bounded quantity / a moving reading — a flow moves by nature; "no mechanism" was the wrong label (wasTREND3)DUP3 duplicates: logically implied by another criterion ( bsd's two rank rows follow from the separation judge;rh:103#P4andrh:106are the same number on the same object);devis the bounded quantity itself ⟹ not counted as an independent criterion
$$198 = 138 + 43 + 7 + 7 + 3$$ ★ 2026-10-29 (§167): the
BOUNDclass is now empty and theSEGclass was merged back into criteria. Re-checking the 4BOUNDrows and the 4SEGrows showed each one does read a judgeable number (not "a bound with no value") ⟹ re-dispatched by whatdevactually reads: 3 ofrh's 4 section-levelRESULT maxdevrows areFORCED_eps, 1 isDUP; the oldBOUND4 areFLOW/FORCED_eps. ⟹ a class on this table is not a label, it is whatdevreads; if it reads 0, it must go toFORCED.★ 2026-10-29 (§173): the
devcolumn has been re-recorded from the run's own strings, and the ledger now matches the run bucket for bucket. Three moves: ①rh8 rows +ym4 rows had been transcribed at 4 significant digits (1.562000e−09) while the run prints 7 (1.562375e−09) ⟹ replaced by the verbatim output string (12replaced,48already identical,0unmatched; source: the stderr count line ofaudit/fix_dev.awk); ② forpvsnp 18 + hl 6 = 24rows the verdict line carries nodev=token at all, so the zero is inferred fromPASS⟹ the ledger now recordsNA(implicit); ③ hence ledger = run, bucket for bucket:114printed zero +60nonzero +24implicit =198(source: the four[L1]rows inmatrix/tol_stat_out.txt, md5921551b4497266cb0ba2a0ab67d79829;audit/reconcile.shrun over all seven papers: 0 unsupported rows in each). One transferable line: every number in a ledger must be greppable, verbatim, in the output — otherwise it is an inference, not a reading.Two reading rules, now fixed: ① the
devcolumn decides the class, thetolcolumn does not (the split, measured: four buckets83 / 31 / 60 / 24—83rowsdev=0 ∧ tol=0(the threshold is zero),31rowsdev=0 ∧ tol>0(exact in practice, sotolis only headroom and never enters the test),60rowsdev≠0(wheretolreally is the threshold, andratio=dev/tollands in[1e−3, 1)for55/60— none is ≥1000× looser than itsdev, none is a ≥0.99 near-copy; the other5sit at1e−9–1e−6, all inrh, becauserh'scrit()uses one1e−08threshold, which againstdev≈1e−15is10⁷×of headroom ⟹ those5are the computable blunt criteria),24rowsdev=NA(implicit)(the verdict line carries nodev=token, so the zero is inferred and enters no test; before §173 this read107/31/60, where107=83+24) ⟹ one column carries two semantics (138threshold /31headroom) while the numbers themselves are sound (source: the four[L1]rows inmatrix/tol_stat_out.txt, measured on the 198-row ledger, md5921551b4497266cb0ba2a0ab67d79829), and classification readsdevonly; source:D:/dpgo/cx/tol_stat.c, runPASS=6 FAIL=0, outputmatrix/tol_stat_out.txt; seeAUDIT_20261027_非被迫结构.md§161/§173); ② "small" ≠ "float-limited" — adevas small as1e−6may still be pure truncation; either name the amplification mechanism or giveh; ③ derived columns must be recomputed — column 6ratiois a function ofdev/tol, not an independent reading, so after§174rewrote 7tolvalues it went stale at once (rh 71#P1says0.0020wheredev/tol = 1.998e−07, off by10⁴) and 7 more rows were hand-copied to two significant digits (bsd 42says0.0016, true1.578e−03) ⟹ the whole column has been rebuilt (audit/fix_ratio.awk:NA(implicit)staysNA(implicit),tol=0becomes0.000000e+00, otherwise%.6e(dev/tol)), re-checked91bit-consistent +83zero-threshold +24implicit =198,0inconsistent (source:matrix/forcedness_ledger.csvre-checked byawk, md5ce11ffff987bd275605291b59f401233); the same pass replaced the last 10 hand-typed0s in column 4 with each row's own printed literal (0.000e+00/0.000000e+00) ⟹ zeros are now greppable too.★ 2026-10-29 (§176): the run's output as quoted inside the papers is now machine-checked too —
bash audit/chk_paper_quote.sh. It reads only inside ``` fences, splits each line at the segment labels ([PASS]/`[sanity]`/`RESULT`/`EXIT`/`k=`/`res=`) so that a joined `RESULT … EXIT=0` line is not misjudged, and compares each segment against `audit/outs/.txt` under the same label (verbatim = `exact`; identical digit sequence = `num`, i.e. translated; otherwise `drift`). Lines starting with `#` go to a separate bucket: any `RESULT PASS=n` or `N · M subtotals` they assert must be a value that really appears in that paper's `RESULT` lines. Current state: 40 segments · 33 verbatim · 7 same digit sequence · 0 drift. It caught and we fixed 4 stale counts (`bsd` five-part subtotals `12 · 13 · 14 · 10 · 8` + `RESULT PASS=57` → `15 · 13 · 14 · 10 · 8` + `60`; `bsd`'s repro block `PASS=12 / k=12 / domain=12` → `PASS=15 / k=15 / domain=15`; `p-vs-np` two-part `17 · 9` + `PASS=26` → `18 · 9` + `27`), 1 hard-coded "17 items" inside `verify_pvsnp.c` (now read live), and 1 truncated `[sanity]` line in `bsd` whose cut-off tail was exactly `dev=0.000e+00`. Tooth test: setting the note back to `PASS=26` prints `DRIFT-C` and exits 1. The reading rule: never judge a quoted block by "does this number appear somewhere" — numbers appear everywhere in the output (that is how `PASS=12` survived next to `tol=1e-12`); judge it by the label-scoped whole segment. See `AUDIT_20261027_非被迫结构.md` §176.
