\documentclass[11pt,a4paper]{article} \usepackage[margin=1in]{geometry} \usepackage{amsmath,amssymb,amsthm} \usepackage{booktabs} \usepackage[hidelinks,breaklinks]{hyperref} \usepackage{enumitem} \theoremstyle{plain}\newtheorem{theorem}{Theorem}\newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma}\newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition}\newtheorem{example}[theorem]{Example} \theoremstyle{remark}\newtheorem{remark}[theorem]{Remark} \newcommand{\N}{\mathbb{N}}\newcommand{\Z}{\mathbb{Z}} \title{Realizer Bounds and Repetition Constraints\\ at the First Coefficient Crossing in the $3x+1$ Problem} \author{Elias De Jes\'us\\ \textit{Independent Researcher}\\ ORCID \href{https://orcid.org/0009-0007-0190-9143}{0009-0007-0190-9143}} \date{14 November 2026} \begin{document}\maketitle \begin{abstract} Terras's coefficient stopping time conjecture asks whether the shortcut Collatz orbit of every integer $n\ge2$ lies below $n$ when its coefficient $\rho_k=3^d/2^k$ first falls below one. Here $k$ counts shortcut steps and $d$ counts their odd inputs. We prove that every eligible realizer descends when the first-crossing word has at most $81$ maximal runs of odd inputs. Any counterexample requires at least $82$ such runs, $10\,439\,860\,591$ steps and $6\,586\,818\,670$ odd inputs. The proof uses a lower bound on the seed from the length of its odd runs, rational approximations to $\log_2 3$, the verified range of Rozier and Terracol, and Rhin's estimate for a linear form in logarithms. The run bound is applied not once but at several interior starts, selected by the size of their contributions to the correction term. We also prove descent for a family with arbitrarily many odd runs. Two further restrictions narrow the remaining words. An initial square $UU$ in a failing word satisfies $|U|<35\log_2 k$. For a counterexample of shortest crossing length, the reciprocal-sum estimate of Nikzad, Frinkle and Bagherzadeh gives a lower bound on $\rho_k$ whose value begins $0.9441733669\ldots$. The open inequality compares the least eligible starting value in each remaining parity class with an explicit threshold. \end{abstract} \section{What is being asked} Write $T(n)=(3n+1)/2$ at odd inputs and $T(n)=n/2$ at even inputs. Fix $n\ge2$. For the parity word $w$ of its first $k$ input states $n,T(n),\dots,T^{k-1}(n)$, let $d$ count the odd inputs. Induction gives \begin{equation}\label{eq:affine} 2^kT^k(n)=3^dn+C_w, \end{equation} where the nonnegative integer $C_w$ depends only on $w$. Write $\rho_k=3^d/2^k$ with $\rho_0=1$. An odd step multiplies the coefficient by $3/2$ and an even step halves it. Let $\tau(n)$ be the first $k$ with $\rho_k<1$, and let $\chi(n)$ be the first $k$ with $T^k(n)M_w . \] Moreover, if the odd runs have lengths $\ell_1,\dots,\ell_s$ and run $i$ starts at time $t_i$, \begin{equation}\label{eq:runsum} M_w=\sum_{i=1}^{s}\frac{1-(2/3)^{\ell_i}}{\rho_{t_i}} . \end{equation} \end{proposition} \begin{proof} The first display is \eqref{eq:affine}. For \eqref{eq:runsum}, each odd step at time $j$ contributes $2^j3^{-d_{j+1}}=1/(3\rho_j)$ to $M_w$, where $d_{j+1}$ counts odd steps through time $j$. Summing a run of $\ell$ consecutive odd steps from $t$ gives $\frac1{3\rho_t}\sum_{r<\ell}(2/3)^r=(1-(2/3)^\ell)/\rho_t$. \end{proof} If a first-crossing word contains $s\ge1$ odd runs, every proper prefix has $\rho_t\ge1$, so each run contributes less than one and each odd step contributes at most $1/3$. Thus \begin{equation}\label{eq:Mbounds} M_wM_w$, since the left side increases with the seed. If a seed $n$ fails, its endpoint satisfies $0\le T^k(n)-n\;\varepsilon_k^{-1}\rho_k(v),\qquad \varepsilon_k=1-\lambda_k(v), \] have $\omega(n)=\sigma(n)$, and rearranging that inequality gives $n(\rho_k^{-1}-1)>M_w$ \emph{exactly}. \textbf{The descent criterion above is therefore Lagarias's, not ours.