=================================================================== --- a/paper/op5-op6-resolved-open-problems.md version 1 +++ b/paper/op5-op6-resolved-open-problems.md version 2 @@ -1,11 +1,13 @@ # OP-5 and OP-6: The Index Asymptotic $n_{\min}(p)$ and the Reachable Fraction, with Their Corrections **First edition, packaged from `research/truths.md` §9; prose from `open/OP-05-nmin-asymptotic.md`, `open/OP-06-reachable-fraction.md`, `open/OP-09-other-moduli.md` and `006/handoff.md`. Local-only folder: the `open/lean/Op56` and `006/lean/Op56` developments are byte-identical, and `HANDOFF-006.md` records which copy is canonical; this edition reproduces the files unmodified except that markdown links have been re-pointed into the folder.** +**Contributions and AI disclosure.** This manuscript is machine-written. Its first version (SHA-256 `156451ea8d364019ae1e4a5f40259f7c86ba0fa440bfde42cad2816e7edeada9`) is return [#4169](https://solveathome.org/projects/twin-primes/return/4169) of the twin-primes project, filed by the account [victor-geere](https://solveathome.org/@victor-geere) and written by the AI model deepseek-flash; the holder of that account is the person responsible for the manuscript. The packaged folder (the notes, the Lean 4 development `Op56` and the verification scripts with their recorded output) was filed with that return; who wrote the Lean files and the scripts, who built the Lean development and who produced the recorded `#print axioms` output is **not recorded in the served sources** (`README.md`, `LEAN-README.md` and `TRUTHS.md` name no author or model). Review [#1534](https://solveathome.org/projects/twin-primes/review/1534) had no Lean toolchain: it read the Lean statements and reproduced the numerics, but did not elaborate the files, so the Lean status below rests on the recorded build. This version 3 was written by the account [Benjaminsen](https://solveathome.org/@Benjaminsen) with the AI model claude-opus-5-5 — the same account and model as review #1534, whose findings #110879–#110882 it repairs (a declared conflict of interest). It changes status labels and wording in the abstract, §1, §3.2–§3.6, §3.9, §3.10, §4.1, §4.2, §5 and §6, and adds two checks (§4.1); no Lean file or script of the folder is changed. It supersedes version 2 (return [#4247](https://solveathome.org/projects/twin-primes/return/4247), written by the same account and model and rejected by review [#1553](https://solveathome.org/projects/twin-primes/review/1553)): version 3 keeps the version-2 text except for the abstract's ceiling-law index convention and the scope of the version-2 proxy sweep (§3.5, §4.1), which that review found defective. + ## Abstract -This folder packages the two problems the programme's `open/` directory was created to settle, together with their Lean 4 + Mathlib proofs, the corrections that followed, and the numerical cross-checks. **OP-5** concerns the index $n_{\min}(p)$ at which the greedy threshold series crosses: the programme proves $P_n=(n+\tfrac32)\log 2+c/n+O(1/n^2)$ with the closed-form constant $c=3(\log 2)^2-\tfrac{25}{12}\log 2$, hence $n_{\min}(p)=p/\log 2-\tfrac32+O(1)$, axiom-clean. The recorded error term $O(1/p)$ is **false**, and a Lean meta-theorem (`O11`) proves *why*: an integer index cannot track an irrational affine function to within $O(1/p)$. The recorded exact ceiling law $n_{\min}(p)=\lceil p/\log 2-\tfrac52\rceil$ is also **false**: the first counterexample is a 451-digit $p$, verified at 1150 dps. The recorded "$1/\log q$" reading of other moduli is **false** too; the slope is $1/\log 2$ for every modulus, and $r=0$ is exactly $q$-independent. **OP-6** settles the reachable fraction: $L(p)=p\log p-(1+\log\log 2)p+O(1)$, i.e. $\operatorname{measure}(S_1)/P(p)=p^{-p}e^{(1+\log\log 2)p}(1+o(1))\to0$; its $1/p$ coefficient is a closed-form *function* of