★ 2026-10-29 (§178): the profile was then asked "is there any new forced structure in here?" — one found, two retracted, one sentence corrected —
D:/dpgo/cx/resid_type.c(all five gatesexit=0,-Wall -Wextra -O2with zero warnings,RESULT PASS=6 FAIL=0 maxdev=2.223404e-07). What was found = the two-type residualR/A: writingdev = k·2^-52(k = m/2^j,m ≤ 64,j ≤ 8), 11 / 60 = 18.33 % land exactly on a lattice point; the same-window background (RELW = 2e-6) within the same binade is 0.008202 rows,sigma = 0.090565⟹ ≈121σ (not noise), surplus 1341× (the old 0.000192 rows / 57292× came from conditioning the background on the observed neighbourhood and is retracted, see §182); spanning four papers (bsd,ns,rh,ym); the round tripR(R(F(dev))) − devhas max relative residual 2.223404e-07 ≤ 1e-6 (the bound comes fromprintf %.6etruncation, not from a guess); both teeth bite (push one R row off the lattice ⟹ R−1; force one A row to8·2^-52⟹ R+1), so it is not the tautology "every double counts as R". Level stated honestly: a reading-level structure — it says these 11 residuals are copies of one binary object across binades; it does not say why these 11. Two retractions: thedev/ulpratio (invariant under powers of two ⟹ tautology, discriminative power 0) and part-level zero rates (parts differ in size ⟹ measuring the grouping, not a law). One correction: §177's "dev≠0⟹ there is a unique key" is false — of the 61 uniquely locatable rows, 54 havedev≠0and 7 havedev=0because theirtoldiffers within the paper, while 6 rows withdev≠0stay ambiguous; the correct statement is a distinct(dev,tol)pair ⟹ a key, anddevalone never carries the weight. One reading, no criterion: the decimal exponent oftolminus that ofdevlands in +1…+3 for 49/60 = 81.7 % (container = the output-sidecensus2, both decade operators nowfloor; on the ledger side only 50 rows carry a numerictol, giving 42/50 = 84.0 % under the same operator; exceptions cluster inrhbecausecrit()uses a uniform1e-08) ⟹ the tolerance column is a shadow ofdevin more than four fifths of rows and carries no independent information. No mechanism says whether it should be +1 or +3, so it stays a reading (the same fact as §174.5 "all 5 blunt criteria sit inrh"). SeeAUDIT_20261027_非被迫结构.md§178.
