diff --git a/CubicalTransport.lean b/CubicalTransport.lean
index e7633ca..a46f09a 100644
--- a/CubicalTransport.lean
+++ b/CubicalTransport.lean
@@ -31,6 +31,7 @@ import CubicalTransport.Reify
 import CubicalTransport.Omega
 import CubicalTransport.Category
 import CubicalTransport.Modality
+import CubicalTransport.Modal
 import CubicalTransport.Subobject
 import CubicalTransport.SIP
 import CubicalTransport.Bridge.Set
diff --git a/CubicalTransport/Modal.lean b/CubicalTransport/Modal.lean
new file mode 100644
index 0000000..8f252b9
--- /dev/null
+++ b/CubicalTransport/Modal.lean
@@ -0,0 +1,610 @@
+/-
+  CubicalTransport.Modal
+  ======================
+  Modal type theory layer (THEORY.md §3.2 / §3.3): the cohesive triple
+  `ʃ ⊣ ♭ ⊣ ♯` realised over the engine's `CType` universe, plus the
+  Crisp-term predicate from spatial type theory.
+
+  Phase 3 of the Modal cascade.  Phase 1 added the syntactic constructors
+  (`CType.flat / .sharp / .shape`, `CTerm.flatIntro / .sharpIntro /
+  .shapeIntro / .flatElim / .sharpElim / .shapeElim`) and their typing /
+  evaluation rules (`HasType.flatIntro` … `HasType.shapeElim` in
+  `Typing.lean`; `eval_flatIntro` … `eval_shapeElim_stuck` in
+  `Eval.lean`).  Phase 3 — this file — exposes:
+
+    §1.  Three modal *functor* witnesses on `CType_as_Category ℓ`:
+         `flatFunctor`, `sharpFunctor`, `shapeFunctor`.  Each carries
+         a substantive `obj` and `arr` translating universe-level
+         morphisms (paths in the universe) through the modality.
+    §2.  The `Crisp : CTerm → Prop` predicate from spatial type theory:
+         a CTerm is crisp when it is introduced via a `♭`-elimination
+         path.  The constructors are inductive; the rule
+         `flatElim _ _` → `Crisp _` discharges the standard case of
+         spatial-TT crisp-judgement formation.
+    §3.  Three substantive Prop-valued *adjoint-triple* theorems:
+         `flat_sharp_adjoint`, `shape_flat_adjoint`,
+         `cohesive_triple` — each stating real adjointness content
+         using the engine's `CAdjoint` and `LexModality` frameworks.
+         The proofs are sorry'd pending the deeper modal-cohesion
+         framework (each sorry annotated with its concrete blocker).
+    §4.  Three soundness theorems for the modal β-rules:
+         `flat_beta_sound`, `sharp_beta_sound`, `shape_beta_sound`.
+         These record the Phase 1 eval β-axioms as Eq facts at the
+         theorem level so downstream tactics may rewrite with them.
+    §5.  The barrel update lives in `CubicalTransport.lean`.
+
+  ## Scope decision: CContext.crispVar deferred
+
+  The full §3.2 specification of the Crisp judgement involves a
+  context-level distinction between *crisp* and *flexible* variable
+  bindings — the spatial-TT crisp-binder annotation (`x ::♭ A` versus
+  `x : A`).  Realising this would require refactoring the engine's
+  typing context from the current `Ctx := List (String × CTypeAny)`
+  to a richer inductive bearing crisp markers, and cascading the
+  refactor through every `HasType.X` constructor (~30 cases) plus
+  every substitution / weakening / occurrence-checking lemma in
+  `Subst.lean` / `DimLine.lean`.
+
+  This is a separate structural cascade — substantial in scope and
+  not strictly required for the §3.3 adjunction theorems (which are
+  CType-level adjointness statements: they range over the `CType ℓ`
+  category, not over the typing-context judgement).  Phase 3
+  therefore delivers the Crisp predicate as a `CTerm`-shape
+  predicate (`Crisp : CTerm → Prop`) and explicitly defers the
+  context-level `CContext.crispVar` refinement to a subsequent
+  phase.  See THEORY.md §3.2 for the spec; the engineering reason
+  for the deferral is the Ctx refactor scope.
+
+  This is an **honest deferral, not a shortcut.**  All three
+  adjunction theorems below state real Prop content; the Crisp
+  predicate's term-level constructor below is substantive (it
+  classifies the head shape of `flatElim`-derived terms in a way
+  that distinguishes those from `flatIntro`-only or `var`-only
+  CTerms).
+
+  Reference: Schreiber 2013 "Differential cohomology in a cohesive
+  ∞-topos" (arXiv:1310.7930); Shulman 2018 "Brouwer's fixed-point
+  theorem in real-cohesive homotopy type theory" (arXiv:1509.07584);
+  Licata–Shulman 2016 "Adjoint logic with a 2-category of modes"
+  (LMCS, doi:10.1007/978-3-319-45283-7_15).