} Sharper still, Theorem~E of \cite{lagarias1985} states that there is an effectively computable $k_0$ such that for admissible $\mathbf v$ of length $k\ge k_0$, every element of $S(\mathbf v)$ except possibly the smallest has stopping time $k$. Its proof is the one we use: a linear forms in logarithms bound (there Baker--Feldman, $|k\log2-l\log3|\ge k^{-c_0}$; here Rhin's $k^{-13.3}$), fed into (2.14). \textbf{``Only the least element of the class can fail'' is thus Theorem~E, asymptotically in $k$.} What the present section adds is only this. The criterion is stated with no ``sufficiently large'' and no threshold $k_0$: it is a single monotone comparison valid at every $k$, so one evaluation at $r(w)$ decides the class outright. And a failure is pinned to the $s$ endpoints $n,\dots,n+s-1$. The quantitative work is in \S\ref{sec:75}, where the run decomposition \eqref{eq:runsum} replaces Lagarias's cruder admissibility estimate $\rho_k(\mathbf v)\le k2^{(1-\theta)k}$. We do \emph{not} make Lagarias's $k_0$ explicit: his $k_0$ is a threshold on the number of \emph{steps} $k$, whereas our hypothesis bounds the number of \emph{runs}, and his conclusion exempts the smallest element of the class while ours covers it. The two statements are therefore distinct. What we do make explicit, in Corollary \ref{cor:least}, is the exemption itself. \end{remark} \begin{example}\label{ex:59} The exact odd-to-odd valuation word $(1,2,1,3,1)$ from \cite[App.~A.6(b)]{dj-floors2026} has eight shortcut input bits $11011001$ and an odd terminal state. The eight input bits give the class $187\bmod 256$. Terminal oddness adds a ninth bit and selects $187\bmod 512$. Indeed, $T^8(187)=179$, while the other lift, $443$, has $T^8(443)=179+3^5=422$. The first coefficient crossing already occurs at step $7$. Its prefix $1101100$ gives the class $59\bmod128$, and \[ 59-T^7(59)=\frac{(2^7-3^4)\cdot59-85}{2^7}=21. \] Thus $T^7(59)=38$ and every eligible seed in $59+128\Z$ descends at step $7$. Within a period of $512$ there are four representatives of this crossing class, of which one realizes the full valuation word. \end{example} \section{Words with at most 81 odd runs}\label{sec:75} \begin{lemma}[run-length realizer bound]\label{lem:runbound} If an odd run starts at $x$ and lasts $\ell$ steps, the endpoint plus one equals $(3/2)^\ell(x+1)$, so $2^\ell\mid x+1$. At each odd step $x+1$ is multiplied by $3/2$, and at each even step it decreases. Writing $D_i$ for the number of odd steps before run $i$ and $L=\log_2(n+1)$, we get $2^{\ell_i}\le(n+1)(3/2)^{D_i}$, hence $D_{i+1}\le L+\alpha D_i$ with $\alpha=\log_23$, and therefore \[ d\le G_sL,\qquad G_s=\sum_{j=0}^{s-1}\alpha^{\,j}. \] \end{lemma} Neither ingredient is new. The divisibility $2^{\ell}\mid x+1$ across a run of $\ell$ odd steps appears explicitly as Lemma 8 of Hercher \cite{hercher2023}, written there as $1\le k_i\le x_i$ with $x_i=\log_2(n_i+1)$; that is the antecedent we are aware of. Garner's contribution is separate and is recorded below: the coefficient stopping time itself \cite[\S4]{garner1981} and the Diophantine study of powers of $2$ against powers of $3$ \cite[\S5--\S6]{garner1981}. The geometric sum $G_s=(\alpha^s-1)/(\alpha-1)$ is likewise the sum $(\delta^{m_2}-1)/(\delta-1)$ appearing in Hercher's Theorem 21, with the \emph{same} constant $\delta=\log_23=\alpha$; his recurrence $x_{i+1}<\delta x_i$ is the purely geometric form that cyclicity permits, while the affine form $D_{i+1}\le L+\alpha D_i$ above re-injects the seed size $L$ at every run because a finite path has a fixed starting point. The method of the next theorem --- a Diophantine bound on powers of $2$ against powers of $3$, combined with a finite verification range --- is Garner's \cite[\S6--\S7]{garner1981}; see Remark \ref{rem:history}. Lemma \ref{lem:runbound} charges