★ 2026-10-29 (§183):
tol的有无现在是判据**,不是约定 —— 判项 X8(audit/x8_face.awk+audit/verify_ledger.sh)。** 口径(用户定案,照引):「tol有数值 ⟺ 该行在对象面(FORCED*,允许非零dev),tol=NA⟺ 该行在过程面(FLOW/TREND,只报"走到哪儿")⟹ 混一面就FAIL,没有一行能例外。」 实测(198 行现算):对象面FORCED 138 + FORCED_eps 43 + FORCED_trunc 6 = 187行 全部有数值tol(187/187);过程面FLOW 7 + TREND 3 = 10行 全部NA(10/10);DUP 1行是重复标记、不属两面;例外 0 行。verify_ledger.sh⟹X8a 187/187 · X8b 10/10 · X8c bad=0 unk=0 · X8d 187+10+1=198 覆盖 · RESULT FAIL=0。 判别力(拿掉它,假的照样过):audit/tooth_x8.sh现造三个只改一行的反例 ——c1过程面带tol(hl:7→3e+00,这个值真出现在输出里,所以 X2 对得上、不算错)、c2对象面缺tol、c3面桶漏格(class=WEIRD)⟹ 旧审计对c1/c3是FAIL=0(判别力 0),加 X8 后FAIL=3 / 3 / 4,真台账仍FAIL=0(齿PASS=4 FAIL=0)。已挂audit/run_chain.sh⟹ 全链ALL_FAIL_ZERO=1(七篇60/35/18/18/27/22/18+audit 25+tooth 6+tooth8 4= 229 PASS / 0 FAIL)。 与 §178.5 不冲突:新增的是有无判据(面划分),至于tol该给多大仍不可判(照样只报z,倍数只作辅助)。详见AUDIT_20261027_非被迫结构.md§183。
★ 2026-10-29 (§183): whether
tolexists is now a criterion, not a convention — item X8 (audit/x8_face.awk+audit/verify_ledger.sh). Rule (user's decision, quoted): "numerictol⟺ the row is on the object face (FORCED*, where a non-zerodevis allowed);tol=NA⟺ the row is on the process face (FLOW/TREND, which only reports how far it got)" — mixing the two faces is aFAIL, with no exception row. Measured (198 rows, recomputed): object faceFORCED 138 + FORCED_eps 43 + FORCED_trunc 6 = 187rows carry a numerictol(187/187); process faceFLOW 7 + TREND 3 = 10rows are allNA(10/10);DUP 1is a duplicate marker, not a face; 0 exceptions.verify_ledger.sh⟹X8a 187/187 · X8b 10/10 · X8c bad=0 unk=0 · X8d 187+10+1=198 covered · RESULT FAIL=0. Discriminative power (remove it and the fakes still pass):audit/tooth_x8.shbuilds three one-line mutants —c1a process-face row with a numerictol(hl:7→3e+00, a value that really is printed in the output, so X2 matches it and sees no error),c2an object-face row missingtol,c3a face bucket that does not exist (class=WEIRD) ⟹ the old audit gaveFAIL=0onc1/c3(power 0); with X8 they giveFAIL=3 / 3 / 4while the real ledger staysFAIL=0(toothPASS=4 FAIL=0). Wired intoaudit/run_chain.sh⟹ whole chainALL_FAIL_ZERO=1(papers60/35/18/18/27/22/18+audit 25+tooth 6+tooth8 4= 229 PASS / 0 FAIL). Not in conflict with §178.5: what was added is a presence criterion (which face the row is on); how largetolshould be remains unjudgeable (still onlyz, with the ratio as an auxiliary). SeeAUDIT_20261027_非被迫结构.md§183.
A third orthogonal axis:
invariance(slice-independence) — column 8 of the ledger:
Value n Meaning theorem23 the statement holds on all slices (proof source + object relation given); we sampled a few points sample153 read on that slice only; behaviour across slices unknown WALL1 the wall's coordinate: hodge:23= the Hodge conjecture itself — not "unfinished", but the marker of this road's only entrancena15 the line points to a blank / header / table / methodology line ⟹ address fell through (see the correction below: falling through ≠ no criterion) cited4 the value is quoted from elsewhere, not measured here ⟹ neither theoremnorsampledup2 duplicates
$$198 = 23 + 153 + 4 + 1 + 15 + 2 \quad(\text{theorem} + \text{sample} + \text{cited} + \text{WALL} + \text{na} + \text{dup})$$ ★ 2026-10-29 (§167): this table used to under-report
citedby 4 and over-reportsegby 4 (segwas merged back into criteria; the CSV says 0) ⟹ the printed total was 198 while two entries were mis-slotted. Now rewritten from the CSV's six values (source:awk -F, '{v[$8]++}' matrix/forcedness_ledger.csv).★★ Correction 2026-10-28 (§158): the
narule does not hold. This round measured that thelinecolumn cannot be reconstructed: it indexes neither the lines ofpaper.mdnor the output lines ofverify_*.exe, and there is no uniform offset. ⟹ "the line number points at a blank" is a property of this broken column, not evidence that that criterion does not exist. The true status of those 15 rows is object unknown, not "not a criterion".Re-cut by whether the object can be named (disjoint, sums to 198):
Bucket n Meaning E object named 76 72 uniquely named + 4 dup rows (dups carry names ⟹ they live inside E, not in a separate bucket) D object unnamed · NA(zero-dev)106 only "dev=0" was recorded, not what is being measured A object unnamed · address fell through 15 lineis broken anddev=0cannot name itB confirmed non-criterion 1 bsd:149= a segmentRESULT maxdev, not a criteriontotal 198 76 + 106 + 15 + 1 = 198 ✓ ★ Correction 2026-10-29 (§167): the old table gave "C duplicates 8" its own bucket while "E 76 = 68 + 8 dup" already contained those 8 ⟹ the five rows summed to 206 ≠ 198 (a phantom balance). Now dups live inside E as a marker; and by §167's re-scan the dup count drops 8 → 4 (two
rhgroups that shared one number turned out to be independent once re-run).
$$198 = \underbrace{193}_{\text{independent readings}} + \underbrace{4}_{\text{duplicates}} + \underbrace{1}_{\text{confirmed non-criterion}}$$ ⟹ The accurate phrase is "193 independent readings + 4 duplicates + 1 non-criterion", of which only 76 have a named object. Not "178 criteria" — that number counted the 15 fell-through rows as nonexistent, and that evidence does not hold.