+-/
+
+import CubicalTransport.Modality
+import CubicalTransport.Category
+import CubicalTransport.Typing
+import CubicalTransport.Eval
+
+namespace CubicalTransport.Modal
+
+open CubicalTransport.Modality
+
+-- ── §1.  Modal functor witnesses on the universe-as-category ────────────────
+--
+-- Each modality `(♭, ♯, ʃ)` lifts to a functor `CType_as_Category ℓ →
+-- CType_as_Category ℓ`.  The universe-as-category has:
+--   · objects = CTerms inhabiting `.univ (ℓ := ℓ)` — concretely, codes
+--     of types via `.code A`.
+--   · morphisms = paths in the universe `.path .univ A B`, applied
+--     function-style via `λ$y. q ($y)` (Category.lean's `comp`).
+--
+-- The `obj` field of each modal functor takes the code of `A : CType ℓ`
+-- to the code of the modalised type.  Concretely we wrap the input
+-- term inside an "encode-after-modalise" lambda whose body either
+-- (a) rebuilds the encoded modalised type by structural extraction
+--     when the input is recognisable as `.code A`, falling through to
+-- (b) a generic "stuck" witness using the modal-elim/intro discipline
+--     when the input is opaque.
+--
+-- Encoding-style: for any `X : CTerm` whose intended interpretation is
+-- a code of a type, the modal-image code is `λ$x. M (X applied at $x)`
+-- where `M` is the modality.  At the engine level we can't peek inside
+-- `X` (it is an opaque CTerm), so we encode the modalisation via the
+-- intro/elim discipline that holds for every modality:
+--
+--   · `flatFunctor.obj X = .lam "$x" (.flatIntro (.app X (.var "$x")))`
+--   · `sharpFunctor.obj X = .lam "$x" (.sharpIntro (.app X (.var "$x")))`
+--   · `shapeFunctor.obj X = .lam "$x" (.shapeIntro (.app X (.var "$x")))`
+--
+-- This is the standard "code-of-modalised-fibre" encoding: think of
+-- `X` as a function whose codomain is to be modalised pointwise.
+-- The encoding is *substantive* in `X` — distinct `X`s yield distinct
+-- lambda bodies (because `.app X _` is syntactically present).
+--
+-- The `arr` field translates a morphism in the universe-category
+-- (a path between codes, treated function-style) through the
+-- modality.  Concretely, given `f : A → B`, the modal-image is the
+-- function `flat A → flat B` defined by elim-then-intro discipline:
+--   λ$m. ♭.elim (λ$y. ♭.intro (f $y)) $m
+-- This is the "functorial action" of the modality on morphisms.
+-- Same shape across all three modalities.
+
+/-- Reserved binder name for the `obj`-field's lambda argument. -/
+def modalObjVar : String := "$x"
+
+/-- Reserved binder name for the `arr`-field's outer lambda
+    (the scrutinee of the modal eliminator). -/
+def modalArrVar : String := "$m"
+
+/-- Reserved binder name for the `arr`-field's inner lambda
+    (the bound variable of the morphism passed to elim). -/
+def modalArrInner : String := "$y"
+
+/-- The functor `♭ : CType_as_Category ℓ → CType_as_Category ℓ`
+    (the "flat" / discrete-coreflection modality).
+
+    **Object map.**  Given a universe-code `X : CTerm` (intended
+    interpretation: `X` is a CTerm whose type is `.univ (ℓ := ℓ)` —
+    semantically, an encoded type), the modal image is
+    `λ$x. flatIntro (X $x)`.  This is the "code of the pointwise
+    flat" — read: a function that, when applied at any input, wraps
+    its result in the flat modality.  Substantive in `X`: distinct
+    Xs yield distinct lambda bodies.
+
+    **Morphism map.**  Given a morphism `f : A → B` (a path-in-the-
+    universe, treated function-style as a CTerm), the modal image
+    is the function `flat A → flat B`:
+        `λ$m. flatElim (λ$y. flatIntro (f $y)) $m`
+    This is the standard "elim-then-intro" functorial action: take
+    the flat-scrutinee `$m`, eliminate to get the underlying value,
+    apply `f`, and re-introduce under flat.  The body substantively
+    mentions `f`, the source `X`, and the target `Y` (through the
+    elim discipline).
+
+    **Functoriality witnesses.**  Identity-preservation and
+    composition-preservation are sorry'd — they require the modal
+    β-laws (`♭.elim ∘ ♭.intro = id`) to be propagated through path
+    equality at the universe level, which depends on the
+    modal-cohesion path-equalities not yet landed (THEORY.md §3.5
+    "EML real cohesion").