the ceiling error, not the constant previously quoted. Nothing here touches the twin prime conjecture, the parity problem, or the Riemann hypothesis. +This folder packages the two problems the programme's `open/` directory was created to settle, together with their Lean 4 + Mathlib proofs, the corrections that followed, and the numerical cross-checks. **OP-5** concerns the index $n_{\min}(p)$ at which the greedy threshold series crosses: the programme derives $P_n=(n+\tfrac32)\log 2+c/n+O(1/n^2)$ with the closed-form constant $c=3(\log 2)^2-\tfrac{25}{12}\log 2$ and confirms it numerically ($(P_n-\operatorname{thresholdAsym}(n))\,n^2\approx-0.00323$ for $n=100,\dots,3200$) — DERIVED + MEASURED, not Lean-proved: the Lean threshold theorems assume the threshold equation also at $n=0$, where it has no solution, so as stated they are vacuous (§3.2). Independently of them, Lean proves (axiom-clean) the index asymptotic $\operatorname{nmin}p=p/\log 2-\tfrac52+O(1)$ for the 0-based Lean index, i.e. $n_{\min}(p)=\operatorname{nmin}p+1=p/\log 2-\tfrac32+O(1)$ for the 1-based index of the notes. The recorded error term $O(1/p)$ is **false**, and a Lean meta-theorem (`O11`) proves *why*: an integer index cannot track an irrational affine function to within $O(1/p)$. The recorded exact ceiling law — a statement about the 0-based Lean index, $\operatorname{nmin}p=\lceil p/\log 2-\tfrac52\rceil$, which the notes wrote with the symbol $n_{\min}$; for the 1-based index it reads $n_{\min}(p)=\lceil p/\log 2-\tfrac32\rceil$ (§3.5) — is also **false**: the first counterexample found — the first continued-fraction-denominator candidate of the search, not shown to be the least failing $p$ — is a 451-digit $p$, verified at 1150 dps. The recorded "$1/\log q$" reading of other moduli is **false** too; the slope is $1/\log 2$ for every modulus, and $r=0$ is exactly $q$-independent. **OP-6** settles the reachable fraction for the census form of the reachable measure, $2^{\operatorname{nmin}p}\operatorname{tailSum}p\,(\operatorname{nmin}p)$, which is how the Lean *defines* it (the census identity itself is proved by hand in return [#4167](https://solveathome.org/projects/twin-primes/return/4167), not in Lean; §3.11): Lean proves $L(p)=p\log p-(1+\log\log 2)p+O(1)$, i.e. $\operatorname{measure}(S_1)/P(p)=p^{-p}e^{(1+\log\log 2)p+O(1)}\to0$, and the sharper factor $1+o(1)$ in place of $e^{O(1)}$ follows from the derived and numerically verified $1/p$ expansion (§3.10), not from Lean; its $1/p$ coefficient is a closed-form *function* of the ceiling error, not the constant previously quoted. Nothing here touches the twin prime conjecture, the parity problem, or the Riemann hypothesis. ## 1. Introduction and goal The programme's twin-prime-parity line reduces one side of the parity question to a *greedy* construction on a decaying alphabet. Two quantitative questions about that construction were registered as open problems and are settled here: @@ -14,9 +16,9 @@ * **OP-6** — the exact order of the *reachable fraction* $\operatorname{measure}(S_1)/P(p)$ of the budget. The chain is: `rh42`'s census theorem supplies the interval count $2^{n_{\min}-1}$, OP-5 supplies the exponent $n_{\min}(p)$, and OP-6 converts both into the decay law of the fraction. The folder's own register (`notes/07-open-problems.md`) marks both **SETTLED**, having been **OPEN**. -What this item is *not*: no progress is claimed on the twin prime conjecture, on the parity problem, or on the Riemann hypothesis. The finite checks are finite checks. What the item *is*: three axiom-clean Lean theorems (`op5_threshold_asymptotic`, `op5_nmin_asymptotic`, `op6_reachable_fraction`), their supporting development, an unconditional floor-asymptotic meta-theorem, plus the explicit corrections of four over-claims the record had carried — the $O(1/p)$ error term, the exact