Why
devcannot rescue those 121 rows (this round's forced negative result):$$\boxed{\texttt{dev}\ne 0 \Longrightarrow \texttt{dev}\text{ is a unique key (24/24 unique hits)} ;\Big|; \texttt{dev}=0 \Longrightarrow \texttt{dev}\text{ is not a key (106/106 ambiguous)}}$$ A zero residual is the common reading of every identity criterion; it carries no object information, so the number cannot be turned back into the object. Howinvarianceis decided:audit/SUBMIT_定理判定方法.md. Distinct count: 198 = 158 distinct + 40 duplicates (parameter repeats, the same identity counted twice, and one object counted across sections — only 2 of the 40 are produced by a runnable rule; the rest is editorial judgment and is labelled as such; see §169). 32 blunt criteria (booleandevor same-source self-comparison) are recorded, not silently dropped; nine of them were made falsifiable in round two; round three (§174) adds a kind of its own, which does not merge into these 32 — the 5 rows withratio=dev/tol<1e−3(all inrh, where one1e−08threshold againstdev≈1e−16…1e−14is10⁴–10⁸×of headroom), previously hand-entered as1e−12/1e−15and therefore disguised as tight criteria; decided bytol_stat's P2 (source:matrix/tol_stat_out.txtL2/L3, md5921551b4497266cb0ba2a0ab67d79829). That is the "loose threshold" axis and cannot be added to the "hand transcription" axis above.NA(manual-read): that 32 is a hand transcription from reading the sources — it is not recomputable from the ledger (source:audit/verify_audit.cL6 now says so). What is recomputable at ledger level:;same ∧ dev=0→ 6 rows and;same ∧ dev≠0→ 18 rows, so the 24-row gap must be settled at source level (source:matrix/tol_stat_out.txtL3b, measured on the 198-row ledger); seeAUDIT_20261027_非被迫结构.md§162.
孪生素数作为 Hardy–Littlewood 的 k=2 特例,写在同篇的附章里(hardy-littlewood/paper.md 的
## Chapter: the twin-prime case),共用 2C₂·Li₂ 常数与同一支验证器 —— 论文仍是七篇。
The twin-prime conjecture is a chapter of Hardy–Littlewood (k=2, same 2C₂·Li₂ constant),
not a separate paper — the count stays at seven.
中文
| 验证器 | 实跑读数 | 它落实的被迫链(机器见证) | 容器(吃什么输入) |
|---|---|---|---|
verify_bsd.c |
57 判据全过 · max dev 7.761e-03(part 2 的 rank-2 读数;part 5 账目另计 8.238e-11) |
j=1728 ⟹ CM;Δ=2⁶ ⟹ 唯一坏素数 2 ⟹ ∏c_p=c₂;Tate (v(Δ),v(c₄))=(6,4) ⟹ I₀*;账 + Cassels 平方类 ⟹ (c₂,|Ш|)=(2,1);Kolyvagin 切循环 ⟹ rank=0 |
32a1(y²=x³−x)与同族有限曲线;L(E,1)、Ω 各读一次 10 位 |
verify_hl.c |
35 判据全过 · maxdev 1.105(P1 是无量纲比值判据,阈值 3.0,不是误差量) |
筛到 √N 完备(对合配对);完备链数闭式 2^{6n−6}(实测 n=1..3);互反账 ∏(q−1)/∏q = 16/77(真约分)。注意:承重链 ΣΔ 在本仓没有代码路径,双账本 S=K₀−ΣK / M_nat=K₀P 是框架陈述、未见证 |
N≤10⁹ 五档;π₂ 与 archive 对表 |
verify_hodge.c |
18 判据全过 · max dev 2.998e-15(段一;段二 0 是整数违例计数,不是浮点残差) |
圆平均 = 投影(Φ_M 整除 ⟹ Σζ^{dk}=M[M\mid d]);rank P=Σ_p C(g,p)²;速率 N/n 与分离门槛 M≥2g+1;术层 N_p=C(g,p)²(阶梯比嵌套) |
g≤4 的模型层;圆作用不动块 |
verify_ns.c |
18 判据全过 · max dev 1.869e-14(段一)/ 4.999e-01(段二,P8 是无量纲比判据 tol=1.0) |
tr S=Σλ=0 ⟹ λ₁=−(λ₂+λ₃);三分恒等 P=λ₂‖ω‖²+Q₊−Q₋;T³ 周期 ⟹ 分部求和无边界项;2s>n+2 ⟹ s≥3;重联严格下降 0.748<1 |
T³ 周期域、牛顿、ν>0、H^s(s≥3) |
verify_pvsnp.c |
