The new word stalk is the reflected pair lattice
Fix a finite-generator Coxeter system, a simple reflection \(s\), a tail word \(\mathbf t\), and \(x\in W\). Let \(A=\operatorname{Sym}_{\mathbb R}(V)\) for its real extended realization \(V\), and let \(\tau=\chi_s\) be the coefficient action. Put \[ X=M_{\mathbf t}^x,\qquad Y=M_{\mathbf t}^{sx},\qquad f=\pi_x,\quad g=\pi_{sx}. \] Let \(L_x\subseteq X\oplus{}^\tau Y\) be their reflected pair lattice.
The map \[ C_x:L_x\xrightarrow{\sim}M_{s\mathbf t}^x,\qquad (p,q)\longmapsto \begin{cases} p_\eta&\text{on the zero-bit block},\\ \tau(q_\eta)&\text{on the one-bit block} \end{cases} \] is an \(A\)-linear isomorphism. Under this identification, the head projections are \[ h_0:C_x(p,q)\mapsto p,\qquad h_1:C_x(p,q)\mapsto q. \] The first is \(A\)-linear and surjective. The second is \(\tau\)-semilinear, and therefore real-linear. Both preserve the degree-one grading.
Relation to the OpenAI notes
Word-image construction before the boundary sheaf.
OpenAI notes: The moment-graph sheaf theorem states the quotient description of an edge module and freeness for the normalized indecomposable boundary sheaf, through the cited moment-graph and Soergel constructions. bg:bmp.
Lean (primary): The tensor word first supplies explicit subexpression evaluations and image stalks. Right coefficient characters give label annihilation; a first-factor reflection proves the exact kernel across the initial simple reflection. Reflected pair lattices then describe stalks after prepending a letter. Frobenius:1558, Frobenius:2018, BottSamelson:214, BottSamelson:964, BottSamelson:1130, BottSamelson:1510, Support:426.
Role in the proof: The word-length induction proves stalk freeness and exact kernels together. Its two tail maps, quotient regularity and lattice saturation prove the non-head reflection step; the ascent/descent split sequences prove new stalk freeness. Coefficient transport then gives the polynomial word-sheaf input to the boundary-summand construction.
Proof
Step 1. Recover each coordinate of a pair
The coordinate map is \(A\)-linear by the calculation in its definition. Suppose \(C_x(p,q)=C_x(p',q')\). Evaluating at every zero-bit coordinate gives \(p_\eta=p'_\eta\), hence \(p=p'\) in the first stalk.
Evaluating at the one-bit coordinates gives \(\tau(q_\eta)=\tau(q'_\eta)\). Injectivity of \(\tau\) gives equality of all coordinates of \(q,q'\), hence \(q=q'\). Thus \(C_x\) is injective on the ambient pair module and on its lattice.
Step 2. Send every lattice pair into the new stalk
If \((p,q)\in L_x\), choose \(n\in M_{\mathbf t}\) with \(f(n)=p\) and \(g(n)=q\). The evaluation recursion gives \[ C_x(p,q)=E_{s\mathbf t,x}(1\otimes n). \] It is therefore in the image defining \(M_{s\mathbf t}^x\). This defines the codomain-restricted map bottHeadPairToStalk.
Step 3. Show that the new stalk lies in this image
For a pure tensor \(a\otimes n\), linearity gives \[ E_{s\mathbf t,x}(a\otimes n) =a\,E_{s\mathbf t,x}(1\otimes n) =a\,C_x(f(n),g(n)). \] The pair \((f(n),g(n))\) belongs to \(L_x\), and scalar closure of the twisted lattice places its multiple there too. Thus every pure-tensor evaluation is in the image of the linear map \(C_x|_{L_x}\).
Every tensor is a finite sum of pure tensors. The image of a linear map is closed under addition, so all word evaluations lie in this image. Since the stalk consists exactly of those evaluations, the map is surjective. Together with Step 1 this proves the isomorphism.
Step 4. Define and check the head projections
Apply \(C_x^{-1}\), then project to its first or second coordinate. The first projection is \(A\)-linear. If \(p\in X\), surjectivity of \(f\) gives \(n\) with \(f(n)=p\); the pair \((f(n),g(n))\) proves surjectivity of \(h_0\).
For the second projection, the lattice scalar law gives \[ h_1(az)=\tau(a)h_1(z). \] Since \(\tau\) fixes real scalars, \(h_1\) is real-linear. This is the scalar structure used in the source's bottHeadRight; its value is the underlying vector of the twisted second stalk.
Step 5. Verify the homogeneous degrees
Choose a basis of the extended realization and use the resulting coefficient grading. A degree-\(n\) stalk vector has every evaluation coordinate in \(A_n\).
The coordinates of \(h_0(z)\) are precisely the zero-bit coordinates of \(z\), so they lie in \(A_n\). The one-bit coordinates of \(z\) are \(\tau((h_1(z))_\eta)\). Apply \(\tau\), which preserves degree and squares to the identity, to conclude \((h_1(z))_\eta\in A_n\).
Thus both head projections preserve degree. These maps give the two orientations used in the later stalk-character recursion.
References
The local source declarations below give the construction and proof.
Lean formalization
OAI.KLInvariance.TitsSpace.bottHeadPairCoordinates_injective: BottSamelson.lean:135.
OAI.KLInvariance.TitsSpace.bottHeadPairCoordinates_seed: BottSamelson.lean:172.
OAI.KLInvariance.TitsSpace.bottHeadPairToStalk: BottSamelson.lean:181.
OAI.KLInvariance.TitsSpace.bottHeadPair_range: BottSamelson.lean:197.
OAI.KLInvariance.TitsSpace.bottHeadPairToStalk_surjective: BottSamelson.lean:207.
OAI.KLInvariance.TitsSpace.bottHeadPairEquiv: BottSamelson.lean:214.
OAI.KLInvariance.TitsSpace.bottHeadLeft: BottSamelson.lean:629.
OAI.KLInvariance.TitsSpace.bottHeadRight: BottSamelson.lean:636.
OAI.KLInvariance.TitsSpace.bottHeadLeft_surjective: BottSamelson.lean:644.
OAI.KLInvariance.TitsSpace.bottHeadPairEquiv_coordinates: BottSamelson.lean:649.
OAI.KLInvariance.TitsSpace.bottHeadLeft_homogeneous: BottSamelson.lean:655.
OAI.KLInvariance.TitsSpace.bottHeadRight_homogeneous: BottSamelson.lean:669.