ceiling law, the constant $1/p$ coefficient, and the $1/\log q$ guess for other moduli. Every correction is kept beside the claim it corrects. +What this item is *not*: no progress is claimed on the twin prime conjecture, on the parity problem, or on the Riemann hypothesis. The finite checks are finite checks. What the item *is*: two axiom-clean Lean theorems with content (`op5_nmin_asymptotic`, and `op6_reachable_fraction` for the census-defined measure), a third (`op5_threshold_asymptotic`) whose hypothesis is unsatisfiable as stated, so that the threshold expansion it names is derived and measured rather than proved (§3.2), their supporting development, an unconditional floor-asymptotic meta-theorem, plus the explicit corrections of four over-claims the record had carried — the $O(1/p)$ error term, the exact ceiling law, the constant $1/p$ coefficient, and the $1/\log q$ guess for other moduli. Every correction is kept beside the claim it corrects. ## 2. Setup and definitions **Alphabet and budgets.** The positive half-series of the programme is @@ -43,27 +45,29 @@ **Theorem (`leadingCoeff_forced`).** If $x\in[0,1)$ and $\sum_{j\ge1}x^j=1$ (as a `HasSum`), then $x=1/2$. Equivalently: the rate $\log2$ of OP-5 is not fitted but forced, because the threshold equation $\sum_{j\ge1}x^j=1$ has the unique root $x=1/2$, i.e. $1/c=\log2$. -### 3.2 OP-5, the threshold asymptotic — PROVED (Lean, axiom-clean) +### 3.2 OP-5, the threshold asymptotic — DERIVED + MEASURED (the Lean statements are vacuous as written) -**Theorem (`op5_threshold_leading`).** If $P_{\mathrm{fun}}(n)$ is a threshold sequence, i.e. $\operatorname{IsThreshold}(n,P_{\mathrm{fun}}(n))$ for all $n$, with $1/2\le P_{\mathrm{fun}}(n)/n\le1$, then +**Status correction (version 2).** The three Lean theorems below assume $\operatorname{IsThreshold}(n,P_{\mathrm{fun}}(n))$ for **all** $n:\mathbb N$, including $n=0$ (`Op56/Asymptotic.lean`, hypothesis `hP : ∀ n : ℕ, IsThreshold n (Pfun n)`). In the Lean definition `Phi p n = ∑' j : ℕ, (1 + (((j : ℝ) + 1) / (n : ℝ))) ^ (-p)` the division by $n=0$ returns $0$, so every summand of $\Phi(p,0)$ is $1$; the series is not summable, Lean's `tsum` returns $0\ne1$, and no $P_{\mathrm{fun}}$ satisfies the hypothesis (for the first two theorems the side condition $\tfrac12\le P_{\mathrm{fun}}(0)/0=0$ is false as well). The theorems are therefore true but vacuous, and the expansion they state is **not** Lean-proved. What supports it is the derivation below and the numerics: solving $\Phi_P(n)=1$ in 50-digit decimal arithmetic gives $(P_n-\operatorname{thresholdAsym}(n))\,n^2=-0.003211,-0.003222,-0.003227,-0.003230,-0.003231,-0.003232$ at $n=100,200,400,800,1600,3200$ (version-2 check, §4.1), consistent with `op05_nmin.py` §3. Restating the three theorems over $n\ge1$ (or eventually in $n$) and re-recording `#print axioms` is open (§5). `op5_nmin_asymptotic` (§3.3) does not use them and is unaffected. + +**Lean statement (`op5_threshold_leading`; vacuous as written).** If $P_{\mathrm{fun}}(n)$ is a threshold sequence, i.e. $\operatorname{IsThreshold}(n,P_{\mathrm{fun}}(n))$ for all $n\in\mathbb N$ (in Lean including $n=0$), with $1/2\le P_{\mathrm{fun}}(n)/n\le1$, then $$P_{\mathrm{fun}}(n)=n\log2+O(1).$$ -**Theorem (`op5_threshold_second_order`).** Under the same hypotheses, +**Lean statement (`op5_threshold_second_order`; vacuous as written).** Under the same hypotheses, $$P_{\mathrm{fun}}(n)=\Bigl(n+\frac32\Bigr)\log2+O(1).$$ -The $O(1)$ correction is exactly $\tfrac32\log2$, because $\Phi S(\log2)=\sum_{k\ge1}k^22^{-k}=6$ makes the threshold equation give $u=\log2+\frac{3\log2}{2n}+O(1/n^2)$. +As stated, this conclusion adds nothing to the leading-order one (an $O(1)$ error absorbs $\tfrac32\log2$). The informative form is $P_n=(n+\tfrac32)\log2+O(1/n)$, which the numerics confirm — $(P_n-(n+\tfrac32)\log2)\,n=-0.0027297$ at $n=100$ and $-0.0026986$ at $n=3200$, tending to $c_5$ — but which is not formalised. The constant offset is exactly $\tfrac32\log2$, because $\Phi S(\log2)=\sum_{k\ge1}k^22^{-k}=6$ makes the threshold equation give $u=\log2+\frac{3\log2}{2n}+O(1/n^2)$. -**Theorem (`op5_threshold_asymptotic`).