26 判据全过 | per 系数恒 +1(相消只能来自元素);mod 2 商秩亏 2n−2;n! 对 n^{cd} 反超(cd=1..21 交叉点 3…29) |
以引理 4.1(VP≠VNP)为条件;GF(2) 有限表 |
verify_rh.c |
22 判据全过 · max dev 1.562e-09 |
|χ|=1 ⟺ σ=1/2(根唯一);Λ 对合反射;Hardy Z 实性;P10 谱分桶 c=(4,4,7)、号差 (0,0,1)、Σm=7、c_0=1 |
五组零点样本;Euler–Maclaurin 截断 |
verify_ym.c |
18 判据全过 · max dev 2.044e-02 |
λ_l=l(l+2)、λ₀ 单重 ⟹ Δ=√3·Λ 是相减;格点规范不变量恒等式;欧拉误差比 |
格点规范场;4D SU(3) 物理扇区的连续极限是引理 4.1 |
一句话:每支跑的是该篇的被迫链 —— 链上每个 ⟹ 对应一条恒等式或计数恒等,逐条打印读数;第四列写的是容器吃什么(输入空间),不是「没做的那部分」。
English
| verifier | what it prints | the forced chain it discharges (machine witness) | the container (what it eats) |
|---|---|---|---|
verify_bsd.c |
57 pass · max dev 7.761e-03 (the part-2 rank-2 reading; the part-5 ledger is 8.238e-11) |
j=1728 ⟹ CM; Δ=2⁶ ⟹ only p=2 is bad ⟹ ∏c_p=c₂; Tate (v(Δ),v(c₄))=(6,4) ⟹ I₀*; ledger + Cassels square class ⟹ (c₂,|Ш|)=(2,1); Kolyvagin cuts the cycle ⟹ rank=0 |
32a1 (y²=x³−x) and its finite family; L(E,1) and Ω read once, 10 digits |
verify_hl.c |
35 pass · maxdev 1.105 (P1 is a dimensionless ratio criterion, tol 3.0 — not an error quantity) |
the sieve to √N is complete (involution pairing); the closed-form chain count 2^{6n−6} (enumerated for n=1..3); the exact 16/77 reduction. Note: the load-bearing ΣΔ chain has no code path in this repo — the two ledgers S=K₀−ΣK / M_nat=K₀P are a framework statement, not witnessed |
five tiers up to N≤10⁹; π₂ tabulated against the archive |
verify_hodge.c |
18 pass · max dev 2.998e-15 (section 1; section 2's 0 is an integer violation count, not a floating-point residual) |
circle average = projection (Φ_M divides ⟹ Σζ^{dk}=M[M\mid d]); rank P=Σ_p C(g,p)²; rate N/n and separation threshold M≥2g+1; technique layer N_p=C(g,p)² by nesting the ladder ratio |
model layers g≤4; invariant blocks of the circle action |
verify_ns.c |
18 pass · max dev 1.869e-14 (section 1) / 4.999e-01 (section 2, P8 is a dimensionless ratio with tol 1.0) |
tr S=Σλ=0 ⟹ λ₁=−(λ₂+λ₃); three-part identity P=λ₂‖ω‖²+Q₊−Q₋; T³ periodicity ⟹ no boundary term in summation by parts; 2s>n+2 ⟹ s≥3; reconnection strictly lowers, 0.748<1 |
periodic T³, Newtonian, ν>0, H^s(s≥3) |
verify_pvsnp.c |
26 pass | per has coefficient +1 throughout (cancellation can only come from the entries); the mod 2 quotient is rank-deficient by 2n−2; n! overtakes n^{cd} (cd=1..21, crossings 3…29) |
conditional on Lemma 4.1 (VP≠VNP); finite GF(2) tables |
verify_rh.c |
22 pass · max dev 1.562e-09 |
|χ|=1 ⟺ σ=1/2 (unique root); Λ involution; reality of Hardy Z; P10 spectrum buckets c=(4,4,7), sign (0,0,1), Σm=7, c_0=1 |
five zero samples; Euler–Maclaurin truncation |
verify_ym.c |
18 pass · max dev 2.044e-02 |
λ_l=l(l+2), λ₀ simple ⟹ Δ=√3·Λ is a subtraction; lattice gauge-invariant identities; Euler error ratio |
lattice gauge fields; the 4D SU(3) continuum limit is Lemma 4.1 |
One line: each binary runs that paper's forced chain — every ⟹ on the chain is an identity or a
count, printed line by line; column four states what the container eats (the input space), not "the part
that was not done".
