% Source unit: introduction \section{Introduction} Let $K$ be a knot in $S^3$, we take it to be understood, and suppress it from the notation. Let $J\colon\mathbb N\to\mathbb Q[q^{\pm1}]$ be its colored Jones function. Set $EJ(n)=J(n+1)$ and $QJ(n)=q^nJ(n)$, so $EQ=qQE$. The annihilator left ideal $\Ann(J)$ becomes principal after localization to $\mathbb Q(q,Q)[E;\sigma]$. We use $\Ann(J)$ for the ideal both before and after localization. The order of an operator is its degree as a polynomial in $E$. A nonzero annihilator is minimal if it has least order. After localization, $\Ann(J)$ has a unique monic minimal generator; clearing its denominators gives $\widehat A$. The specialization map is $\varepsilon(q,E,Q)=(1,L,M^2)$. Let $M=S^3\setminus N(K)$, where $N(K)$ is an open tubular neighborhood of $K$.\footnote{We assume the reader can distinguish the meaning of letters from context. E.g. here $M$ is the 3-manifold, and an eigenvalue of $\rho(\mu)$.} Thus $\partial M\cong S^1\times S^1$, with $\pi_1(\partial M)=\langle\lambda,\mu\rangle\cong\mathbb Z^2$. Write $X(M)$ for its $\mathrm{SL}_2(\mathbb C)$-character variety. The character variety of the boundary, $X(\partial M)$, has coordinate ring $\mathbb C[M^{\pm1},L^{\pm1}]^\eta$, where $\eta(M,L)=(M^{-1},L^{-1})$. For a representation $\rho\colon\pi_1(M)\to\mathrm{SL}_2(\mathbb C)$, the commuting matrices $\rho(\mu)$ and $\rho(\lambda)$ have a common eigenline, with eigenvalues $M,L$~\cite[§1, p.~1145]{DunfieldGaroufalidis2004}. These pairs lift the image of the restriction map $X(M)\to X(\partial M)$ to eigenvalue coordinates. Let $D\subset\mathbb C^2$ be the plane curve obtained by taking the Zariski closure of the image.\footnote{Points of the Zariski closure are discarded, as in~\cite[§§2.2--2.3]{CCGLS1994}.} Its ideal $I(D)$ is principal, and its generator $A$ is the $A$-polynomial\footnote{Lê denotes this polynomial by $A'_K=(L-1)A_K$~\cite[§3.2]{Le2006}. This is a standard normalization, since $L-1$ divides $A$ for a knot in a homology sphere~\cite[§2.5]{CCGLS1994}.} \cite[§§2.2--2.3]{CCGLS1994}. For nonzero $f,g\in\mathbb C[M,L]$, we say that $f$ and $g$ are $M$-essentially equal, written $f\asymp g$, if $f/g\in\mathbb C(M)^\times$. We say that $f$ is $M$-essentially divisible by $g$ if $g$ divides $h(M)f$ for some nonzero $h(M)\in\mathbb C[M]$. Two algebraic subsets of $(\mathbb C^*)^2$ are $M$-essentially equal if they become equal after adding finitely many lines parallel to the $L$-axis. A set is $M$-essentially contained in another if it is contained after adding such lines to the latter \cite[§§0.2, 3]{Le2006}. We write $\Red(f)$ for the product of the distinct irreducible factors of $f$. \begin{conjecture}[Strong AJ~\cite{Garoufalidis2004,Le2006}] For every knot, $\varepsilon^*(\widehat A)\asymp A$. \end{conjecture} \begin{conjecture}[Weak AJ~\cite{Le2006}] The zero sets of $\varepsilon^*(\widehat A)$ and $A$ are $M$-essentially equal. \end{conjecture} Equivalently, weak AJ asserts $\Red(\varepsilon^*(\widehat A))\asymp A$, which is Zhang's reformulation of the AJ conjecture~\cite{Zhang2026}. A framing of a knot is a continuous, nowhere-zero normal vector field along it, considered up to homotopy. We say that a framed knot in $S^3$ has zero framing if the disjoint parallel copy obtained by pushing it in the framing direction has linking number zero with the original knot. We use this framing for $K$. Put $R=\mathbb C[t^{\pm1}]$,\footnote{We use $q=t^{-4}$ throughout. Here $t$ is the Kauffman bracket parameter and $q$ is the Jones variable. Thus $t=-1$ gives $q=1$. With this substitution, we write $\widehat A$ in the operators $LJ(n)=J(n+1)$ and $MJ(n)=t^{-2n}J(n)$, and the specialization in the AJ conjecture is $\varepsilon(t,L,M)=(1,L,M)$.} and for an oriented 3-manifold $M$, let the Kauffman bracket skein module $\mathcal S_t(M)$ be the quotient of the free $R$-module on isotopy classes of framed unoriented links by $L_\times=tL_0+t^{-1}L_\infty$ and $L\sqcup\bigcirc=-(t^2+t^{-2})L$. Its elements are called skeins. \input{assets/f29c235d40b1-skein_relations.tex} The first relation replaces a crossing by its two smoothings, with the links unchanged outside a ball. In the second, the dashed box denotes $L$, and $\bigcirc$ is a zero-framed unknot in a ball disjoint from $L$~\cite[§1.1]{Le2006}. Every framed link in $S^3$ has a blackboard-framed diagram. Therefore $\mathcal S_t(S^3)=R\langle\varnothing\rangle$ by resolving its crossings and removing trivial circles~\cite[§1.2]{Le2006}. For a framed link $L$ in $S^3$, write $\langle L\rangle$ for the coefficient of $\varnothing$ in its skein class. Gluing $N(K)$ to $M$ gives the pairing $\mathcal S_t(N(K))\otimes_R\mathcal S_t(M)\longrightarrow R$, with $\langle x,y\rangle=\langle x\cup y\rangle$ \cite[§§1.2, 4.4]{Le2006}. The boundary skein algebra $\mathcal S_t(\partial M\times I)$ acts on $\mathcal S_t(M)$. For $a\in\mathcal S_t(\partial M\times I)$, put $\Theta(a)=a\cdot\varnothing$. The cyclic submodule generated by $\varnothing$ is $\mathcal S_t(\partial M\times I)\varnothing \cong\mathcal S_t(\partial M\times I)/\Ann(\varnothing)$ by the first isomorphism theorem for left $R$-modules. Since a boundary skein can be glued on either side, the pairing factors through the map $\mathcal S_t(N(K))\otimes_{\mathcal S_t(\partial M\times I)} \mathcal S_t(M)\longrightarrow\mathcal S_t(S^3)$. The peripheral ideal of Frohman--Gelca--Lofaro is $\ker\Theta=\Ann(\varnothing)$. Following Lê, define the orthogonal ideal $\mathcal O$ to consist of $a\in\mathcal S_t(\partial M\times I)$ such that $\langle u,\Theta(a)\rangle=0$ for every $u\in\mathcal S_t(N(K))$~\cite[§4.4]{Le2006}. Their construction gives the following~\cite[§3.3]{FrohmanGelcaLofaro2002}. \[ \begin{tikzcd}[column sep=large,row sep=large] \Ann(\varnothing) \arrow[r,hook] \arrow[d,hook,"\Phi"'] & \mathcal S_t(\partial M\times I) \arrow[r,"\Theta"] \arrow[d,hook,"\Phi"] & \mathcal S_t(M) \\ \mathcal O \arrow[r,hook] & \mathcal T. \end{tikzcd} \] Here $\Phi$ identifies the boundary skein algebra with the invariant subalgebra $\mathcal T^\eta$ of the quantum torus $\mathcal T$. We regard $\mathcal O$ as a subideal of $\mathcal T^\eta$ via $\Phi$. Therefore peripheral relations annihilate the colored Jones function.\footnote{They kill the colored Kauffman brackets, which equal the colored Jones function after $t\mapsto it$, where $i=\sqrt{-1}$~\cite[p.~1235]{Gelca2002}.} \Needspace{5\baselineskip} Let $\Cheb_n$ be the sequence of orthogonal polynomials defined by $\Cheb_{n+1}(z)=z\Cheb_n(z)-\Cheb_{n-1}(z)$, with initial conditions $\Cheb_0(z)=1$ and $\Cheb_1(z)=z$. For $j\ge0$, let $K^j$ be $j$ parallel copies of $K$ in this framing, with $K^0=\varnothing$. For a polynomial $f(z)=\sum_j a_jz^j$, its cabling is the skein $f(K)=\sum_j a_jK^j\in\mathcal S_t(N(K))$. For this framing, we use the normalization $J(n+1)=(-1)^n\langle\Cheb_n(K)\rangle/[n+1]$, where $[m]=(t^{2m}-t^{-2m})/(t^2-t^{-2})$. Thus $J_{\bigcirc}(n)=1$~\cite[§1.3]{Le2006}. Gelca proves that a nonzero peripheral relation and finitely many initial values determine the sequence~\cite{Gelca2002}. At $t=-1$, the Kauffman bracket skein module $\mathcal S_{-1}(M)$ has a natural commutative algebra structure, and its quotient by the nilradical is $\mathbb C[X(M)]$, the coordinate ring of the character variety~\cite{Bullock1997,PrzytyckiSikora2000}. A choice of spin structure on $M$ gives an algebra isomorphism $\mathcal S_1(M)\cong\mathcal S_{-1}(M)$ by changing the signs of framed links, so the same quotient description holds at $t=1$ \cite[Theorem 2.2, Proposition 2.4]{PrzytyckiSikora2000}. The specialization diagram is \[ \begin{tikzcd}[column sep=large,row sep=large] \mathcal S_t(\partial M\times I) \arrow[r,"\Theta"] \arrow[d,"\varepsilon"'] & \mathcal S_t(M) \arrow[d,"\varepsilon"] \\ \mathbb C[X(\partial M)] \arrow[r,"\theta"'] & \mathbb C[X(M)]. \end{tikzcd} \] Here $\theta$ is induced by restriction of