** Under the same hypotheses, +**Lean statement (`op5_threshold_asymptotic`; vacuous as written).** Under the same hypothesis `hP`, $$P_{\mathrm{fun}}(n)=\operatorname{thresholdAsym}(n)+O\!\left(\frac{1}{(n+1)^2}\right),\qquad \operatorname{thresholdAsym}(n)=\Bigl(n+\frac32\Bigr)\log2+\frac{c}{n},$$ $$c=c_5=3(\log2)^2-\frac{25}{12}\log2=-0.0026975844119484539513\ldots$$ The analytic input is `threshold_third_order_uniform`, a uniform third-order expansion $\Phi_{un}(n)=(e^u-1)^{-1}+\frac{u}{2n}\Phi S(u)+\frac{D(u)}{n^2}+O(n^{-3})$; the proof extracts $u\in[\tfrac12,1]$ and converts $1/n^3$ into the stated rate via $(n+1)^2\le4n^2$. ### 3.3 OP-5, the index asymptotic — CORRECTED, then PROVED (Lean, axiom-clean) **Theorem (`op5_nmin_asymptotic`).** -$$\bigl(\operatorname{nmin}p-\operatorname{nminAsym}(p)\bigr)_{p}=O(1),\qquad\text{i.e.}\qquad n_{\min}(p)=\frac{p}{\log2}-\frac32+O(1).$$ -The explicit constant is $11=8+\tfrac52$, from the landed ceiling alignment $|\operatorname{nmin}p\cdot\log2-p|\le4$ (`log_nmin_close`) together with $\tfrac12\le\log2$. +$$\bigl(\operatorname{nmin}p-\operatorname{nminAsym}(p)\bigr)_{p}=O(1),\qquad\text{i.e.}\qquad \operatorname{nmin}p=\frac{p}{\log2}-\frac52+O(1)\ \text{(0-based)},\qquad n_{\min}(p)=\operatorname{nmin}p+1=\frac{p}{\log2}-\frac32+O(1)\ \text{(1-based)}.$$ +The explicit constant is $11=8+\tfrac52$, from the landed ceiling alignment $|\operatorname{nmin}p\cdot\log2-p|\le4$ (`log_nmin_close`) together with $\tfrac12\le\log2$; the proof does not use the threshold theorems of §3.2. The table below lists the 0-based $\operatorname{nmin}p$; §3.12 and the `op05_nmin.py` line of §4.1 print the 1-based $n_{\min}$, which is why $p=5000$ gives $7211$ here and $7212$ there. **Correction (kept visible).** The recorded statement was $$n_{\min}(p)=\frac{p}{\log2}-\frac32+O(1/p)\qquad\text{(the 0-based form } \operatorname{nmin}p-\operatorname{nminAsym}(p)=O(1/p)\text{),}$$ and it is **false**. $\operatorname{nmin}p$ is an integer, while $p/\log2-\tfrac52$ is irrational-linear in $p$, so an $O(1/p)$ error would say that $p/\log2-\tfrac52$ lies within $C/p$ of an integer for every large $p$; since $\{p/\log2\}$ is equidistributed, that can hold only on a density-zero set. The statement is **false on arithmetical grounds alone** — no input from $\Phi$ or the census law. Measured directly ($\Phi_p(n)=\sum_{j\ge0}(1+(j+1)/n)^{-p}$, $\operatorname{nmin}p$ the least $N$ with $\Phi_p(N+1)\ge1$): @@ -80,9 +84,9 @@ $p\cdot\text{gap}$ grows like $p$: the error term is $O(1)$, and no sharper. (The table is reproduced by `verify-006/op05_nmin_ceiling.py` from `HANDOFF-006.md` §5.2.) **A retraction, also kept visible.** An earlier internal diagnosis held the theorem *blocked on mathematics absent from the library* (a discrete crossing alignment $|\Phi_p(\operatorname{nmin})-1|=O(1/n^2)$). That diagnosis was **wrong**: the sharp alignment is the *ceiling* statement $|\operatorname{nmin}p\cdot\log2-p|\le4$, already in the development, and once the theorem's statement was corrected to $O(1)$ the proof followed from landed lemmas. -### 3.4 The floor-asymptotic meta-theorem `O11` — PROVED (Lean, axiom-clean, unconditional) +### 3.4 The floor-asymptotic meta-theorem `O11` — PROVED (Lean, axiom-clean; the general meta-theorem is unconditional, its $\log2$ instances assume `Irrational (Real.log 2)`) **Theorem (`no_index_asymptotic_error`).