the seed size $L$ once per run, so applying it from the seed alone costs a factor $G_s$. The next two lemmas apply it instead at several interior starts. Which starts are available is decided by the correction term: a run whose contribution $m_i$ to $M_w$ is large has small $\rho_{t_i}$, hence small $x_i$. \begin{lemma}[splitting]\label{lem:split} Let a first-crossing word have $s\le81$ odd runs, let its realizer fail to descend, and suppose $0<\lambda<\Lambda$. Then \begin{equation}\label{eq:split0} n+1<1+81/\lambda . \end{equation} If moreover $\lambda>r/B$ for an integer $r\ge1$, then $s>r$, and among the $s-1$ noninitial runs the $r$ with the largest contributions $m_i$ all begin at integers satisfying \begin{equation}\label{eq:split} x_i+1<1+\frac{(81-r)\,(1+81/B)}{\lambda-r/B}. \end{equation} The two bounds are proved together but neither contains the other: putting $r=0$ in \eqref{eq:split} would leave a spurious factor $1+81/B$, so \eqref{eq:split0} is the sharper statement for the initial run. \end{lemma} \begin{proof} A failure gives $(e^\Lambda-1)n\le M_w$ by Proposition \ref{prop:comparator}, and $e^t-1>t$, so $n\LambdaB$ and $\lambda>r/B$, \[ s>M_w>n\Lambda>B\lambda>r, \] so $s\ge r+1$ and at least $r$ noninitial runs exist. Order the noninitial contributions $m_{(1)}\ge\dots\ge m_{(s-1)}$ and put $m^*=m_{(r)}$, the smallest selected one. Bound by $1$ exactly $r$ contributions, namely the initial $m_1$ together with the $r-1$ larger selected ones $m_{(1)},\dots,m_{(r-1)}$. Every one of the remaining $s-r$ contributions, which are $m_{(r)}=m^*$ itself and $m_{(r+1)},\dots,m_{(s-1)}$, is at most $m^*$. Hence \[ n\Lambda\frac{n\Lambda-r}{s-r}. \] It matters that $m_{(r)}$ is counted in the second group: moving it to the first gives $r+1$ discards and replaces $\lambda-r/B$ by $\lambda-(r+1)/B$, which is negative for three of the five rows of Theorem \ref{thm:75}. For a selected index $i$ we have $m_i\ge m^*$, and $m_i=(1-(2/3)^{\ell_i})/\rho_{t_i}$ gives $\rho_{t_i}<1/m_i$. Since $x_i=\rho_{t_i}(n+M_i)$ with $M_i\le M_w<81$, \[ x_i<\frac{n+81}{m^*}<\frac{(n+81)(s-r)}{n\Lambda-r} =\frac{(1+81/n)(s-r)}{\Lambda-r/n}. \] The seed enters both parts, and $n>B$ improves both: $81/n<81/B$ shrinks the numerator while $-r/n>-r/B$ enlarges the denominator, as does $\Lambda>\lambda$. With $s\le81$ this is \eqref{eq:split}. \end{proof} The conclusion of Lemma \ref{lem:split} is an \emph{absolute} height bound, not a statement about the relative drift $\rho_{t_i}$. Only one implication is used: a \emph{selected} run has $m_i\ge m^*$, hence $\rho_{t_i}<1/m^*$. The converse fails, because $m_i$ depends on the run length as well as on $\rho_{t_i}$: a run with small $\rho_{t_i}$ but small $\ell_i$ also has small $m_i$ and may well be discarded. \begin{lemma}[restart]\label{lem:restart} Order the $r$ selected starts of Lemma \ref{lem:split} chronologically and prepend the initial start. If each of these $r+1$ starts satisfies $x+1<2^H$, then \[ dB$. Put $\Lambda=k\ln2-d\ln3>0$. By Proposition \ref{prop:comparator} a failure gives $(e^\Lambda-1)n\le M_w$, hence with \eqref{eq:Mbounds} \begin{equation}\label{eq:lambda} 0<\Lambda1/(3B)$. With the second half of \eqref{eq:lambda} this forces $a/b<\alpha0$. Since $k/d>\alpha$, $q\Lambda+de=(qk-pd)\ln2\ge\ln2$. Four distinct convergents are used below: one for the lower bound on $d$ and three in the table, namely \[ \begin{array}{llll} \mathrm{I} & (9\,881\,527\,843\,552\,324, & 6\,234\,549\,927\,241\,963), & e_{\mathrm{I}}\in(5.22,5.24)\cdot10^{-18},\\ \mathrm{II} & (79\,641\,170\,620\,168\,673\,833, & 50\,247\,984\,153\,525\,417\,450), & e_{\mathrm{II}}\in(6.71,6.72)\cdot10^{-21},\\ \mathrm{III} & (6\,724\,555\,128\,221\,608\,268, & 4\,242\,721\,909\,926\,539\,673), & e_{\mathrm{III}}\in(1.24,1.25)\cdot10^{-19},\\ \mathrm{IV} & (423\,372\,672\,964\,960\,618, & 267\,118\,416\,222\,671\,843), & e_{\mathrm{IV}}\in(1.64,1.65)\cdot10^{-18}. \end{array} \] Convergent I gives $d>(0.69-q\cdot81/B)/(5.24\cdot10^{-18})>1.28\cdot10^{17}$. This is the lower bound that the rest of the proof must contradict. \emph{The head.