characters~\cite[§4.2]{Le2006}. The left ideal of $\mathcal T$ generated by $\Phi(\Ann(\varnothing))$ becomes principal upon localization of the coefficient ring to $\mathbb C(t,M)$, its unique monic minimal generator is $\widehat P$~\cite[§4.3]{Le2006}. A relation that kills $\varnothing$ pairs to zero with every $\Cheb_n(K)$, so $\widehat P\in\Ann(J)$. Therefore $\widehat P=C\widehat A$ for some operator $C$~\cite[§§4.4--4.5]{Le2006}. Lê defines $B$ as a generator of the extension of $\ker(\mathbb C[M^{\pm1},L^{\pm1}]^\eta \longrightarrow\mathbb C[X(M)])$ to $\mathbb C(M)[L^{\pm1}]$~\cite[§§3.3--3.4]{Le2006}. Lê--Tran show that $B$ is obtained from $A$ by removing its maximal factor depending only on $M$. For two-bridge knots, $B=A$ \mbox{\cite[Corollary~2.3, Proposition~2.4]{LeTran2015}}. \begin{conjecture}[{Lê's conjecture~\cite[Conjecture 3]{Le2006}}] For every knot $K\subset S^3$, this pairing has trivial right kernel. If $y\in\mathcal S_t(M)$ satisfies $\langle x,y\rangle=0$ for every $x\in\mathcal S_t(N(K))$, then $y=0$. \end{conjecture} Nondegeneracy identifies the peripheral and orthogonal ideals. The latter is the invariant part of the recurrence ideal, so localization gives the following formulation, which we attribute to Lê \cite[§§4.4--4.5]{Le2006}. \begin{conjecture} For every knot $K\subset S^3$, $\widehat P=\widehat A$. \end{conjecture} Zhang gave the first family of composite-knot counterexamples to strong AJ, while establishing weak AJ for the same family~\cite{Zhang2026}. \begin{theorem}[{Zhang~\cite[Theorem 1.1]{Zhang2026}}] Let $K=T(a,b)\mathbin{\#}T(c,d)$, where the pairs are coprime, $|a|>b\ge2$, $|c|>d\ge2$, and $a,c$ have the same sign. Then Conjecture~2 holds for $K$, but repeated factors of $A$ occur in $\varepsilon^*(\widehat A)$ precisely when $ab=cd$ and $a\ne c$. \end{theorem} His proposed correction is precisely the weak AJ conjecture. We give the first prime-knot counterexample to the strong AJ conjecture. \begin{theorem}\label{thm:main} The strong AJ conjecture for hyperbolic prime knots is false. \end{theorem} \begin{pf} Let $K=8_{11}$, we check $\widehat P=\widehat A$ by showing it has minimal order. The specialization satisfies $\varepsilon^*(\widehat A)\asymp (L+M^6)A$. \end{pf} The same computation disproves Lê's assertion that $\varepsilon^*(\widehat P)\asymp B$ for every two-bridge knot \mbox{\cite[Theorem~3]{Le2006}}. We explain the gap in his proof in \mbox{\S~\ref{sec:811}}. Since $L+M^6$ divides $A$, the factorization also proves weak AJ for $8_{11}$. After we uploaded our initial preprint to GitHub on November 13, 2026, Li, Li, and Sun found a family of counterexamples to the weak AJ conjecture, namely marked Whitehead doubles of torus knots $T(r,s)$ with coprime parameters $2\le rq\ge1$ both odd \cite[\S5.1]{Le2006}. Place the two arcs in $B$ vertically, with endpoints $0,\infty$ and $0',\infty'$. Push the two arcs in $B'$ onto $S^2=\partial B$, so that they join $0$ to $0'$ and $\infty$ to $\infty'$. Removing open tube neighborhoods of the vertical arcs from $B$ gives the genus 2 handlebody $H\cong\Sigma\times I$, where $\Sigma$ is a disk with two holes. In Lê's box presentation, $B=[0,6]\times[0,2]\times[0,1]$, with spatial coordinates $(x,y,z)$. The points $0=(1,1,1)$ and $0'=(5,1,1)$ lie on the top face. The points $\infty=(1,1,0)$ and $\infty'=(5,1,0)$ lie directly below them. The vertical segments are the arcs of $K$ inside $B$. Lê's construction specifies the boundary arcs from $p,q$ \cite[\S5.3, Proposition~5.1]{Le2006}. \begin{wrapfigure}{L}{.312\textwidth} \centering \includegraphics[width=\linewidth]{assets/1a9150e2b8ea-basis_handlebody.png} \captionsetup{justification=raggedright,singlelinecheck=false} \caption{The loops $\alpha$, $\alpha'$, and $\gamma$.} \label{fig:basis-handlebody} \end{wrapfigure} For $4_1=\mathfrak b(5,3)$, Figure~\ref{fig:four-one-le-box} shows $H$ with the original knot retained as a reference. Its vertical segments run through the removed tubes. The remaining arcs lie on $S^2$. Let $F:=\mathbb Q(t)$. The skein module of $H$ is $F[\alpha,\alpha',\gamma]$. The loops $\alpha,\alpha'$ surround the individual holes and $\gamma$ surrounds both. Figure~\ref{fig:basis-handlebody} shows them in a middle slice $\Sigma\times\{\tfrac12\}$. To recover the knot exterior $M=S^3\setminus N(K)$, we must attach one 2-handle to $H$. Its location is determined by the boundary arc $\ArcSeg{00'}$ of $K$. Take a narrow neighborhood of this arc in the original sphere $S^2$, including its endpoints, and call its boundary $\omega$. After the tubes have been removed, $\omega$ lies on $\partial H$. Let $D$ be a disk. The 2-handle is $D\times I$, whose side annulus $\partial D\times I$ is glued to an annular neighborhood of $\omega$. We identify $D$ with the core disk $D\times\{\tfrac12\}$ and depict it as the ribbon bounded by $\omega$ in Figure~\ref{fig:knot-slide-closures} \cite[\S\S5.2--5.4]{Le2006}. \begin{wrapfigure}{R}{.47\textwidth} \centering \includegraphics[width=\linewidth]{assets/32e76b8112e2-knot-slide-initial.png} \captionsetup{justification=raggedright,singlelinecheck=false} \caption{The core disk $D$ depicted as a ribbon in the trefoil box. The blue arc $\ArcSeg{00'}$ runs along its center. The left slide arc is yellow. The teal arc $\ArcSeg{0'0}$ returns through $\infty'$ and $\infty$. Juxtaposition denotes concatenation, so $\ArcSeg{00'}\,\ArcSeg{0'0}=K$.} \label{fig:knot-slide-closures} \end{wrapfigure} \paragraph{The slide relations.}\label{sec:slide-relations} Lê uses four relative arcs, all with endpoints $P,Q'$. The points $P,Q,Q',P'$ lie on $\omega$ at $x=3$ on the top face $z=1$. The direction up the page in Figure~\ref{fig:slide-arc} is the $y$-direction. We write $\ArcSeg{AB}$ for the depicted unoriented segment with endpoints $A,B$. A longer point word lists successive marked points without repeating junctions. The segment $\ArcSeg{PQ}$ follows $\omega$ around $0$, and $\ArcSeg{P'Q'}$ follows $\omega$ around $0'$. The other segments lie inside $H$, below the top face. The four arcs are $a_1:=\ArcSeg{PQ'}$, $a_2:=\ArcSeg{PQQ'}$, $a_3:=\ArcSeg{PP'Q'}$, and $a_4:=\ArcSeg{PQP'Q'}$. Lê chooses framing vectors normal to the boundary sphere $\partial B$ on these arcs and on the closing arcs of $\omega\subset\partial B$. We use this convention, extending the vectors over the small inward push into $H$. When joining the arcs, we match the endpoint framings and carry the resulting framing through the disk slide \cite[\S\S5.5--5.6.2]{Le2006}. \WFclear \begin{wrapfigure}{l}{.40\textwidth} \centering \resizebox{\linewidth}{!}{\input{assets/7a935ded1da8-slide_arc_strip.tex}} \captionsetup{skip=3pt} \caption{Top view of the handlebody.} \label{fig:slide-arc} \end{wrapfigure} We use $a_1$, shown in blue in Figure~\ref{fig:slide-arc}. Its endpoints divide $\omega$ into two arcs. Call the one passing around $0$ by $\omega_l$ and the other by $\omega_r$. The visible portions of $\omega_l$ include the turns around both $0$ and $0'$ and are purple. The orange portions belong to $\omega_r$. They leave the strip to the right of $P$ and to the left of $Q'$. The omitted portions continue outside the strip. Closing $a_1$ along $\omega_l$ gives $\mathcal L_0$. Closing it along $\omega_r$ gives $\mathcal R_0$. In $M$, the closing arc can move across $D$ from $\omega_l$ to $\omega_r$, while $a_1$ stays fixed. Thus $\mathcal L_0$ and $\mathcal R_0$ are isotopic as framed links in $M$, and $\langle \mathcal R_0\rangle=\langle \mathcal L_0\rangle$ in $\mathcal S_t(M)$. Their skein polynomials in $H$ can nevertheless be different. This difference is the slide relation we compute. \WFclear \noindent\begin{minipage}[t]{\textwidth} \setlength{\parindent}{15pt} For $k\ge0$, place $k$ parallel copies of $\gamma$ in a fixed interior collar, disjoint from the boundary collar in which the slide occurs. The link $\mathcal R_k$ is $a_1$ closed along $\omega_r$, together with these copies of $\gamma$. The link $\mathcal L_k$ uses $\omega_l$ and the same copies in the same positions. The same disk move shows $\langle \mathcal R_k\rangle=\langle \mathcal L_k\rangle$ in $\mathcal S_t(M)$. Figure~\ref{fig:slide-stills} shows three stages of this move. The box keeps the common tangle, including the copies of $\gamma$, unchanged throughout. \par\medskip \centering \input{assets/e0507e60c1dd-slide_stills.tex} \captionof{figure}{A slide across the attached disk $D$, shown in three stills. The closing arc moves from $\omega_l$ (purple) across $D$ to $\omega_r$ (orange). The boxed tangle is fixed. The first and last links are $\mathcal L_k$ and $\mathcal R_k$.