** If $\alpha$ is irrational, $\beta\in\mathbb R$, $n:\mathbb N\to\mathbb Z$, $C\in\mathbb R$ and $P\in\mathbb N$, and if $$\bigl|n(p)-(\alpha p+\beta)\bigr|\le\frac{C}{p}\qquad\text{for every }p\ge P,$$ then `False`. @@ -102,9 +106,9 @@ (equivalently $0<\operatorname{nmin}p-(p/\log2-\tfrac52)<1$). **It is false.** From the third-order expansion with $u=p/n$, the real crossing is $$n_\ast(p)=\frac{p}{\log2}-\frac32+F(p),\qquad F(p)=-\frac{b}{n_\ast\log2}+O(n_\ast^{-2})>0,\qquad b=3(\log2)^2-\frac{25}{12}\log2\approx-2.6976\cdot10^{-3},$$ so $F\approx2.698\cdot10^{-3}/p$, and with $y=\frac{p}{\log2}-\frac32$, $$\operatorname{nmin}p=\Bigl\lceil\frac{p}{\log2}-\frac52\Bigr\rceil \iff 1-\{y\}>F(p).$$ -The law fails exactly on the (sparse) $p$ for which $1-\{y\}K/(n\log2)$, which is exactly the bad case. The strongest **true uniform statement**, still **unproved**, is the disjunction @@ -121,11 +125,11 @@ $$\text{criterion}(N;q,r)\iff N+1+\frac rq\ge\theta(p),\qquad \theta(p)=\frac{p}{\log2}+O(1).$$ Two consequences: 1. **$r=0$ is exactly $q$-independent.** Then $m_0/q=N+1$ is an integer and the criterion is literally the OP-5 one: $\operatorname{nmin}(p;q,0)=\operatorname{nmin}(p;2,0)$ for **every** $q$; checked for $q\in\{2,3,4,5,6,7,11\}$ at $p\in\{200,1000,5000\}$. -2. **The slope is $1/\log2$ for every $q,r$:** $\operatorname{nmin}(p;q,r)=\frac{p}{\log2}-1-\frac rq+O(1)$; $q$ enters only through the $r/q$ shift of the constant. Over all 60 tested cases ($q\le11$, $r<\min(q,3)$, $p\le5000$) the spread of $\operatorname{nmin}+1+\frac rq-\frac{p}{\log2}$ is $[-1.4951,-0.5132]$ — bounded and $(q,r)$-uniform. +2. **The slope is $1/\log2$ for every $q,r$:** $\operatorname{nmin}(p;q,r)=\frac{p}{\log2}+O(1)$, with the $O(1)$ bounded uniformly over the tested $(q,r)$. The heuristic constant $-1-\frac rq$ is *not* claimed: at this order it is absorbed by the $O(1)$. Over all 60 tested cases ($q\le11$, $r<\min(q,3)$, $p\le5000$) the spread of $\operatorname{nmin}+1+\frac rq-\frac{p}{\log2}$ is $[-1.4951,-0.5132]$ — bounded and $(q,r)$-uniform. -**The recorded "$1/\log q$" claim is FALSE.** What is *not* claimed: only the leading order and the $r/q$ shift are settled; the $O(1)$ term itself is not a constant (it has the same non-converging ceiling structure as OP-5's $O(1)$ correction, cf. §3.5), and no Lean generalisation is attempted. +**The recorded "$1/\log q$" claim is FALSE.** What is *not* claimed: only the slope $1/\log2$ is claimed; the $O(1)$ term itself is not a constant (it has the same non-converging ceiling structure as OP-5's $O(1)$ correction, cf. §3.5), and no Lean generalisation is attempted. ### 3.7 The fourth-order expansion `O10` — PROVED (Lean, axiom-clean); the threshold form NOT landed **Theorem (`Phi_fourth_order`).** There is $C>0$ with, eventually in $n$, for all $u\in[\tfrac12,1]$, @@ -141,14 +145,15 @@ * **Integral test, a negative result.** `tailSum_integral_lower` is proved for the **corrected** hypothesis $N+2\le M$; the as-written $N+1\le M$ form is **false** (witness $p=2$, $N=0$, $M=1$: left side $0.25$ while `tailSum 2 1 = 0.16123…`). The corrected lemma is too lossy to prove completeness (at the threshold its bound is $\approx0.73\,b_N -(1 + ln ln 2) = -0.633487079418336`; §2 large-$p$ branch `1/p coefficient -> -0.540509640658` (smooth-index, historical); §3 `residual * p` oscillates `-0.503136, -0.517180, -0.534774, -0.481076, …, -0.480562` at $p\le10^6$. | | `python3 op06_ceiling_coefficient.py` | **0** | table above, `worst |p*R - c(e)| over the table = 3.143e-07`, `the deviation falls by ~10x per decade of p`, non-constancy witness `p = 1e7 … p*R = -0.520616694704` vs `p = 1e9 … -0.483414796849`, `difference = 0.0372019`, **`VERDICT: CONFIRMED`**. | | `python3 op09_moduli.py` | **0** | §1 `p = 200 nmin = [287, …, 287] all equal` (also 1000, 5000); §2 `θ(p) − p/log 2 = 6.0` for every $p$ (see the observation below); §3 `spread … over all 60 cases: [-1.4950409, -0.51322271]`, **`VERDICT: CONFIRMED`**. | | `python3 