} Read the same identity in the other direction: for a convergent $p/q<\alpha$ and any upper bound $d\le d^{\dagger}$ one gets $\Lambda\ge(\ln2-d^{\dagger}e)/q=:\lambda$. Each such $\lambda$ feeds Lemma \ref{lem:split}, whose bound and the initial-start bound give an $H$ with $x+1<2^H$ at all $r+1$ segment starts, and then Lemma \ref{lem:restart} returns a new upper bound $d1.28\cdot10^{17}$. Each $\lambda$ exceeds $r/B$, as Lemma \ref{lem:split} requires, and each is computed from the \emph{printed} ceiling of the row above, so the table may be checked line by line. \emph{The tail.} For $d\ge2^{65}$, Rhin's effective estimate \cite{rhin1987} (in the form used by \cite[Prop.~6.3]{rozier2026}) gives $\Lambda\ge k^{-13.3}$, so $n<81(2d)^{13.3}$ and $\log_2(n+1)<7+13.3\log_2(2d)$. With $2^j\le d<2^{j+1}$, Lemma \ref{lem:runbound} applied from the seed alone forces $2^j\le dB$. Combined with \eqref{eq:lambda}, the Farey interval then gives the two integer bounds. Rhin's estimate enters only in the case $d\ge2^{65}$; its hypothesis is $H=\max(|u_1|,|u_2|)\ge2$ \cite[Prop.~6.3]{rozier2026}, and with $u_1=k$, $u_2=-d$ and $k>d$ one has $H=k$, so the form used is $\Lambda\ge k^{-13.3}$. Beyond these, the proof of Theorem \ref{thm:75} uses only Proposition \ref{prop:comparator} with \eqref{eq:runsum} and \eqref{eq:Mbounds}, Lemma \ref{lem:runbound}, and certified enclosures of $\ln2$ and $\ln3$. It uses no formalised or externally computed bound on individual runs; in particular it is independent of the two-run bounds discussed in Remark \ref{rem:history}. Whether $81$ is optimal for this method is open: the thresholds depend on the available convergents and on the choice of $r$ in each row, and we have not searched that space exhaustively. \end{remark} \begin{remark}[Historical context, and what is claimed here]\label{rem:history} Every method used above is older than this note; what is new is the combination and the resulting explicit constants for \emph{finite first crossings}. \emph{Garner \cite{garner1981}} is the direct precedent. He defined the coefficient stopping time, also formulated the conjectured equality with the stopping time --- he cites Terras \cite{terras1976} as his reference [3] --- and verified it for $n<1\,150\,000$, identified the realizing residue class modulo $2^{1+p(m)}$ with $p(m)=\lfloor m\log_23\rfloor$, and proved (his \S6) that $\sigma(n)=\kappa(n)$ whenever the number of odd \emph{terms} $m$ is below a bound assembled from a study of the powers of $2$ and $3$ --- his tabulated breakpoints are denominators of convergents of $\log_23$ --- together with a verified range. His \S7 applies this to cycles, obtaining at least $35\,400$ terms from $n>6\cdot10^7$. Theorem \ref{thm:75} is that argument with two changes: the hypothesis bounds the number of odd \emph{runs} rather than odd terms, which is strictly weaker because a single run may contain arbitrarily many odd steps, and the modern verification range and an explicit Farey sandwich replace his tabulated constants. \emph{Simons and de Weger \cite{simonsdeweger2005}} developed the run-growth approach for nontrivial cycles, bounding the number of local minima; \emph{Hercher \cite{hercher2023}} strengthened that line to $m\ge92$. Their counts and ours are the same kind of quantity but not the same quantity: in a cycle the odd runs close up, so the number of local minima equals the number of runs, whereas in a finite first-crossing word the first run may begin at the seed, which is not a local minimum, and the path stops instead of closing. The last odd run is complete: were step $k$ odd we would have $2^{k-1}\le3^{d-1}$ and $2^k>3^d$, whence $3^d<2^k\le2\cdot3^{d-1}<3^d$. The crossing therefore occurs in the even block that follows the last run, as the proof of Lemma \ref{lem:restart} uses. Their bounds also use cyclicity, through the closing of the product around the cycle, which a finite path does not provide. With a modern verification range their cycle bound becomes $m\ge83$ \cite{hercher2023}. A different route to a run threshold proceeds from a bound on the first two runs of every small seed. Justin Alick has formalised such local two-run bounds in Lean~4 \cite{alick2026}, among them the declaration \texttt{allPairs70}, which establishes $\ell_1+\ell_2\le76$ for every odd $n<2^{70}$ and is sharp at $n=806\,134\,805\,793\,934\,253\,823$ with runs of lengths $8$ and $68$. Combined with the present estimates that bound yields a threshold of $76$ runs. The proof of Theorem \ref{thm:75} does not use it, and the two routes are independent. \textbf{Nothing in this note raises the first-crossing threshold to $92$, and Hercher's cycle conclusion must not be read as applying to finite first crossings.} \emph{Lagarias \cite{lagarias1985}} is the closest predecessor of all, and Remark~\ref{rem:classical} gives the locators. His (2.14) is our descent criterion verbatim after rearrangement, and his Theorem~E already proves, by exactly our route --- linear forms in logarithms fed into that criterion --- that for admissible words of length $k\ge k_0$ only the smallest element of the class can fail. Two things are left to us, and they are different from each other: Corollary~\ref{cor:least} removes the length restriction $k\ge k_0$ from that exemption, while Theorem~\ref{thm:75} proves outright descent, for the smallest element included, under a hypothesis on the run count rather than on $k$. Lagarias also records \cite[\S2.3]{lagarias1985} that Terras \cite{terras1976} and ``later Garner'' \cite{garner1981} conjectured the equality, which settles the question of independence raised above. \emph{Rozier and Terracol \cite{rozier2026}} supply the verification range and, in their Remark~5.5, credit Simons with proposing classification by local maxima. Three classes of object must be kept apart. A \emph{nontrivial cycle} is the subject of \cite{garner1981,simonsdeweger2005,hercher2023}. A \emph{paradoxical sequence} \cite[Def.~1.1]{rozier2026} is a finite trajectory with $T^j(n)\ge n$ and coefficient below one at step $j$, for \emph{any} such $j$. A \emph{first-crossing failure}, the object here, is a paradoxical sequence with $j=\tau(n)$, hence a proper subclass. Their Theorem 5.3 excludes paradoxical sequences with $93\le j\le301\,993$ outright, and all with $4615\le n\le2.8\cdot10^{19}$; the $301\,993$ is the extent of a finite computation, whereas the bound $k\ge10\,439\,860\,591$ of Theorem \ref{thm:75} is a lower bound on a counterexample derived from the Farey sandwich. A counterexample must satisfy both, and ours is the stronger constraint; there is no conflict. \emph{Hikawa \cite{hikawa2026}} reviews the problem and singles out exactly the triple studied here --- the correction term $q_v$, the coefficient difference $2^k-3^d$, and the least positive representative $r_v$ of the residue class. \textbf{Priority for the specific first-crossing statements of this note is therefore not established}; we have not located a match, which is not the same as there being none. \end{remark} \section{A family with arbitrarily many runs} \begin{proposition}\label{prop:110} Repeat the block $110$ exactly $h\ge1$ times and append zeros up to the first coefficient crossing. Every eligible realizer descends at that crossing. \end{proposition} \begin{proof} One block acts by $H(x)+5=9(x+5)/8$, so $-5$ is its fixed point. The block's full coefficient is $9/8$ and its two nonempty