} \label{fig:slide-stills} \end{minipage}\par\medskip % Source unit: peripheral/diagram-polynomials \subsection{From diagrams to exact polynomial relations}\label{sec:diagram-polynomials} Let $D\subset H$ be a framed link. Represent $D$ by a planar diagram on the disk with two holes. Its skein is $P=\langle D\rangle\in\mathcal S_t(H)=F[\alpha,\alpha',\gamma]$. A \textbf{state} is a choice of one of the two smoothings at every crossing. This leaves disjoint simple circles. Remove each contractible circle with a factor $-t^2-t^{-2}$, and assign $\alpha$, $\alpha'$, or $\gamma$ to each remaining circle according to which holes it surrounds. Each smoothing contributes $t$ or $t^{-1}$. The contribution of a state is the product of its smoothing factors, circle factors, and loop variables. The \textbf{state sum} adds these contributions over all choices of smoothings. It uses only finitely many additions and multiplications of integer Laurent polynomials, so $P\in\mathbb Z[t^{\pm1}][\alpha,\alpha',\gamma]$. This applies in particular to $\langle \mathcal L_k\rangle$ and $\langle \mathcal R_k\rangle$. We compute $\langle \mathcal L_k\rangle$ and $\langle \mathcal R_k\rangle$ using the recoupling rules of Appendix~\ref{app:tl}. We apply these rules to diagrams on the disk with two holes, with each local move supported in a disk disjoint from the holes. The crossing, fusion, and bubble identities are unchanged. The circle relation removes only a circle bounding such a disk, with factor $\delta=-t^2-t^{-2}$. A circle surrounding either hole or both is an \textbf{essential loop}. It cannot be contracted to a point within the disk with two holes. Such a loop does not bound a disk in this surface, so the circle relation does not remove it. \begin{wrapfigure}{R}{.43\textwidth} \centering \resizebox{\linewidth}{!}{\input{assets/b8b491bb243f-degree_cut_arcs.tex}} \caption{The arcs $a_0$ and $a_{0'}$ (red) in a middle slice of $H\cong\Sigma\times I$. The displayed $\gamma$-loop meets each arc once.} \label{fig:degree-cut-arcs} \end{wrapfigure} To use the TL calculus, fill the two removed tubes back into the ball $B$ and mark their cores by fixed strands. The link remains disjoint from these strands. A circle surrounding a hole now encircles the corresponding core, so keeping the cores marked retains the obstruction to contracting that circle. This represents the handlebody diagram as an ordinary tangle with two additional strands whose endpoints remain fixed. Write $a\cdot D$ for the unsigned intersection number of an arc $a$ with a diagram $D$ in $\Sigma$. Choose disjoint arcs $a_0,a_{0'}$ from the boundaries of holes $0,0'$ to the outer boundary, transverse to $D$ and avoiding its crossings. Figure~\ref{fig:degree-cut-arcs} shows an example with $a_0\cdot\gamma=a_{0'}\cdot\gamma=1$, but the arcs need not meet $\gamma$ only once. \WFclear \Needspace{8\baselineskip} \begin{theorem}[Diagrammatic degree bound]\label{thm:diagram-degree-bound} For a diagram $D$ in $\Sigma$, \[ \deg_\gamma\langle D\rangle\le\min\{a_0\cdot D,a_{0'}\cdot D\}. \] The bound also holds after setting $\alpha'=\alpha$. \end{theorem} \begin{pf} Smooth crossings in disks disjoint from the arcs, so every state retains the same intersection counts. Each state circle representing $\gamma$ separates both inner boundary circles from the outer boundary and therefore meets each arc. Thus each state has at most $\min\{a_0\cdot D,a_{0'}\cdot D\}$ $\gamma$-circles. Removing contractible circles and summing the states cannot increase this degree, nor can the substitution $\alpha'=\alpha$. \end{pf} Color the two fixed core strands $a$ and $b$, and choose an admissible total color $c$ for the pair. The colors let us evaluate the tangle by the fusion rules of Appendix~\ref{app:tl}. Keep the auxiliary strands throughout the calculation and close only the strands belonging to the link. The resulting diagram is an endomorphism of $f_a\otimes f_b$. Its restriction to the $c$-channel is a scalar multiple of the identity. An uncolored circle encircling a strand colored $c$ acts by $\chi_c=-\bigl(t^{2c+2}+t^{-2c-2}\bigr)$. This is the encircling identity \eqref{eq:encircling-scalar}, proved in Appendix~\ref{app:tl}. In particular, $\chi_0=\delta$. Therefore, the scalar computed for $D$ in the $c$-channel is $P(\chi_a,\chi_b,\chi_c)$. A loop surrounding both holes encircles their fused $c$-strand, which explains the third argument. \begin{theorem}[Symbolic recovery]\label{thm:symbolic-recovery} Let $P=\langle D\rangle$ be the skein of a framed link $D\subset H$. With formal color variables $x=t^{2a+2}$, $x'=t^{2b+2}$, and $y=t^{2c+2}$, the characters $\chi_a=-(x+x^{-1})$, $\chi_b=-(x'+(x')^{-1})$, and $\chi_c=-(y+y^{-1})$ are Laurent polynomials fixed by inversion. Expressing the channel evaluation in these three characters and replacing them by $\alpha$, $\alpha'$, and $\gamma$ recovers $P$ uniquely. \end{theorem} \begin{pf} The state sum gives $P\in\mathbb Z[t^{\pm1}][\alpha,\alpha',\gamma]$, so its channel evaluation has the stated form and all intermediate recoupling denominators cancel. Let $\upsilon(z)=z^{-1}$ denote inversion. The Laurent polynomial subring invariant under inversion is $F[z^{\pm1}]^\upsilon=F[z+z^{-1}]$. Applying the preceding identity to each variable proves uniqueness. \end{pf} Put $d=(p+1)/2$. Attaching the 2-handle identifies $\alpha'$ with $\alpha$, so define \begin{equation}\label{eq:slide-closures} S_{d+k}:=\left.\bigl(\langle \mathcal R_k\rangle-\langle \mathcal L_k\rangle\bigr) \right|_{\alpha'=\alpha}\in F[\alpha,\gamma]. \end{equation} Lê proves that $\alpha'=\alpha$ and the relations $S_m=0$, $m\ge d$, generate the kernel of $\mathcal S_t(H)\to\mathcal S_t(M)$. Moreover, $S_m$ has degree $m$ in $\gamma$ and can be written $S_m=s_m\gamma^m+\sum_{i=0}^{m-1}s_{m,i}(\alpha)\gamma^i$. Here $s_m$ is the leading coefficient and $s_{m,i}(\alpha)\in F[\alpha]$ is the coefficient of $\gamma^i$. The fact that $s_m$ is a signed power of $t$ comes from the braid closures. Lê's box diagram represents $\mathcal R_k$ and $\mathcal L_k$ as closures of braids on $2m=(p+1)+2k$ and $2(m-1)=(p-1)+2k$ strands, respectively, where $m=d+k$. For a closure on $2m$ strands, exactly one state has $m$ circles surrounding both holes. This state uses the horizontal smoothing at every crossing, which keeps the strands running across the braid. Its coefficient is a product of $t$ and $t^{-1}$, and every other state has lower $\gamma$-degree. This state sum uses the blackboard framing of the braid diagram. To apply it to $\mathcal R_k$ and $\mathcal L_k$, we compare that framing with Lê's chosen framings, whose vectors are normal to the boundary sphere $\partial B$ containing $\omega$, and which are carried through the slide. Different choices when matching the endpoint framings can introduce twists, but any difference from the braid diagram's blackboard framing is represented by curls \cite[\S5.6.2]{Le2006}. One smoothing of a curl produces a contractible circle, and the other straightens the strand. The two curl types therefore multiply the skein by $t\delta+t^{-1}=-t^3$ and $t^{-1}\delta+t=-t^{-3}$, respectively, as shown in Figure~\ref{fig:framing-curl}. \noindent\begin{minipage}[t]{\textwidth} \setlength{\parindent}{15pt} Thus any required framing correction multiplies the leading coefficient by a power of $-t^3$. It remains $\pm t^\ell$ for some integer $\ell$ \cite[Lemma~1.1]{Le2006}. Thus $\langle \mathcal R_k\rangle$ has $\gamma$-degree $m$, while $\langle \mathcal L_k\rangle$ has $\gamma$-degree $m-1$. Subtracting the latter cannot cancel the highest term of the former. The substitution $\alpha'=\alpha$ leaves that coefficient unchanged because it contains neither puncture variable \cite[\S5.7]{Le2006}. Hence $s_m=\pm t^\ell$ is invertible, and the relation replaces $\gamma^m$ by lower powers of $\gamma$. We explain below why these relations suffice. \par\medskip \centering \input{assets/f384e2644aa4-framing_curl.tex} \captionof{figure}{Resolving a framing curl gives $t\delta+t^{-1}=-t^3$.