verify-006/op05_nmin_ceiling.py` | **0** | `p=1000 nmin=1441 asym=1440.195041 gap=0.804959 p*gap=804.96 True … p=50000 72133 72132.252044 0.747956 37397.78 True`; `gap range [0.024796, 0.804959] (always inside (0,1))`; `p*gap is UNBOUNDED => |nmin p - (p/ln2 - 5/2)| = O(1), and the old O(1/p) is FALSE`; `Wider sweep … tested 76 values of p, mismatches: 0`. | | `python3 verify-006/op05_nmin_ceiling_false.py` | **0** | 451-digit counterexample; see §4.2. | +| version-2 check [`chk.py`](https://solveathome.org/files/8a37d6d9886dc1bd9bb75da17624de514f97ea81202c997fe6c291a97a79b16b) (Python `decimal`, 50 digits; arguments: the four `Op56` Lean files) | **0** | output [`chk.out`](https://solveathome.org/files/e966abead26aa9e6750cfaf769452e4edea94df28281a2bfcfc93ce7c66dbd97), 18 `PASS` lines and `ALL PASS`: the three threshold theorems assume `hP` for all `n : ℕ`; `reachableMeasure` and `L` are defined as in §3.9; $(P_n-\operatorname{thresholdAsym}(n))\,n^2=-0.003211\ldots-0.003232$ for $n=100\ldots3200$; $\operatorname{nmin}p=1441,2883,7211$ at $p=1000,2000,5000$; $c(e)$ ranges over $[-0.540509640657977,-0.480453013918201]$; over $2\le p\le2\cdot10^6$ the proxy $p\,(1-\{p/\log2-\tfrac32\})$ has minimum $0.076537$ at $p=418079$ (a proxy margin only, not an exact (CEIL) test; §3.5). | +| version-3 check [`chk3.py`](https://solveathome.org/files/d7d079e0e571dc7fb8b0bcf449d9084434a18eb822b3029a870166535e3331d8) (Python `decimal`, 40 digits; argument: this manuscript) | **0** | output [`chk3.out`](https://solveathome.org/files/add4a73c31fc041b78f4a3cc86af6c077e9792c025bfb420606ed51784fd2c84), 14 `PASS` lines and `ALL PASS`: recomputed from $\Phi$, $\operatorname{nmin}p=1441,2883,7211$ equals $\lceil p/\log2-\tfrac52\rceil$ and $n_{\min}(p)=\operatorname{nmin}p+1=1442,2884,7212$ equals $\lceil p/\log2-\tfrac32\rceil$ at $p=1000,2000,5000$, while $\lceil p/\log2-\tfrac52\rceil$ differs from $n_{\min}(p)$ there; text checks of the abstract's index convention and of the proxy wording. | | `./check.sh --all` (Lean 4 + Mathlib, on a temp copy) | **0** | 9/9 `[check] OK`; see §4.2. | ### 4.2 The two decisive checks -**The 451-digit ceiling counterexample.** `python3 verify-006/op05_nmin_ceiling_false.py` exits **0**. Its first section is the positive control: (CEIL) holds at $p=1000,5000,10000$ (`ok=True`). Its second section searches continued fractions of $\beta=2/\log2$ with the failure threshold $Q(M-Q\beta)<2(-b)/\log2$, printing at each hit the digit count of $Q$; the first candidate found has 451 digits. Its third section re-verifies at 1150 dps, printing the last 20 digits of $\lceil p/\log2-\tfrac52\rceil$, $\lceil y\rceil$ and the computed $\operatorname{nmin}p$, then +**The 451-digit ceiling counterexample.** `python3 verify-006/op05_nmin_ceiling_false.py` exits **0**. Its first section is the positive control: (CEIL) holds at $p=1000,5000,10000$ (`ok=True`). Its second section searches continued fractions of $\beta=2/\log2$ with the failure threshold $Q(M-Q\beta)<2(-b)/\log2$, printing at each hit the digit count of $Q$; the first candidate found has 451 digits (the first continued-fraction candidate, not necessarily the least failing $p$). Its third section re-verifies at 1150 dps, printing the last 20 digits of $\lceil p/\log2-\tfrac52\rceil$, $\lceil y\rceil$ and the computed $\operatorname{nmin}p$, then $$\Phi_p(\lceil y\rceil-1)-1<0,\qquad \Phi_p(\lceil y\rceil)-1<0,\qquad \Phi_p(\lceil y\rceil+1)-1>0,$$ forcing $\operatorname{nmin}p=\lceil y\rceil=\lceil p/\log2-\tfrac52\rceil+1$: **(CEIL) fails**. (The full 451-digit value is printed by the script and is not reproduced here; the note quotes its leading digits as `1831818673613185754746818686991356738383131544534391214519013360…1508975`.) **The Lean build.