proper-prefix coefficients are $3/2$ and $9/4$. Consequently the first crossing occurs among the appended zeros at $k=\lceil h\log_2 9\rceil\le4h$. Iteration gives $H^h(n)+5=9^h(n+5)/8^h$. Integrality implies $8^h\mid n+5$, hence $n\ge8^h-5$. The normalized memory is $5(1-(8/9)^h)<5$ and the appended zeros leave it unchanged. For $h\le63$, $k\le252$ is below the crossing-length bound in Theorem \ref{thm:75}. For $h\ge64$, Rhin's estimate gives $n\Lambda\ge(8^h-5)/(4h)^{13.3}>5$. Indeed, $8^{64}-5>5\cdot256^{14}$, and $((h+1)/h)^{14}\le(65/64)^{14}<8$, so $8^h-5>5(4h)^{14}$ persists by induction. A failure requires $n\Lambdan$, so $x-n$ is a positive multiple of $2^\ell$. Equivalently $2^\ell(x-n)$ is a positive multiple of $2^{2\ell}$. At each step $T(y)+1\le(3/2)(y+1)$, whence \[ 2^\ell\le x-n0$. Then $n\Lambda\exp(-B_i/3). \] \end{proposition} \begin{proof} We use the external height record $H(y)\le2456$ through $D$, reported in \cite[eq.~(34)]{nfb2026}. Here $H$ counts ordinary Collatz operations to first arrival at $1$. For $2\le y\le D$, this gives $\tau(y)\le\chi(y)\le H(y)\le2456$, since shortcut steps combine ordinary operations. The lower bound on $k$ in Theorem~\ref{thm:75} therefore gives $n>D$. For $0\le jD$. If two of $x_0,\ldots,x_{k-1}$ coincided, the intervening segment would contain a cycle of length $L\frac{C_w}{2^k-3^d}=\frac{M_w}{\rho_k^{-1}-1}. \] Equality already gives a counterexample. A lower bound for $r(w)$ or an upper bound for the right side can certify a word if it preserves this strict inequality. The rational-shadow estimates of \cite{dj-shadow2026} supply bounds for individual blocks. Applying one to this question requires transporting it back to the starting seed and comparing it with the threshold of the actual first-crossing prefix. \newpage \section*{Status of the results} It is worth separating three things explicitly. \emph{Established before this note.} The residue-class correspondence for parity words (Terras \cite{terras1976}); the affine form, the descent criterion $n>\varepsilon_k^{-1}\rho_k(v)$, and --- in Theorem~E --- the statement that for admissible words of length $k\ge k_0$ only the smallest element of a class can fail, proved from linear forms in logarithms (Lagarias \cite[\S2.2--\S2.3, Thm.~E]{lagarias1985}); the coefficient stopping time, the conjecture $\kappa=\sigma$, and the method of combining a Diophantine bound on $2^a$ against $3^b$ with a finite verification range, including a descent theorem under a bound on the number of odd \emph{terms} (Garner \cite{garner1981}); run-growth bounds for the number of local minima in a nontrivial cycle (Simons--de Weger \cite{simonsdeweger2005}, Hercher \cite{hercher2023}); verification of $t=\tau$ to $2.8\cdot10^{19}$ and the exclusion of paradoxical sequences with $93\le j\le301\,993$ (Rozier--Terracol \cite{rozier2026}); the reciprocal-sum estimate (Nikzad--Frinkle--Bagherzadeh \cite{nfb2026}). \emph{Offered here.} The transfer of Garner's method to finite first crossings with the hypothesis weakened from odd \emph{terms} to odd \emph{runs}, which is what Lemma \ref{lem:runbound} buys; the resulting explicit constants $k\ge10\,439\,860\,591$, $d\ge6\,586\,818\,670$ and a threshold of $82$ runs; the splitting and restart lemmas, which apply the run bound at several interior starts rather than once from the seed; descent for the whole repeated-$110$ family, which has arbitrarily many runs; the finite-word square bound $|U|<35\log_2k$; Corollary \ref{cor:least}, which removes the length restriction from the conclusion that only the least eligible seed can fail; and the reciprocal-sum restriction $\rho_k>0.9441733669$ on a shortest counterexample. Subject to the priority caveat of Remark \ref{rem:history}. \emph{Unresolved.