} \label{fig:framing-curl} \end{minipage}\par\medskip The four arcs $a_1,a_2,a_3,a_4$ from $P$ to $Q'$, defined above, use neither boundary detour, the detour around $0$, the detour around $0'$, or both detours, respectively. The endpoint framings are normal to the boundary. Resolving crossings expresses every relative skein with these endpoints as a linear combination of the four arcs with additional closed components. For each arc $a_i$, the two complementary arcs of $\omega$ give two closures. Lê denotes their difference by $\operatorname{sl}(a_i)$. The kernel of $\mathcal S_t(H)\to\mathcal S_t(M)$ is spanned by these differences and those formed after adding closed components to the relative tangles. These are relations of type $i$, where $i$ specifies the arc. Lê's Lemmas 5.4--5.6 reduce these four families to the slides of $a_1$ together with $\alpha'=\alpha$. After making that substitution, the kernel is generated over $F[\alpha]$ by the slides of $a_1$ with $k$ parallel copies of $\gamma$, for $k\ge0$. The generator with $k$ copies is $S_{d+k}$, up to sign, by \eqref{eq:slide-closures}. Setting $m=d+k$ identifies these generators with the family $S_m$, $m\ge d$, stated above. The degree calculation above gives $\deg_\gamma S_{d+k}=d+k$, with an invertible leading coefficient. Successive reduction therefore leaves degree less than $d$. Conversely, a nonzero finite $F[\alpha]$-linear combination of these slides has degree at least $d$. The largest degree present occurs in exactly one slide and has a nonzero leading coefficient. Thus no nonzero combination of $1,\gamma,\ldots,\gamma^{d-1}$ vanishes in $\mathcal S_t(M)$, proving the basis theorem. Starting with the highest power and repeating this reduction gives coordinates in $1,\gamma,\ldots,\gamma^{d-1}$. These coordinates are unique by Lê's basis theorem. \begin{theorem}[{Lê~\cite[Theorem~2]{Le2006}}]\label{thm:le-basis} Let $K=\mathfrak b(p,q)$ be a 2-bridge knot, let $M=S^3\setminus N(K)$ be its exterior, and put $d=(p+1)/2$. Then the skein module $\mathcal S_t(M)$ is free over $F[\alpha]$ with basis $1,\gamma,\ldots,\gamma^{d-1}$. \end{theorem} For $m\ge d$, set $\widetilde s_{m,i}(\alpha):=s_{m,i}(\alpha)/s_m$. We define the \textbf{slide relations} by \begin{equation}\label{eq:normalized-slide-relations} \widetilde S_m :=\gamma^m+\sum_{i=0}^{m-1}\widetilde s_{m,i}(\alpha)\gamma^i. \end{equation} All coefficients still lie in $F[\alpha]$, since $s_m$ is a signed power of $t$. \Needspace{4\baselineskip} \begin{theorem}[Reduction]\label{thm:reduction} Every polynomial $P\in F[\alpha,\gamma]$ has a unique representative $P^\sharp\in F[\alpha,\gamma]$ of $\gamma$-degree less than $d$ in $\mathcal S_t(M)$. There are finitely many non-zero polynomials $h_m(\alpha)\in F[\alpha]$ such that \begin{equation}\label{eq:slide-polynomial-identity} P=P^\sharp+\sum_{m\ge d}h_m(\alpha)\widetilde S_m \qquad\text{in }F[\alpha,\gamma]. \end{equation} The coefficients of $P^\sharp$ are the unique coordinates of the class of $P$ in Lê's basis. \end{theorem} \begin{pf} If $P$ has highest term $b(\alpha)\gamma^m$ with $m\ge d$, replace $P$ by $P-b(\alpha)\widetilde S_m$. The highest term cancels exactly. Repeating in decreasing degree terminates at a polynomial $P^\sharp$ of $\gamma$-degree less than $d$. Collecting the subtracted terms gives \eqref{eq:slide-polynomial-identity}. Each degree $m$ uses its own evaluated slide relation $\widetilde S_m$. The slide relations vanish in $\mathcal S_t(M)$, so $P$ and $P^\sharp$ represent the same element. Uniqueness follows from Theorem~\ref{thm:le-basis}. \end{pf} % Source unit: peripheral/peripheral-matrices \subsection{Constructing the peripheral matrices}\label{sec:peripheral-matrices} Let $\lambda\subset\partial M$ be the parallel copy of $K$ determined by its zero framing, and let $\mu$ be the meridian. We equip these curves with the framings induced by $\partial M$. The notation $\lambda+\mu$ denotes the simple boundary curve with that homology class. We compute the actions of $\lambda$ and $\lambda+\mu$ on $\mathcal S_t(M)$ as matrices $U,V\in\operatorname{Mat}_d(F[\alpha])$. Their columns in Lê's basis satisfy \begin{equation}\label{eq:peripheral-columns} \langle\lambda+\gamma^j\rangle^\sharp=\sum_{i=0}^{d-1}U_{ij}(\alpha)\gamma^i, \qquad \langle(\lambda+\mu)+\gamma^j\rangle^\sharp=\sum_{i=0}^{d-1}V_{ij}(\alpha)\gamma^i. \end{equation} The meridian alone gives $\langle\mu+\gamma^j\rangle^\sharp=\alpha\gamma^j$, so its matrix is just $\alpha I_d$. The mixed slope $\lambda+\mu$ supplies the additional information needed to separate the two longitude orientations and recover the forward action in Section~\ref{sec:recurrence}. % Source unit: peripheral/peripheral-matrices/cut-evaluation \subsubsection{Left-right recombination}\label{sec:cut-evaluation} A \textbf{regular cut} in a Morse diagram is a slice transverse to the strands which contains no crossing, cup, or cap. Let $X$ be the ordered tensor product of the boundary projectors of a colored tangle $D\colon X\to X$, and let $W$ be the corresponding tensor product at the cut. The two halves are $L\colon X\to W$ and $R\colon W\to X$, with $D=R\circ L$. \begin{theorem}[Left-right recombination]\label{thm:cut-recombination} At a regular cut of a colored framed Morse tangle $D=R\circ L\colon X\to X$, let $i_{c,k}\colon f_c\to W$ be labelled fusion trees with $i_{c,k}^*i_{d,\ell}=\delta_{cd}\delta_{k\ell}n_{c,k}f_c$. Write $L=\sum_{c,k}i_{c,k}\circ L_{c,k}$ and $R^*=\sum_{c,k}i_{c,k}\circ R_{c,k}$, where $L_{c,k},R_{c,k}\colon X\to f_c$. For a half $A=L$ or $R$, choose fusion trees $i'_{c,\eta}\colon f_c\to W$, $\eta\in\mathbf L_c$, and define $C_c$ by $i_{c,k}=\sum_{\eta\in\mathbf L_c}(C_c)_{\eta k}i'_{c,\eta}$, using successive applications of \eqref{eq:recoupling}. An insertion acting on coefficient $\eta$ by a scalar $\upsilon_{c,\eta}$ changes $A_c=(A_{c,k})$ to $\widetilde A_c=C_c^{-1}\operatorname{diag} (\upsilon_{c,\eta})C_cA_c$. The modified tangle is $\widetilde D=\sum_{c,k}n_{c,k}\widetilde R_{c,k}^*\circ\widetilde L_{c,k}$, with the unaffected half unchanged. The sums run over admissible total colors and internal labels. Without scalar insertions, the changes of basis cancel. \end{theorem} \begin{pf} The fusion resolution of $\mathbb{1}_W$ in Appendix~\ref{app:tl} gives the expansions. Recoupling takes either coefficient vector to $C_cA_c$; scalar multiplication in that tree and conversion back give the stated conjugation. Substituting the modified expansions into $\widetilde R\circ\widetilde L$ and applying orthogonality gives the recombination formula, with the original norms $n_{c,k}$. \end{pf} For a closed link, $X=f_0$ and only $c=0$ contributes. The coefficients are then scalars, and reflection fixes scalars in $F$, so the formula reduces to $\langle D\rangle=\sum_\ell L_{0,\ell}R_{0,\ell}n_{0,\ell}$. Evaluate $L$ and $R^*$ by the Morse-word algorithm of Appendix~\ref{app:tl}, reading the right half backwards with cups and caps interchanged. Express their outputs in the same fusion-tree basis at the cut, using recoupling when necessary. % Source unit: peripheral/peripheral-matrices/algebraic-insertions \subsubsection{Algebraic insertions}\label{sec:algebraic-insertions}\label{sec:gamma-insertion} Write