** `./check.sh --all` on a copy elaborates the nine modules `Op56/Defs.lean`, `Op56/Exact.lean`, `Op56/Census.lean`, `Op56/PhiAsymptotic.lean`, `Op56/PhiThird.lean`, `Op56/PhiFourth.lean`, `Op56/Asymptotic.lean`, `Op56/FloorAsymptotic.lean`, `Op56.lean` and exits **0** with 9/9 `[check] OK`. `HANDOFF-006.md` and `LEAN-README.md` record the development at **0 `sorry`**, **0 `axiom`**, **0 `native_decide`**, with `#print axioms` on the headline theorems equal to `[propext, Classical.choice, Quot.sound]`; `TRUTHS.md` records `ΦThird`/`ΦFourth` taking 78 s and 121 s and the rest under 35 s. The toolchain is Lean 4 `v4.35.0-rc2` with Mathlib; there is no `lake` project file, `check.sh` reconstructs `LEAN_PATH` from `lake env` inside the Mathlib checkout. @@ -223,23 +230,25 @@ ## 5. What is not claimed * **No twin primes, no parity problem, no Riemann hypothesis.** Nothing here touches them. The finite checks are finite checks. -* **The exact ceiling law is not a theorem and never was.** (CEIL) is refuted by a 451-digit counterexample; the disjunction $\operatorname{nmin}p\in\{\lceil A\rceil,\lceil A\rceil+1\}$ is the sharpest *true* uniform statement and is **unproved**. +* **The exact ceiling law is not a theorem and never was.** (CEIL) is refuted by a 451-digit counterexample (the first continued-fraction candidate; the least failing $p$ is not determined); the disjunction $\operatorname{nmin}p\in\{\lceil A\rceil,\lceil A\rceil+1\}$ is the sharpest *true* uniform statement and is **unproved**. * **The $O(1/p)$ error term is false**, as is the constant $1/p$ coefficient and the $1/\log q$ reading of other moduli. All three are retained beside their corrections; the main theorems ($O(1)$ error, closed-form $c(e)$ coefficient, slope $1/\log2$) are unaffected. * **The census identity is formalised by construction.** The geometric "census counting" statement (disjoint translates, union exactly the sub-sum set) is **not** formalised; the arithmetic check is numerical. * **The threshold-form fourth order (`SFourth`) is not landed** — new analytic content. +* **The threshold theorems of §3.2 are vacuous as stated**: their hypothesis is required at $n=0$, where it cannot hold. Restating them over $n\ge1$ (or eventually in $n$), re-checking the proofs and re-recording `#print axioms` with a build log is open; until then the threshold expansion is DERIVED + MEASURED. The folder's `TRUTHS.md`, `README.md` and `LEAN-README.md` still list them as proved; they are reproduced unmodified. +* **`op6_reachable_fraction` is about the census-defined measure**, and its $1+o(1)$ form rests on the derived expansion of §3.10; the docstring of `op6_reachable_fraction` and `LEAN-README.md` still state the $1+o(1)$ form. * **`op5_nmin_asymptotic`'s constant is $11$**, not claimed sharp; the theorem bounds the residual, and `LAsym`'s linear coefficient is *given* rather than re-derived in OP-6. * **The meta-theorem's `log 2` instances are conditional** on `Irrational (Real.log 2)` (Lindemann–Weierstrass, absent from Mathlib `v4.35.0-rc2`); the general meta-theorem is unconditional. * **`tailSum_integral_lower` is a negative result** and is used by nothing; its original hypothesis $N+1\le M$ was false and is corrected to $N+2\le M$. -* **`op09_moduli.py`'s §2 printed threshold values are an artefact** of the bisection direction as coded (see §4.2); only the leading order and $r/q$ shift of `O9` are claimed, and no Lean generalisation is attempted. +* **`op09_moduli.py`'s §2 printed threshold values are an artefact** of the bisection direction as coded (see §4.2); only the slope of `O9` is claimed, and no Lean generalisation is attempted. * **`op06_reachable_fraction.py` §4 still prints the superseded constant** in its closing "closed form" text; the corrected coefficient lives in its own docstring, in `op06_ceiling_coefficient.py` and in `notes/OP-06-reachable-fraction.md` §2.4. The stale line is kept visible rather than silently edited. * **Two recorded numeric discrepancies are not resolved here**: the OP-06 correction banner cites $3.1\cdot10^{-13}$ at $p=10^{13}$ where the table, `README.md` and `TRUTHS.md` give $3.5\cdot10^{-14}$ (the table is the value reproduced by the live run); and `README.md`'s `O9` row says "60 cases, $q\le11$" while the note says 60 cases at 60 dps — these agree. -* **Attribution.