} Combining Theorem \ref{thm:75} with Corollary \ref{cor:least}, what remains is exactly this: \textbf{the least eligible seed of a first-crossing word with $82$ or more odd runs}. A lower bound on that seed for a word with a shallow crossing and no long initial repetition --- the inequality of \S\ref{sec:open} --- is open. Whether $81$ is optimal for the present method is also open; the thresholds depend on the available convergents and on the choice of $r$ in each row, and that space has not been searched exhaustively. Whether Hercher's cycle techniques transfer to finite first-crossing trajectories is a separate question, not addressed here. Terras's conjecture itself is open, as is the Collatz conjecture. \section*{Reproducibility} The companion verifier checks the numerical inequalities used in the proofs, recomputes the finite bracket count, and verifies the worked examples using integer and rational arithmetic. It lives in the companion repository \href{https://github.com/innerlightr-wq/terras-first-crossing}{\texttt{github.com/innerlightr-wq/\allowbreak terras-first-crossing}}, with the verification code pinned at the immutable commit \href{https://github.com/innerlightr-wq/terras-first-crossing/commit/a93b1e82d13fab7024c380ae36320e4ec62a73f9}{\texttt{a93b1e8}}. Run the following commands from the project root. \begin{verbatim} python3 scripts/verify_crossing.py python3 scripts/verify_81_runs.py \end{verbatim} The first script is the pinned release verifier and is unchanged. The second is new and checks the inputs of Theorem \ref{thm:75} independently: the certified enclosures of $\ln2$ and $\ln3$, the sign of $e=q\ln3-p\ln2$ for each convergent used, the lower bound $d>1.28\cdot10^{17}$, the whole tail $d\ge2^{65}$ together with the failure at $j=64$, and each row of the five-row table with its $\lambda$, $H$ and resulting bound on $d$, each row's $\lambda$ being recomputed from the \emph{printed} ceiling of the row above. It also checks the four printed $e$-enclosures, the hypothesis of Corollary \ref{cor:least}, and that the crossing step is even, and that the displayed rounded form of the lower bound on $d$ also closes. It reports $31$ checks. The script checks the numerical inequalities with integers and rationals, encloses $\ln2$ and $\ln3$ by a 96-term series with a geometric tail, recomputes the finite bracket sum in (164) of \cite{nfb2026}, and encloses each exponential cutoff by alternating Taylor bounds. The SHA-256 digest of \texttt{verify\_crossing.py} begins \texttt{f852c88311071a0d}; that of \texttt{verify\_81\_runs.py} begins \texttt{ca338694ac39ff93}. Both scripts are included in the deposit bundle. The orbit and residue-class examples are finite checks. The proofs for arbitrary word lengths are in the text. The finite seed verification of Rozier--Terracol, Rhin's estimate, the high-path accounting of Nikzad--Frinkle--Bagherzadeh and the height record cited above are external inputs. The general case of Terras's conjecture remains open. \section*{AI assistance and author responsibility} The author directed this research and used Claude (Anthropic) and ChatGPT (OpenAI) for proof exploration and revision, cross-checking the reciprocal-sum estimates against the source paper, developing and testing the exact-arithmetic verifiers, and drafting the manuscript. For Theorem \ref{thm:75} specifically, the splitting and restart lemmas were developed in author-directed work with ChatGPT, and the numerical certificate was then reconstructed independently by Claude from the lemma statements rather than from the author's script: the discard count in Lemma \ref{lem:split} was confirmed to be load-bearing, the convexity step in Lemma \ref{lem:restart} was checked against every composition of $s\le13$, and the five rows were recomputed by optimising over all certified convergents, which improved the first three height bounds and reproduced the last two. The author