the base longitude diagram as $R\circ L$ at a regular cut. The ribbon between $K$ and its parallel $\lambda$ lets us move the extra meridional turn to the cut. To compute both insertions from the same halves, use the cut constructed in Theorem~\ref{thm:box-insertion-cut}. \begin{theorem}[Algebraic meridian insertion]\label{thm:meridian-insertion} Use the notation and coefficient maps of Theorem~\ref{thm:cut-recombination} for $D=R\circ L$. For two adjacent strands of colors $a,b$ at the cut, let $\tau=B_{b,a}\circ B_{a,b}$ be their full twist, acting as the identity on the remaining strands. Choose a fusion tree that first joins this pair, and let $C_c$ and $\mathbf L_c$ be as in Theorem~\ref{thm:cut-recombination}. Write $e(\eta)$ for the label on the output edge of the subtree with leaves $a,b$. For $\epsilon=\pm1$, compute $R\circ\tau^\epsilon\circ L$ by Theorem~\ref{thm:cut-recombination} with $A=L$ and $\upsilon_{c,\eta}=(R^{ab}_{e(\eta)}R^{ba}_{e(\eta)})^\epsilon$. For the extra meridional turn, the pair has colors $a,1$ and $\epsilon$ is fixed by the crossing convention. \end{theorem} \noindent\begin{minipage}[c]{.34\linewidth} \centering \begin{tikzpicture}[x=.85cm,y=.72cm,line width=.9pt,line cap=round] \draw (0,0) node[below] {$e$} -- (0,.45); \draw (-.55,1.05) -- (0,.45) -- (.55,1.05); \fill (0,.45) circle (1.4pt); \foreach \h/\braid in {1.05/{B_{a,b}},2.05/{B_{b,a}}}{ \begin{scope}[shift={(0,\h)}] \draw (.55,0) .. controls (.55,.14) and (.32,.31) .. (.09,.45); \draw (-.09,.55) .. controls (-.32,.69) and (-.55,.86) .. (-.55,1); \draw (-.55,0) .. controls (-.55,.3) and (.55,.7) .. (.55,1); \node[left] at (-.65,.5) {$\braid$}; \end{scope} } \node[above] at (-.55,3.05) {$a$}; \node[above] at (.55,3.05) {$b$}; \node at (2.1,1.55) {$=\;R^{ab}_e R^{ba}_e$}; \begin{scope}[shift={(4.2,0)}] \draw (0,0) node[below] {$e$} -- (0,.45); \draw (-.55,3.05) node[above] {$a$} -- (-.55,1.05) -- (0,.45) -- (.55,1.05) -- (.55,3.05) node[above] {$b$}; \fill (0,.45) circle (1.4pt); \end{scope} \end{tikzpicture} \captionof{figure}{The full twist $\tau=B_{b,a}\circ B_{a,b}$ on the fusion tree $i_{e,a,b}$, read from bottom to top.} \label{fig:meridian-full-twist} \end{minipage}\hfill \begin{minipage}[c]{.63\linewidth} \begin{pf} The ribbon between $K$ and its parallel $\lambda$ moves the meridional turn to the cut as a full twist of the paired strands. On $i'_{c,\eta}$ the two crossings contribute $R^{ab}_{e(\eta)}$ and $R^{ba}_{e(\eta)}$, whose values are given in Appendix~\ref{app:tl}, \eqref{eq:colored-crossing}. The inverse twist contributes the reciprocal of their product. \end{pf} We can prescribe geometrically the placement of loops in $S^3$ within a given linking configuration with our knot, link, or tangle. We may then need to evaluate the resulting skeins under the pairing, including pairings of their cablings. The following theorem provides a way to compute these loop insertions algebraically. \end{minipage}\par \begin{theorem}[Algebraic $\gamma^j$ insertion]\label{thm:gamma-insertion} Use the notation and coefficient maps of Theorem~\ref{thm:cut-recombination} for $D=R\circ L$. Suppose the strands piercing the disk bounded by $\gamma$ form a consecutive block at the regular cut. Write $a_1,\ldots,a_m$ for their colors in order. Choose a fusion tree that fuses these strands before joining them to the remaining strands, and let $C_c$ and $\mathbf L_c$ be as in Theorem~\ref{thm:cut-recombination}. For $\ell\in\mathbf L_c$, define $e(\ell)$ to be the label on the output edge of the subtree with leaves $a_1,\ldots,a_m$. For $j$ parallel copies of $\gamma$ on the right, compute $D+\gamma^j$ by Theorem~\ref{thm:cut-recombination} with $A=R$ and $\upsilon_{c,\eta}=\chi_{e(\eta)}^j$. \end{theorem} \begin{pf} Draw the successive fusions of the enclosed strands inside the disk bounded by $\gamma$. Moving its boundary past these fusion vertices leaves a loop around the outgoing cable $f_{e(\eta)}$. The encircling identity \eqref{eq:encircling-scalar} multiplies the coefficient in row $\eta$ by $\chi_{e(\eta)}$ for each copy, giving $\chi_{e(\eta)}^j=\bigl[-(t^{2e(\eta)+2}+t^{-2e(\eta)-2})\bigr]^j$ for $j$ copies. \end{pf} From Theorem~\ref{thm:gamma-insertion} it is not clear that such a Morse word exists for a given diagram. However, in the case of 2-bridge knots, Lê's box presentation gives us a clear-cut way to find one. \begin{theorem}[An insertion cut in Lê's box]\label{thm:box-insertion-cut} For the longitude and for each of the two slide closures in Lê's box, a side projection admits a regular cut satisfying the geometric hypothesis of Theorem~\ref{thm:gamma-insertion} for the given loop $\gamma$. \end{theorem} \begin{pf} Use the horizontal slice through $\gamma$ and view the box along the $x$-axis. The puncture strands, and their nearby parallel copies in the longitude diagram, meet this slice inside the disk bounded by $\gamma$ near $y=1$. The other intersections lie on the side faces $y=0,2$. For the slide closures, the attaching arc follows these side faces, and the common closing arc lies above the slice. Thus the enclosed intersections are separated from the others by a gap on each side in the projection. Rotate the view slightly about the $z$-axis. The horizontal coordinate becomes $y\cos\theta+x\sin\theta$. A sufficiently small generic $\theta$ separates the finitely many cut points while preserving both gaps. The enclosed strands therefore form a consecutive block. Choose a nearby regular height within which the strands remain disjoint and have the same projected order, and move $\gamma$ to that height with its disk framing. Sweeping the resulting diagram gives the required colored Morse word and cut. \end{pf} Fix puncture color $a$ and total color $c$, as in Section~\ref{sec:diagram-polynomials}, and put $\iota=i_{c,a,a}$. Define scalars $L_\ell,R_\ell\in F$ by $L_{c,\ell}\circ\iota=L_\ell f_c$ and $R_{c,\ell}\circ\iota=R_\ell f_c$. Write $n_\ell=n_{c,\ell}$. The channel inclusion has norm $\iota^*\iota=\nu_{a,c}f_c$, where $\nu_{a,c}=\theta(a,a,c)/\Delta_c$. For the base diagram $D=R\circ L$, insertion and left--right recombination give \[ \langle D+\gamma^j\rangle(\chi_a,\chi_c) =\frac{1}{\nu_{a,c}}\sum_k n_k L_k\sum_{\eta,\ell}(C_c^{-1})_{k\eta}\chi_{e(\eta)}^j(C_c)_{\eta\ell}R_\ell. \] \begin{wrapfigure}{r}{.45\textwidth} \centering \resizebox{\linewidth}{!}{\begin{tikzpicture}[scale=.85] \draw[dashed] (-.5,1.5)--(3.5,1.5); \draw[orange,thick] (1.5,1.5) ellipse (2.1 and .35); \draw[thick,preaction={draw=white,line width=3pt}] (0,0)--(0,3) (1,0)--(1,3) (2,0)--(2,3) (3,0)--(3,3); \draw[orange,thick,preaction={draw=white,line width=3pt}] (-.6,1.5) arc[start angle=180,end angle=360,x radius=2.1,y radius=.35]; \node at (1.5,-.4) {strands enclosed by $\gamma$}; \node at (5,1.5) {$\longmapsto$}; \draw[orange,thick] (7.5,1.55) ellipse (.6 and .25); \draw[thick,preaction={draw=white,line width=3pt}] (6,0)--(7.5,1.1)--(9,0) (7.5,1.1)--(7.5,2)--(6,3) (7.5,2)--(9,3); \draw[orange,thick,preaction={draw=white,line width=3pt}] (6.9,1.55) arc[start angle=180,end angle=360,x radius=.6,y radius=.25]; \node[anchor=west] at (8.22,1.55) {$e(\ell)$}; \node at (7.5,-.4) {$j$ loops multiply by $\chi_{e(\ell)}^j$}; \end{tikzpicture}} \caption{The algebraic basis insertion. Fuse the strands enclosed by $\gamma$ once, then obtain every power of $\gamma$ by scalar multiplication on the right-half coefficients.