中文
三层答复,逐层都能当场核。能被问倒的只有一种情况:把 B/C 层的数说成 A 层。
| 层 | 含义 | 机器判据 | 本仓实例(盒子 → 解数) |
|---|---|---|---|
| A 枚举唯一 | 该数是某个整数方程组的解,且在给定值域内解集基数 = 1 | 把盒子完整枚举、数解数;解数 >1 就不算被迫 | bsd 重数向量 (4,3,2,2,2,2):15504 组 → 1 组(只留迹 232、只留积 100,两条都在承重);bsd 的 (c₂,|Ш|)=(2,1):48 组 → 1 组;ym 的 min|3q²−p²|(q≤200)= 1,而 0 不可能(3 非平方)⟹ 间隙被迫为 1 |
| B 恒等式 | 该数由对所有参数成立的恒等式给出 | 参数扫一遍,逐点残差 0 |
hodge 的 Σ_p C(g,p)² = C(2g,g)(g=1..4:2/6/20/70);pvsnp 的 n! > n^{cd} 交叉点(cd=1..21 → 3…29);ns 的 2s>n+2 ⟹ s≥3;rh 的 |χ|=1 ⟺ σ=1/2 |
| C 读数 | 该数是测出来的,不承担被迫性 | 每个数都要有出处(本仓实跑 / 归档探针 / 外部锚) | hl 的 ρ 三档、ns 的 0.748、ym 的 2.044e-02 等(audit/verify_audit.c 的 L2 层逐条核声明值是否 ≥ 实跑值且真在输出里出现,make audit 可复跑) |
一句话:A 层回答「为什么是这个整数」——把盒子列出来、把解数数出来;B 层回答「为什么是这个式子」——参数扫一遍全零;C 层不说被迫,只说读数。 换成别的数字:A 层解数变 0,B 层残差非零,C 层根本数不出来 —— 所以三层都守得住「为什么不换个数字」。
A 层的边界(2026-10-27 审计后补,逐条可核):A 层只回答「在给定盒子里解唯一」,不回答「盒子为什么是这个盒子」。
bsd的重数向量一例:盒子的三条整数(15线 /Σλ=6/∏λ=−6¹¹)来自onestep领域的 15 线谱容器; 论文 §5bis.1 已逐字写明「matrix construction ofSfrom the mother lawW_6is not given in this note」。 故该条的准确层级是条件性 A 层(盒子给定后解唯一,判别力1−1/15504=0.99994), 不是「无条件被迫」。audit/verify_audit.c的 L4 层把这条前提标为断言层(候选集为空 ⟹ 判别力0)。ym的min|3q²−p²|=1:q=1,p=2即取到1,而0不可能由√3∉ℚ给出 ⟹ 这是初等证明(B 层), 不是「盒内解唯一」(A 层);本仓代码里也没有q≤200的枚举。README 旧文案把它列在 A 层,已改判。
共用容器(一处必须知道的重复):bsd §5bis、hl §5bis.1、rh P10 用的是同一个 15 线谱容器
(15 线 / Σλ=6 / ∏λ=−6¹¹ / (|2|,|3|,|6|)=(4,4,7)),三处各自把它计成自己的判据 ⟹ 「198 条」里有
2 处是同一对象的第二、第三次计数。每篇内部的恒等式不受影响,但跨篇不可当三条独立证据用;
audit/verify_audit.c 的 L4 层专门判这一条(重复处必须如实声明 + 前提必须标断言层)。
English
Three tiers, each checkable on the spot. Only one thing can be caught out: calling a B/C-tier number an A-tier one.
| Tier | Meaning | Machine test | Instances here (box → solutions) |
|---|---|---|---|
| A, enumeratively unique | the number is a solution of an integer system, and within the stated range the solution set has cardinality 1 | enumerate the box completely and count; more than one solution and it is not forced | bsd multiplicity vector (4,3,2,2,2,2): 15504 tuples → 1 (trace only 232, product only 100, both carry weight); bsd's (c₂,|Ш|)=(2,1): 48 → 1; ym's min|3q²−p²| (q≤200) = 1, while 0 is impossible (3 is not a square), so the gap is forced to 1 |
| B, identity | the number is given by an identity that holds for all parameters | sweep the parameter, residual 0 at every point |
hodge's Σ_p C(g,p)² = C(2g,g) (g=1..4: 2/6/20/70); pvsnp's crossing of n! over n^{cd} (cd=1..21 → 3…29); ns's 2s>n+2 ⟹ s≥3; rh's |χ|=1 ⟺ σ=1/2 |
| C, reading | the number is measured and carries no forcing | every number must state its source (this repo's run / archived probe / external anchor) | hl's three ρ values, ns's 0.748, ym's 2.044e-02, and so on (audit/verify_audit.c, layer L2, checks each declared value against the actual run; make audit re-runs it) |
One line: tier A answers "why this integer" — list the box, count the solutions; tier B answers "why this formula" — sweep the parameter and get zero everywhere; tier C does not claim forcing, only a reading. Change the number and tier A's count drops to 0, tier B's residual turns non-zero, and tier C's number simply cannot be produced — so all three tiers survive "why not some other number".