takes responsibility for the mathematical claims, their interpretation, and any errors. \begin{thebibliography}{10} \bibitem{alick2026} Justin Alick. \newblock Collatz $m$-cycles: a {L}ean~4 formalisation, 2026. \newblock Declaration \texttt{allPairs70 : AllPairs 70 76} in \texttt{Collatz/MCycleTwoRiseHi2.lean} lines 15--19, proved by \texttt{decide +kernel}, at commit \texttt{d1895945}. \url{https://github.com/justinalick/collatz-m-cycles-101-lean}. \bibitem{eliahou1993} Shalom Eliahou. \newblock The $3x+1$ problem: new lower bounds on nontrivial cycle lengths. \newblock {\em Discrete Mathematics}, 118:45--56, 1993. \bibitem{garner1981} Lynn~E. Garner. \newblock On the {C}ollatz $3n+1$ algorithm. \newblock {\em Proceedings of the American Mathematical Society}, 82(1):19--22, 1981. \bibitem{hercher2023} Christian Hercher. \newblock There are no {C}ollatz-$m$-cycles with $m\le91$. \newblock {\em Journal of Integer Sequences}, 26(3), 2023. \newblock Article 23.3.5. Corrigendum (Theorem 21), \url{https://cs.uwaterloo.ca/journals/JIS/VOL26/Hercher/corrigendum.pdf}; the conclusion $m\ge92$ is unaffected. \bibitem{hikawa2026} Kazunobu Hikawa. \newblock Research notes on {Terras}' coefficient stopping time problem {I}: Previous results and open problems, 2026. \newblock English translation, November 2026; Japanese original, Zenodo, 13 November 2026, DOI 10.5281/zenodo.24187839. \bibitem{dj-sturmian2026} Elias~De Jes\'us. \newblock Beyond {S}turmian words in the $3x+1$ periodicity conjecture: A periodic-approximation and arithmetic-height bridge, 2026. \newblock Revision 5, 13 November 2026; Revision 4 deposited as Zenodo DOI 10.5281/zenodo.23070093. \bibitem{dj-shadow2026} Elias~De Jes\'us. \newblock Exact realizer congruences and rational-shadow bounds for {C}ollatz words, 2026. \newblock Zenodo preprint, 31 October 2026. \href{https://doi.org/10.5281/zenodo.23847289}{\texttt{doi:10.5281/zenodo.23847289}}. \bibitem{dj-floors2026} Elias~De Jes\'us. \newblock Exact realizer floors for one- and two-defect valuation words in the accelerated $3x+1$ map, 2026. \newblock Zenodo preprint, 22 October 2026. \href{https://doi.org/10.5281/zenodo.23579193}{\texttt{doi:10.5281/zenodo.23579193}}. \bibitem{dj-indep2026} Elias~De Jes\'us. \newblock Independent verification of high-path reciprocal bounds for nonrepeating {C}ollatz trajectories, 2026. \newblock Unpublished manuscript, 3 November 2026. \bibitem{lagarias1985} Jeffrey~C. Lagarias. \newblock The $3x+1$ problem and its generalizations. \newblock {\em American Mathematical Monthly}, 92(1):3--23, 1985. \bibitem{nfb2026} Behzad Nikzad, Karl Frinkle, and Fatemeh Bagherzadeh. \newblock On bounds for reciprocal sums and residues in {C}ollatz orbits, and their consequences, 2026. \newblock Hexagon preprint 2611.00033v1, \url{https://hexagonmath.org/2611.00033v1}. \bibitem{rhin1987} Georges Rhin. \newblock Approximants de {P}ad\'e et mesures effectives d'irrationalit\'e. \newblock In {\em S\'eminaire de Th\'eorie des Nombres, Paris 1985--86}, volume~71 of {\em Progress in Mathematics}, pages 155--164. Birkh\"auser Boston, 1987. \bibitem{rozier2026} Olivier Rozier and Claude Terracol. \newblock Paradoxical behavior in {C}ollatz sequences. \newblock {\em Discrete Mathematics}, 349:115167, 2026. \newblock arXiv:2502.00948v5. \bibitem{simonsdeweger2005} John Simons and Benne de~Weger. \newblock Theoretical and computational bounds for $m$-cycles of the $3n+1$ problem. \newblock {\em Acta Arithmetica}, 117(1):51--70, 2005. \newblock The bound was subsequently raised; Hercher \cite{hercher2023} reports $m\ge76$ for Simons--de Weger, and $m\ge83$ on substituting a modern verification range. \bibitem{terras1976} Riho Terras. \newblock A stopping time problem on the positive integers. \newblock {\em Acta Arithmetica}, 30:241--252, 1976. \end{thebibliography} \end{document}