} \end{wrapfigure} Let $P_j^\lambda(\alpha,\gamma)$ and $P_j^{\lambda+\mu}(\alpha,\gamma)$ be the skeins of $\lambda+\gamma^j$ and $(\lambda+\mu)+\gamma^j$, respectively, in the diagram's framing, after setting $\alpha'=\alpha$. Write $L=(L_k)$ and $R=(R_k)$ in the basis at the cut. For multiple insertions on nested or disjoint blocks, we can use a common fusion tree. Here choose the tree that first fuses the meridional pair of colors $a,1$ and then the block enclosed by $\gamma$. Let $C_c$ be the change-of-basis matrix of Theorem~\ref{thm:cut-recombination} from the basis at the cut to this tree, at fixed total color $c$. Write $e_\mu(\eta)$ and $e_\gamma(\eta)$ for the output colors of the pair and the block, respectively. The corresponding diagonal matrices are $D_\gamma=\operatorname{diag}(\chi_{e_\gamma(\eta)})$ and $D_\mu=\operatorname{diag}(R^{a1}_{e_\mu(\eta)}R^{1a}_{e_\mu(\eta)})$. Applying Theorems~\ref{thm:meridian-insertion} and~\ref{thm:gamma-insertion} and recombining in the basis at the cut gives \begin{equation}\label{eq:cut-weights} \bigl(P_j^\lambda(\chi_a,\chi_c),P_j^{\lambda+\mu}(\chi_a,\chi_c)\bigr) =\frac1{\nu_{a,c}}\sum_k n_k \bigl(C_c^{-1}D_\gamma^j C_c R\bigr)_k \Bigl(L_k,\bigl(C_c^{-1}D_\mu^\epsilon C_c L\bigr)_k\Bigr). \end{equation} \Needspace{4\baselineskip} \begin{theorem}[Peripheral matrix construction]\label{thm:peripheral-matrices} The following finite procedure computes $U,V$. \begin{enumerate} \setcounter{enumi}{-1} \item Let $N:=\max_{0\le j0$, satisfies $|u|,v<\sqrt{N/2}$, Wang's rational reconstruction theorem recovers it exactly \cite{WangGuyDavenport1982}. Forming the polynomials from the recovered rational numbers gives $f_j\in\mathbb Q[q,Q]\subset\mathbb K$. We verify $\sum_{j=0}^s f_jv_j(I)=0$ for every $I$ by clearing denominators and comparing polynomial coefficients. Let $\varphi_1,\ldots,\varphi_k$ solve $\widehat P$ and let $W$ be their Casoratian. Right division gives $E^h\varphi_l=\sum_{b=0}^{r-1}u_{h,b}E^b\varphi_l$. By Cauchy--Binet, $E^jW=\sum_I v_j(I)D_I$, where $D_I$ is the minor with rows $I$ of $(E^b\varphi_l)_{0\le b1$ and $0\le c0$, $\gcd(v,N)=1$, $|u|,v<\sqrt{N/2}$, and $uv^{-1}\equiv c\pmod N$. If such a fraction exists, rational reconstruction by the extended Euclidean algorithm recovers it. \end{theorem} After clearing denominators, we obtain an operator with coefficients in $\mathbb Z[q^{\pm1},Q]$. We substitute it into all the independently computed Laurent-polynomial windows. This tests the reconstructed relation exactly on the available values. Specialization keeps the linear systems small. At each $g$, there is one unknown for each index $i\in I$, regardless of the number of powers of $q$ in its coefficient. Only after these systems have been solved do we recover the powers of $q$ by simultaneous interpolation. For $7_4$, we obtain the order-three candidate $\widehat H=h_0+h_1E+h_2E^2+h_3E^3$, where \begin{align*} h_0={}&q^2Q(1-qQ)^2(1-q^2Q) \bigl(1-q^3Q+q^5(1+q)Q^2\bigr),\\ h_1={}&(1-q^2Q)\bigl(1-q^2Q+q^3(1+q)Q^2\bigr)\\ &\quad\cdot\bigl(1-q^2(1+q)Q +q^3(1+q+2q^2+q^3)Q^2 -q^6(1+q+q^2)Q^3+q^8(1+q+q^2)Q^4\bigr),\\ h_2={}&-q^4Q^2\bigl(1-q^3Q+q^5(1+q)Q^2\bigr)\\ &\quad\cdot\bigl(1-q^2(1+q)Q +q^3(1+q+2q^2+q^3)Q^2 -q^6(1+q+q^2)Q^3+q^8(1+q+q^2)Q^4\bigr),\\ h_3={}&-q^{20}Q^7\bigl(1-q^2Q+q^3(1+q)Q^2\bigr). \end{align*} The expanded operator has $76$ nonzero monomials. % Source unit: discovery/the-discovery \subsection{Guessing \texorpdfstring{$\widehat A$}{A-hat}}\label{sec:discovery-pipeline} We compute an initial segment of $J$ by fusion and use \eqref{eq:habiro-inversion} to obtain $C$. The solver then produces a candidate $\widehat H=\sum_{i=0}^r h_i(q,Q)E^i$. At an index $k$ where $h_r(q,q^k)\ne0$, this gives the continuation \[ C(k+r)=-\frac{\sum_{i=0}^{r-1}h_i(q,q^k)C(k+i)}{h_r(q,q^k)}. \] If the leading coefficient vanishes, the recurrence does not determine that value, which must be computed independently. We use the recurrence to extend $C$ and transform back to Jones values by \eqref{eq:habiro-newton-expansion}. In the computation these steps are performed directly at the interpolation points, so extending the sequence does not require expanding every Jones polynomial. We apply the same solver to the extended Jones values, obtaining a candidate for $\widehat A$. For $8_5$, the resulting operator has order $10$, $Q$-degree $103$ and $97{,}522$ nonzero monomials after clearing denominators. Let $A_C$ be Culler's tabulated polynomial for this knot, which has $L$-degree $8$ \cite{CullerApolynomials}. Reversing the longitude and multiplying by $L^{10}$ to match Culler's convention shows $L^{10}\left.\varepsilon^*(\widehat A)\right|_{L\mapsto L^{-1}} \asymp(LM^{12}-1)(L-1)A_C$. We initially took $(L-1)A_C$ to be the full $A$-polynomial. Since $LM^{12}-1$ is not a factor of this polynomial, the calculation appeared to contradict weak AJ. However, the missing factor comes from the trefoil group. By work of Kitano and Suzuki~\cite[Theorem~1.1]{KitanoSuzuki2005}, there is a surjection $\pi_1(S^3\setminus8_5)\mathrel{\relbar\joinrel\twoheadrightarrow}\pi_1(S^3\setminus3_1)$. To determine its effect on the peripheral curves, we use the presentation \[ \pi_1(S^3\setminus8_5)=\left\langle a,b,c\ \middle|\ a^2ca^{-2}b^{-1}a^{-1}ca^2b,\quad abc^{-1}b^{-1}acb\right\rangle, \] with peripheral words \[ \mu=a^{-1}b^{-1}c^{-1}a^{-1},\qquad \lambda=aca^{-2}c^{-1}b^{-1}c^{-1}aba^2c^9b^{-1}. \] Put $\pi_1(S^3\setminus3_1)=\langle r,s\mid rsr=srs\rangle$ and $\zeta=(rs)^3$. The assignment \[ \phi(a)=(rs)^{-1},\qquad \phi(b)=(rs)^2,\qquad \phi(c)=r \] satisfies both relators. It is surjective since $r=\phi(c)$ and $s=\phi(c)^{-1}\phi(a)^{-1}$. Substitution into the peripheral words gives \[ \phi(\mu)=s^{-1}r^{-1}s,\qquad \phi(\lambda)=\phi(\mu)^{-12}\zeta^{-2}. \] For $M$ transcendental, set $\eta=M^2+M^{-2}-1$ and take the trefoil representation \[ \rho(r)=\begin{pmatrix}M&1\\0&M^{-1}\end{pmatrix},\qquad \rho(s)=\begin{pmatrix}M&0\\-\eta&M^{-1}\end{pmatrix}. \] These matrices satisfy the braid relation and $\rho(\zeta)=-I$. Their commutator has trace $2+\eta\ne2$, so the representation is irreducible. Its pullback along the surjection is also irreducible, and the peripheral calculation gives $(\rho\circ\phi)(\lambda)=(\rho\circ\phi)(\mu)^{-12}$. On a common eigenline with meridian eigenvalue $M$, the longitude eigenvalue is therefore $M^{-12}$. The peripheral image is Zariski dense in the curve $LM^{12}-1=0$, so $LM^{12}-1$ divides the $A$-polynomial of $8_5$, and $A_C$ is missing this factor. For $\varepsilon^*(\widehat A)=\sum_{j=0}^r a_j(M)L^j$, its content with respect to $L$ is $\mathfrak C(M)=\gcd_j a_j(M)$ in $\mathbb Q[M^{\pm1}]$. We choose its representative with primitive integer coefficients, positive leading coefficient and nonzero constant term. For $7_4$, the candidate has $3{,}364$ nonzero monomials. Its specialization has $150$ nonzero monomials, and its content is $\mathfrak C(M)=(M^4-1)^5\bigl(2M^8-5M^6+8M^4-5M^2+2\bigr)$. Our largest stored $\widehat A$ is for $8_{12}$, with order $15$, $Q$-degree $172$ and $927{,}668$ nonzero monomials. Its specialization has $2{,}424$ nonzero monomials, and its content has $129$. Needless to say, we won't display it. % Source unit: computerization \section{Computerization}\label{sec:computerization} I started researching how to compute $\widehat A$ some time in spring of 2026. Around that time I had the idea for the procedure in section~\ref{sec:discovery}, it's a natural thing you to do once you know a sequence is $q$-holonomic (or just holonomic, it works the same way) write an ansatz, and interpolate. I was inspired by Culler's calculation of the $A$-polynomial, he sampled roots of unity on a circle of radius about 1 on the $M$ line, and computed its fiber in $V(A)$ by choosing square roots on Thurston's gluing equations variety, and used fast Fourier transform to recover the coefficients of $A$. The $q$-holonomic solver was implemented in code with the use of agents pretty easily. I hit a computational bottleneck with Hajij--Levitt's code. But then I checked whether FGL's formula for the colored Jones polynomial \cite[\S4]{FrohmanGelcaLofaro2002} had been implemented by directly resolving the skein, and I could not find an implementation. After the agent had done it directly by smoothings, it suggested fusion as a faster alternative to the standard smoothings of skeins. Once we had working software that reproduced Hajij--Levitt, I asked it to formalize it in Lean, and it did it