中文
| 件 | 内容 | 出现处 |
|---|---|---|
| 容器 | 有限格,判据只读特征多项式系数(相似变换不变量) | 七篇全部 |
| 对称 | 对合 σ(反读)、τ(时间反演)、w=(−1)^k |
BSD / HL / RH / YM |
| 闭环 | R(F(x))=x 的双向复算,残差 res 必须跑出来而非声明 |
七篇全部 |
三件事同构但分别陈述,不合并成一条跨篇统一定理 —— 三件各自的证据分散在七支验证器里, 接口留白处即已知缺口(见第五节)。
English
| piece | content | where it appears |
|---|---|---|
| container | a finite grid; criteria read only characteristic-polynomial coefficients (similarity invariants) | all seven |
| symmetry | involution σ (reverse reading), τ (time reversal), w=(−1)^k |
BSD / HL / RH / YM |
| closure | two-way recomputation R(F(x))=x; the residual res must be computed, never declared |
all seven |
The three pieces are isomorphic yet stated separately; each is witnessed inside the seven verifiers, and the interface is deliberately left open rather than closed by a cross-paper theorem.
| 文件 | 内容 | 行数 |
|---|---|---|
audit/INDEX_AUDIT.md |
审计索引(按主题十二组,逐节定位) | 145 |
AUDIT_20261027_非被迫结构.md |
审计正文(§1–§147) | 7745 |
audit/STACKS.md |
六条投影链:母 → 子 → 孙 → 测值 | 131 |
audit/TWO_LEDGERS.md |
两本账:船票账(旧体系认)· 闭环账(两投影合) | 150 |
EXTERNAL_MATCH.md |
三处对外对上:值 · 阶 · 常数 | 103 |
matrix/forcedness_ledger.csv |
198 行逐条台账(9 列) | 198 |
cx/probes/onestep/DEPRECATED.md |
未签到探针处置表(弃用 6 / 可修未修 2 / 值的问题 2) | 56 |
读法:先 INDEX_AUDIT.md 定位节号 → 再读正文那一节 → 数值一律回 forcedness_ledger.csv 与探针实跑核。
make run # 七支:编 + 跑(应全 rc=0)
gcc -Wall -Wextra -O2 -D_GNU_SOURCE -o t <paper>/code/verify_<x>.c -lm && ./t中文:判据输出约定 —— 每条判据打一行 [PASS]/[FAIL],每段结束打 RESULT PASS=n FAIL=m,
整个文件末尾打一行总帐 RESULT PASS=<总数> FAIL=0 (all criteria in this file) 加 EXIT=0;
进程退出码 = 是否存在失败(有失败即非 0)。每支验证器自带若干段,各段小计相加即该支总数。
另:[sanity] 类判项是恒等式自检,只报读数、不计入判据数。
English: output contract — one [PASS]/[FAIL] line per criterion, RESULT PASS=n FAIL=m per
section, and a closing aggregate RESULT PASS=<total> FAIL=0 (all criteria in this file) plus
EXIT=0; the process exit code is non-zero if any criterion fails. Each verifier ships several
sections and the section subtotals add up to the file total. Items tagged [sanity] are identity
checks: they print a reading and are not counted.
矩阵读数随仓 / matrix readings ship here:两张表都是 containers(action=matrix, …) 的实跑产物 ——
matrix/onestep_layer_gap.csv(51 层 / 预算 448 / 已发 51 / 缺口 [],逐层 D1=0 · S2=1 · CLOSED=1;
cross15 那层 9/9,md5=97af70b8c4d5ed4a418720376a273dd6)与 matrix/abacus_layer_gap.csv
(13 层 / 预算 106 / 已发 13 / 缺口 [])。论文里引用的「51 层 / 预算 448」「13 层 / 预算 106」就是这两张表的行;
矩阵工具本体在归档工作区(D:\dpgo\),本仓只带这两张表与每篇那一支 code/verify_*.c。
Matrix readings ship here: both tables are artifacts of real containers(action=matrix, …) runs —
matrix/onestep_layer_gap.csv (51 layers / cell budget 448 / 51 cells emitted / gaps [], every layer
D1=0 · S2=1 · CLOSED=1; the cross15 layer is 9/9 with md5=97af70b8c4d5ed4a418720376a273dd6) and
matrix/abacus_layer_gap.csv (13 layers / cell budget 106 / 13 emitted / gaps []). The "51 layers / budget
448" and "13 layers / budget 106" quoted in the papers are the rows of these two tables. The matrix tool
itself lives in the archive workspace (D:\dpgo\); this repository carries the two tables plus one
code/verify_*.c per paper.