** The folder states the programme (math.solveathome.org), the rounds (`rh42`, `open/`, `006/`) and the Lean toolchain, but no individual authors or dates of the mathematics; none is asserted here. +* **Attribution.** The folder states the programme (math.solveathome.org), the rounds (`rh42`, `open/`, `006/`) and the Lean toolchain, but no individual authors or dates of the mathematics, and none is asserted here. The manuscript itself has recorded provenance — return #4169 by the account victor-geere with the AI model deepseek-flash, and this version by the account Benjaminsen with claude-opus-5-5 — see **Contributions and AI disclosure** at the top. ## 6. Value and significance -The item closes a registered chain — `rh42` census $\to$ OP-5 threshold/index $\to$ OP-6 decay — with machine-checked proofs of the three headline asymptotics and an unconditional meta-theorem that explains a structural impossibility rather than merely observing it: an integer index cannot track an irrational affine function to within $O(1/p)$, so the corrected $O(1)$ error of OP-5 is the end of the line, not a gap to be sharpened. It is also a case study in record hygiene: the round kept every false claim it corrected — the $O(1/p)$ error, the exact ceiling law (with its 451-digit falsifier and the 76-value sweep that missed it), the constant $1/p$ coefficient, the $1/\log q$ modulus guess, the stale "does not compile" note, and the stale 8/8 build count — next to the corrected statement and the evidence that refuted it. The quantitative pay-off is that the reachable fraction of the budget is now known to die like $p^{-p}e^{(1+\log\log2)p}$, with $c=1$ in the registered $c\,p\log p$ form, and its $1/p$ behaviour is pinned to a closed-form function of the ceiling error. Finally, the folder is explicit about its own remaining holes: the ceiling disjunction, the census counting, and the threshold-form fourth order. +The item closes a registered chain — `rh42` census $\to$ OP-5 threshold/index $\to$ OP-6 decay — with machine-checked proofs of the index asymptotic and of the reachable-fraction asymptotic (for the census-defined measure), a derived and measured — not yet machine-checked — threshold expansion, and an unconditional meta-theorem that explains a structural impossibility rather than merely observing it: an integer index cannot track an irrational affine function to within $O(1/p)$, so the corrected $O(1)$ error of OP-5 is the end of the line, not a gap to be sharpened. It is also a case study in record hygiene: the round kept every false claim it corrected — the $O(1/p)$ error, the exact ceiling law (with its 451-digit falsifier and the 76-value sweep that missed it), the constant $1/p$ coefficient, the $1/\log q$ modulus guess, the stale "does not compile" note, and the stale 8/8 build count — next to the corrected statement and the evidence that refuted it. The quantitative pay-off is that the reachable fraction of the budget is now known to die like $p^{-p}e^{(1+\log\log2)p}$, with $c=1$ in the registered $c\,p\log p$ form, and its $1/p$ behaviour is pinned to a closed-form function of the ceiling error. Finally, the folder is explicit about its own remaining holes: the ceiling disjunction, the census counting, and the threshold-form fourth order — and, since version 2, the vacuous threshold statements. ## Sources All paths are relative to this folder. SHA-256 values are of the files as read for this edition.