formalized all the required mathematical machinery that our code uses to produce the colored Jones function. This gave me unwarranted confidence in its abilities. We used our existing $q$-holonomic solver on the new sequence, and we were getting results. We found all the operators mentioned in the introduction, and the counterexample. However it was clear to me that having a reconstructed operator doesn't establish neither membership of the annihilator, nor minimality. So I asked it to prove both. I asked it to test the operator on our symbolic Jones function in Mathematica and look for simplifications. The computation crashed my computer. However, instead of using the provided operator coefficients and Jones sequence to finish or simplify that calculation. It began constructing auxiliary recurrence and telescoping systems. Whenever a calculation timed out, it changed the framework and introduced more auxiliary coordinates. The construction grew from a seven-variable sum to a 42-component system and eventually a 1,319-dimensional Lean state vector. To my shame this went on for way too long, I had it running on \texttt{/goal} mode, and it consumed in a few days the allowance for the whole of the 2 weeks. In my defense it was giving me periodic progress reports. I had been doing other unrelated research stuff, so I wasn't too worried. It wasn't until the second time it spent its whole weekly limit that I gave a close reading of the Lean. I almost fell off my chair, when I realized it was just lying to me. The vocabulary changed along with the detours. ``Source coefficient budgets,'' ``Habiro source budgets'' and ``generating towers'' described bookkeeping about coefficients and bounds while the required cancellation remained uncomputed. ``Flux algebra'' and ``numerical cutoff costs'' gave names to further constructions without producing anything. The terminology made it difficult to distinguish work on auxiliary machinery from progress on the identity I had asked it to prove. In a nutshell it was running in circles. But then I did some more research, and landed on Lê's work. Then I asked the agent if it had any ideas on how to compute $\widehat P$, and it said that we could do it with fusion. So it did for a few knots, and I asked it how it did it. It described the procedure, and I realized it had been doing it inefficiently so I suggested the insertion and recombination theorems, which is how $8_{11}$ was finally computed after about a day of intense calculation in a rented machine with 750 GB of RAM in Amazon Web Services, which I had to ask permission to use, which ended up costing \textdollar114.81 bucks. While this work would definitely have been impossible to finish without AI. This section highlights just the most relevant snippets, of the myriad of hurdles one encounters with agentic assistance. Only your training, and your intuition can protect you against such tomfoolery. More than an AI assisted paper, half the time I felt I was assisting the AI. \FloatBarrier \appendix % Source unit: tl \section{Temperley--Lieb evaluation}\label{sec:tl}\label{app:tl} Work over $F=\mathbb Q(t)$ and put $\delta=-t^2-t^{-2}$. The \textbf{Temperley--Lieb category} $\mathrm{TL}(\delta)$ has the nonnegative integers as objects. A planar matching joins $n$ points on the bottom edge of a rectangle and $m$ points on its top edge in pairs by disjoint arcs inside the rectangle. Every boundary point is the endpoint of exactly one arc, and the arcs do not cross. \input{assets/cc499d52af46-planar_matchings.tex} We identify matchings that deform into one another while keeping the boundary fixed. Such a deformation cannot change which boundary points are paired, and the arcs must remain disjoint throughout. \input{assets/ba9367b9e3f1-planar_deformation.tex} Denote the middle matching in Figure~\ref{fig:planar-matchings} by $d:3\to3$. A morphism from $n$ to $m$ is an $F$-linear combination of these matchings. The \textbf{Temperley--Lieb algebra} $\operatorname{End}(n)=\operatorname{Hom}_{\mathrm{TL}(\delta)}(n,n)$ has multiplication given by composition. For example, the first two matchings in Figure~\ref{fig:planar-matchings} give the morphism\footnote{$\mathbb{1}_o$ denotes the identity morphism of an object $o$.} \[ D=t\mathbb{1}_3+t^{-1}d\in\operatorname{End}(3). \] To compose $f:n\to m$ and $g:m\to r$, stack $g$ above $f$ and join their $m$ common boundary points. Remove each closed circle produced by this operation and multiply by $\delta$ for each circle removed. Figure~\ref{fig:tl-stacking} shows this for $d\circ d$. Composition extends bilinearly to all morphisms. \input{assets/2e02dbad5f64-tl_stacking.tex} The tensor product places one rectangle to the right of the other. Thus $f:n\to m$ and $g:r\to s$ give $f\otimes g:n+r\to m+s$. For example, placing a single identity strand to the left of $d\in\operatorname{End}(3)$ gives $\mathbb{1}_1\otimes d\in\operatorname{End}(4)$. \input{assets/a033047281c9-tl_tensor.tex} The morphism $\operatorname{cup}:0\to2$ joins its two output points, and $\operatorname{cap}:2\to0$ joins its two input points. \input{assets/32b7781cd58c-tl_cup_cap.tex} Number the boundary points on each horizontal edge from left to right. For $1\le i1). \] The relative Kauffman bracket identifies framed tangles with morphisms of this category. In particular, the two crossings are $t\mathbb{1}_n+t^{-1}e_i$ and $t^{-1}\mathbb{1}_n+te_i$ \cite{KauffmanLins1994}. Let $V=F\langle\omega_0,\omega_1\rangle$, with basis $\omega_0,\omega_1$. We evaluate diagrams by an $F$-linear tensor functor $\mathcal E:\mathrm{TL}(\delta)\to\operatorname{FinVect}(F)$. It sends the object $n$ to $V^{\otimes n}$, with $V^{\otimes0}=F$. A matching $D:n\to m$ therefore gives a linear map $\mathcal E(D):V^{\otimes n}\to V^{\otimes m}$. The images of cup and cap under $\mathcal E$ are the linear maps \[ \operatorname{Cup}=\mathcal E(\operatorname{cup}):F\longrightarrow V\otimes V, \qquad \operatorname{Cap}=\mathcal E(\operatorname{cap}):V\otimes V\longrightarrow F, \] defined by $\operatorname{Cup}(1)=t\,\omega_0\otimes\omega_1-t^{-1}\omega_1\otimes\omega_0$, $\operatorname{Cap}(\omega_0\otimes\omega_1)=-t$, $\operatorname{Cap}(\omega_1\otimes\omega_0)=t^{-1}$, and $\operatorname{Cap}(\omega_i\otimes\omega_i)=0$ for $i=0,1$. The circle relation gives $\operatorname{Cap}\operatorname{Cup}=\delta\mathbb{1}_F$. Straightening the bends in a single arc gives \[ (\operatorname{Cap}\otimes\mathbb{1}_V)(\mathbb{1}_V\otimes\operatorname{Cup}) =\mathbb{1}_V. \] Figure~\ref{fig:tl-straightening} shows this identity. \begin{figure}[htbp] \input{assets/f0e80752871d-tl_straightening.tex} \end{figure} Applying $\mathcal E$ to the definition of $e_i$ gives $\mathcal E(e_i)=\mathbb{1}_V^{\otimes(i-1)}\otimes(\operatorname{Cup}\operatorname{Cap}) \otimes\mathbb{1}_V^{\otimes(n-i-1)}$. % Source unit: tl/jones-wenzl-projectors \subsection{Jones--Wenzl projectors} Put \[ [a]=\frac{t^{2a}-t^{-2a}}{t^2-t^{-2}},\qquad \Delta_a=(-1)^a[a+1],\qquad [a]!=\prod_{i=1}^a[i]. \] The Jones--Wenzl projector $f_a\in\operatorname{End}(a)$ is the unique linear combination of planar matchings whose coefficient of $\mathbb{1}_a$ is one and which satisfies $e_i f_a=f_a e_i=0$ for $1\le i0$. Inserting the resolution of $\mathbb{1}_W$ between two diagrams gives matrix multiplication. The inverse sends this sequence to $\sum_{c,\ell,\ell'}D^{(c)}_{\ell,\ell'}n_{c,\ell'}^{-1}i_{c,\ell}i_{c,\ell'}^*$. Closure on either side gives $\sum_c\Delta_c\operatorname{tr}(D^{(c)})$. Thus cable evaluation and fusion evaluation give the same bracket. Changing a fusion tree changes the basis in this isomorphism. For three boundary colors $a,b,c$ and total color $d$, let $\alpha_e=(i_{e,a,b}\otimes\mathbb{1}_{f_c})\circ i_{d,e,c}$ and $\beta_f=(\mathbb{1}_{f_a}\otimes i_{f,b,c})\circ i_{d,a,f}$. The first tree joins $a,b$ through $e$, and the second joins $b,c$ through $f$. Both maps have domain $f_d$ and codomain $f_a\otimes f_b\otimes f_c$. The \textbf{$6j$ coefficients} are defined, in our normalization, by \begin{equation}\label{eq:recoupling} \alpha_e=\sum_f \begin{Bmatrix}a&b&f\\c&d&e\end{Bmatrix}\beta_f, \end{equation} where the sum runs over $f$ for which $(b,c,f)$ and $(a,f,d)$ are admissible \cite[\S2, p.