中文
- 承重引理分两类,写在纸上:五篇的承重引理在容器内被迫 —— HL(
HL ⟺ ΣΔ→log(e^{2γ}/4),两常数互倒、乘积1.000000000000000)· Hodge(圆投影恒等式Φ_M整除 + 阶梯比嵌套N_p=C(g,p)²)· NS(tr S=Σλ=0+ 三分恒等P=λ₂‖ω‖²+Q₊−Q₋)· RH(Σm_λ=7的偏离无处安放)· YM(λ_l=l(l+2)、λ₀单重 ⟹Δ是相减);两篇的承重引理是另一条猜想,因此正文写成条件式 —— BSD(引理 4.1:每对在s=1恰吞一个(s−1))与 P≠NP(引理 4.1:per无多项式大小算术电路 = ValiantVP≠VNP)。条件式不是让步:它把「哪一步是别人的猜想」写清楚,其余每一步都由恒等式承担。 - 中英两版是同一份说明的两种语言,逐节对应:节号与子节一一相同(§0′ / §0″ / §1–§11、§5bis / §5ter / §5quater、§7bis、附录 A/B/C、附章),公式、表格、读数也一一相同;判据数、段小计、
RESULT总帐、[sanity]口径两边一致。 - 判据都落在有限格上,阈值由口径给定(不事后调参):
p ≤ 97、n ≤ 30、cd ≤ 30、g ≤ 4、M ∈ {2,3,4,6,8,12}、k ≤ 12。超出窗口的那一格是换容器,不是判据失效 —— 窗口外要重新给容器(例如cd ≥ 22的交叉点在n > 30之外,pn_form就把它们记为边界而不断言)。 - 程序的输出是判据,不是猜想的真值:每个判据在整数/有理数上给 0/1,文件末行
RESULT PASS=n FAIL=0;读法是「这一层的恒等式零违例」,见第二节表。 - 孪生是 HL 的一章,不是第八篇。
English
- Load-bearing lemmas come in two kinds, and both are on the page. Five papers carry a lemma that
is forced inside the container — HL (
HL ⟺ ΣΔ→log(e^{2γ}/4), the two constants reciprocal, product1.000000000000000), Hodge (the circle-projection identity,Φ_Mdivides, plus the nested ladder ratioN_p=C(g,p)²), NS (tr S=Σλ=0together withP=λ₂‖ω‖²+Q₊−Q₋), RH (the deviationΣm_λ=7has nowhere to sit), YM (λ_l=l(l+2),λ₀simple⟹ Δis a subtraction). Two papers rest on a lemma that is another conjecture, so their theorem is stated conditionally — BSD (Lemma 4.1: each pair eats exactly one(s−1)ats=1) and P≠NP (Lemma 4.1:perhas no polynomial-size arithmetic circuit, i.e. Valiant'sVP≠VNP). The conditional form is not a hedge: it names whose conjecture a step is, and every other step is carried by an identity. - The two notes are one document in two languages, section by section. Section numbers and
subsections match one to one (§0′, §0″, §1–§11, §5bis / §5ter / §5quater, §7bis,
appendices A/B/C, the twin chapter), as do formulas, tables and readings. Counts, section
subtotals,
RESULTtotals and the[sanity]rule are identical on both sides. - Every criterion sits on a finite grid with a threshold fixed by the container (no post-hoc
tuning):
p ≤ 97,n ≤ 30,cd ≤ 30,g ≤ 4,M ∈ {2,3,4,6,8,12},k ≤ 12. A cell outside the window is a different container, not a failed criterion — e.g. thecd ≥ 22crossings lie beyondn = 30, sopn_formrecords them as a boundary and asserts nothing there. - The program prints criteria, not a truth value for the conjecture: every criterion resolves to
0/1 over integers or rationals, and the last line is
RESULT PASS=n FAIL=0. Read it as "the identities of this layer hold with zero violations"; see the §2 table. - The twin-prime case is a chapter of HL, not an eighth paper.
- zjaizlh8:说结构、指方向;
- DeepSeek:计算与校验;
- 腾讯元宝 hy 模型(专业版):校验 —— 免费使用,特此致谢。
zjaizlh8 set the structure and direction; DeepSeek ran the computations and checks; Tencent Yuanbao hy model (professional edition) provided independent verification.
Author / 作者: zjaizlh8 · 2026-10-27 License / 许可: CC BY 4.0(文本与代码 / text and code);猜想表述本身为公有数学问题 / the conjectures themselves are public mathematical statements.