~368]{MasbaumVogel1994}. The coefficient has six labels $a,b,c,d,e,f$, hence the name $6j$.\footnote{Our integer colors are $2j$. In quantum mechanics, $j$ labels the eigenvalue $\hbar^2j(j+1)$ of squared angular momentum. The $6j$ coefficients change the order in which three angular momenta are coupled.} \input{assets/7e3c8c00b92b-tl_recoupling.tex} The bubble identity gives $\beta_f^*\beta_f= \frac{\theta(b,c,f)\theta(a,d,f)}{\Delta_f\Delta_d}\mathbb{1}_{f_d}$. Pairing \eqref{eq:recoupling} with $\beta_f^*$ therefore gives \begin{equation}\label{eq:sixj-tetrahedron} \begin{Bmatrix}a&b&f\\c&d&e\end{Bmatrix} =\frac{\mathcal E(\operatorname{cl}(\beta_f^*\alpha_e))(1)} {\mathcal E(\operatorname{cl}(\beta_f^*\beta_f))(1)} =\frac{\Delta_f\left\langle\begin{matrix}a&b&e\\c&d&f\end{matrix}\right\rangle} {\theta(b,c,f)\theta(a,d,f)}. \end{equation} \begin{wrapfigure}{R}{.34\textwidth} \centering \input{assets/db356273acc5-tl_tetrahedron.tex} \end{wrapfigure} Following Masbaum--Vogel, let $\left\langle\begin{matrix}a&b&e\\c&d&f\end{matrix}\right\rangle =\mathcal E(\operatorname{cl}(\beta_f^*\alpha_e))(1)$ denote the tetrahedron coefficient. Closing this composite joins the two trees into the tetrahedral graph in Figure~\ref{fig:tl-tetrahedron}. For completeness, the tetrahedron formula \cite[Theorem~2, p.~364]{MasbaumVogel1994} is as follows. Set \[ \begin{aligned} u_1&=(a+b+e)/2,& u_2&=(b+c+f)/2,\\ u_3&=(c+d+e)/2,& u_4&=(a+d+f)/2,\\ v_1&=(a+b+c+d)/2,& v_2&=(a+c+e+f)/2,\\ v_3&=(b+d+e+f)/2. \end{aligned} \] \WFclear Then \begin{equation}\label{eq:tetrahedron-factorials} \left\langle\begin{matrix}a&b&e\\c&d&f\end{matrix}\right\rangle =\frac{\displaystyle\prod_{i=1}^4\prod_{j=1}^3[v_j-u_i]!} {[a]![b]![c]![d]![e]![f]!} \sum_{k=\max_i u_i}^{\min_j v_j} \frac{(-1)^k[k+1]!} {\displaystyle\prod_{i=1}^4[k-u_i]!\prod_{j=1}^3[v_j-k]!}. \end{equation} The four vertex triples are admissible, so all the $u_i,v_j$ and the factorial arguments in the stated range are nonnegative integers. We use $[0]!=1$. Fix the over/under convention by assigning $\sigma_i=t\mathbb{1}_n+t^{-1}e_i$ to one elementary crossing of adjacent strands. The opposite crossing is $\sigma_i^{-1}=t^{-1}\mathbb{1}_n+te_i$. The Temperley--Lieb relations and $\delta=-t^2-t^{-2}$ imply $\sigma_i\sigma_i^{-1}=\mathbb{1}_n$, $\sigma_i\sigma_{i+1}\sigma_i=\sigma_{i+1}\sigma_i\sigma_{i+1}$, and $\sigma_i\sigma_j=\sigma_j\sigma_i$ for $|i-j|>1$. Thus different decompositions of the same braid into adjacent crossings give the same morphism. To form a \textbf{crossing of colors $a,b$}, replace its two strands by cables of $a$ and $b$ parallel strands along their framings, using this same over/under convention at every elementary crossing. Let $b_{a,b}$ denote the resulting braid, which exchanges the two blocks and crosses each of the $a$ strands with each of the $b$ strands once. Attaching the projectors gives $B_{a,b}=(f_b\otimes f_a)b_{a,b}(f_a\otimes f_b) :f_a\otimes f_b\to f_b\otimes f_a$. This is independent of the chosen decomposition of the cabled braid by the relations above. The opposite block crossing gives its inverse. If $V_a=\operatorname{im}\mathcal E(f_a)$, its image under the functor is the linear map $\mathcal E(B_{a,b}):V_a\otimes V_b\to V_b\otimes V_a$. For an admissible triple $(a,b,c)$, the \textbf{$c$-channel} in $f_a\otimes f_b$ is the summand isomorphic to $f_c$ included by $i_{c,a,b}$. Its projector is $p_c^{a,b}=\frac{\Delta_c}{\theta(a,b,c)}i_{c,a,b}i_{c,a,b}^*$. We use the same name for its image under $\mathcal E$. Similarly, the $c$-channel in $V_b\otimes V_a$ is $\operatorname{im}\mathcal E(i_{c,b,a})$. The map $\mathcal E(B_{a,b})$ carries the $c$-channel in $V_a\otimes V_b$ to this subspace. At the level of diagrams, this is expressed by \begin{equation}\label{eq:colored-crossing} B_{a,b}\circ i_{c,a,b}=R^{ab}_c\,i_{c,b,a}, \qquad R^{ab}_c=(-1)^u t^{vw-u(u+v+w+2)}. \end{equation} Here $u,v,w$ are the admissibility integers. So we say that the crossing of colors $a,b$ acts by $R^{ab}_c$ on the $c$-channel because it sends the inclusion, by composition, to the exchanged inclusion multiplied by this scalar \cite[Theorem~3 and \S2]{MasbaumVogel1994}. For example, when $a=b=1$ and $c=0$, the inclusion is the cup, and \[ B_{1,1}\circ\operatorname{cup} =(t+t^{-1}\delta)\operatorname{cup} =-t^{-3}\operatorname{cup}, \] which is the scalar $R^{11}_0$ in \eqref{eq:colored-crossing}. \begin{figure}[htb] \centering \begin{minipage}{.43\textwidth} \centering \input{assets/b2a123dbc88e-tl_colored_crossing.tex} \end{minipage} \end{figure} % Source unit: tl/encircling-strand \subsection{Encircling a colored strand}\label{sec:encircling-strand} An uncolored circle means a circle colored $1$. Place such a circle around a straight strand colored $c$, with zero framing on the circle. This gives an endomorphism of $f_c$. Since $\operatorname{End}(f_c)=Ff_c$, it equals $\chi_c f_c$ for a scalar $\chi_c$. Close the $c$-strand. The resulting diagram is a Hopf link. Fuse the colors $1$ and $c$ before evaluating its two crossings. In channel $e$, the crossings contribute $R^{1c}_eR^{c1}_e$ and the closure contributes $\Delta_e$. Closing $\chi_c f_c$ gives $\chi_c\Delta_c$, so \[ \chi_c=\frac{1}{\Delta_c} \sum_e\Delta_e R^{1c}_eR^{c1}_e, \] where $e=c-1,c+1$, with the first channel omitted when $c=0$. For $c\ge1$, the crossing formula \eqref{eq:colored-crossing} gives $R^{1c}_{c+1}=R^{c1}_{c+1}=t^c$ and $R^{1c}_{c-1}=R^{c1}_{c-1}=-t^{-c-2}$. Substituting $\Delta_e=(-1)^e[e+1]$ therefore yields \begin{equation}\label{eq:encircling-scalar} \chi_c =\frac{\Delta_{c+1}t^{2c} +\Delta_{c-1}t^{-2c-4}}{\Delta_c} =-\bigl(t^{2c+2}+t^{-2c-2}\bigr) \qquad(c\ge1). \end{equation} For $c=0$, the sole channel $e=1$ gives $\chi_0=\Delta_1=\delta=-t^2-t^{-2}$, so the same final formula holds. If a collection of strands has first been fused to color $b$, an uncolored circle around that collection consequently acts by $\chi_b$. Adding $j$ parallel circles gives $\chi_b^j$. % Source unit: tl/morse-words \subsection{Morse words}\label{sec:morse-words} Choose a height direction in the diagram so that its local minima, local maxima, and crossings have distinct heights. Sweeping from bottom to top, record a cup at each minimum, a cap at each maximum, and a crossing whenever two strands cross. This ordered list is a \textbf{Morse word}. Number the current strands from left to right. Each entry specifies the two adjacent strand positions involved and, for a crossing, which strand passes over the other. A cup creates two strands, a cap removes two strands, and a crossing exchanges their positions. Stacking these elementary pieces in the recorded order reconstructs the diagram \cite{KnotAtlasMorseLink}. Since we read from left to right, we'll write the word from left to right. Write $\MorseCup{i}{j}$ and $\MorseCap{i}{j}$ for a cup and a cap on adjacent positions $i0$ strands have color $1$, write $\MorseColorBoxes{1_1^n}$. We write $\varnothing$ only when there are no strands. Together with the Morse letters and the initial strand count, these records determine the colored tangle. For the formulas below, $\mathbf a_k=(a_{k,1},\ldots,a_{k,m_k})$ denotes the expanded list of colors at the $k$th cut. We write the colored word from left to right, inserting its sparse color records between consecutive letters. At a cut with $m$ strands and expanded color list $\mathbf a$, put $a_i$ for its $i$th entry. The permitted letters are as follows. \begin{itemize} \item A cup $\MorseCup{i}{i+1}$ is permitted for $1\le i\le m+1$. If its color is $a$, it inserts $a,a$ at the new positions $i,i+1$, giving $m+2$ strands. \item A cap $\MorseCap{i}{i+1}$ is permitted for $1\le i