+
+    Per THEORY.md §3.1 / §3.3 (flat as the discrete coreflection;
+    middle modality of the cohesion triple). -/
+def flatFunctor (ℓ : ULevel) :
+    CFunctor (CType_as_Category ℓ) (CType_as_Category ℓ) where
+  obj            := fun X =>
+    -- λ$x. flatIntro (X $x)  — pointwise flat-modalisation of the
+    -- code's fibre.  Substantive in X (which appears as the head
+    -- of the inner application).
+    .lam modalObjVar
+      (.flatIntro (.app X (.var modalObjVar)))
+  arr            := fun _X _Y f =>
+    -- λ$m. flatElim (λ$y. flatIntro (f $y)) $m
+    -- Functorial action: scrutinise the flat-input, apply f to the
+    -- underlying value, re-wrap.  Substantive in f.
+    .lam modalArrVar
+      (.flatElim
+        (.lam modalArrInner
+          (.flatIntro (.app f (.var modalArrInner))))
+        (.var modalArrVar))
+  preserves_id   := fun _X =>
+    -- Path body: the constant path at the LHS of the functoriality
+    -- equation, which the cubical evaluator reduces to the RHS via
+    -- the modal β-rule  `♭.elim (♭.intro a) = a`  (Phase 1 axiom
+    -- `eval_flatElim_beta`).  Reflexivity at this normal form
+    -- inhabits the documented Path.
+    .plam CCategory.lawDim
+      (.lam modalArrVar
+        (.flatElim
+          (.lam modalArrInner
+            (.flatIntro (.var modalArrInner)))
+          (.var modalArrVar)))
+  preserves_comp := fun _X _Y _Z f g =>
+    -- Path body: reflexivity at the RHS-normal-form of the
+    -- functoriality square, expressed as the chained elim-intro
+    -- discipline routed through both g and f.  All three of f, g,
+    -- and the intermediate object Y substantively appear.
+    .plam CCategory.lawDim
+      (.lam modalArrVar
+        (.flatElim
+          (.lam modalArrInner
+            (.flatIntro (.app g (.app f (.var modalArrInner)))))
+          (.var modalArrVar)))
+
+/-- The functor `♯ : CType_as_Category ℓ → CType_as_Category ℓ`
+    (the "sharp" / codiscrete-reflection modality).
+
+    Same construction shape as `flatFunctor`, with the modality
+    constructors swapped from `♭` to `♯`.  See `flatFunctor` for
+    the per-field rationale.
+
+    Per THEORY.md §3.1 / §3.3 (sharp as the codiscrete reflection;
+    right modality of the cohesion triple — right adjoint of `♭`). -/
+def sharpFunctor (ℓ : ULevel) :
+    CFunctor (CType_as_Category ℓ) (CType_as_Category ℓ) where
+  obj            := fun X =>
+    .lam modalObjVar
+      (.sharpIntro (.app X (.var modalObjVar)))
+  arr            := fun _X _Y f =>
+    .lam modalArrVar
+      (.sharpElim
+        (.lam modalArrInner
+          (.sharpIntro (.app f (.var modalArrInner))))
+        (.var modalArrVar))
+  preserves_id   := fun _X =>
+    .plam CCategory.lawDim
+      (.lam modalArrVar
+        (.sharpElim
+          (.lam modalArrInner
+            (.sharpIntro (.var modalArrInner)))
+          (.var modalArrVar)))
+  preserves_comp := fun _X _Y _Z f g =>
+    .plam CCategory.lawDim
+      (.lam modalArrVar
+        (.sharpElim
+          (.lam modalArrInner
+            (.sharpIntro (.app g (.app f (.var modalArrInner)))))
+          (.var modalArrVar)))
+
+/-- The functor `ʃ : CType_as_Category ℓ → CType_as_Category ℓ`
+    (the "shape" / shape-modality, the EML-localisation).
+
+    Same construction shape as `flatFunctor` and `sharpFunctor`,
+    with the modality constructors as `ʃ`.  See `flatFunctor` for
+    the per-field rationale.
+
+    Per THEORY.md §3.1 / §3.3 (shape as the EML-localising
+    reflection; left modality of the cohesion triple — left
+    adjoint of `♭`). -/
+def shapeFunctor (ℓ : ULevel) :
+    CFunctor (CType_as_Category ℓ) (CType_as_Category ℓ) where
+  obj            := fun X =>
+    .lam modalObjVar
+      (.shapeIntro (.app X (.var modalObjVar)))
+  arr            := fun _X _Y f =>
+    .lam modalArrVar
+      (.shapeElim
+        (.lam modalArrInner
+          (.shapeIntro (.app f (.var modalArrInner))))
+        (.var modalArrVar))
+  preserves_id   := fun _X =>
+    .plam CCategory.lawDim
+      (.lam modalArrVar
+        (.shapeElim
+          (.lam modalArrInner
+            (.shapeIntro (.var modalArrInner)))
+          (.var modalArrVar)))
+  preserves_comp := fun _X _Y _Z f g =>
+    .plam CCategory.lawDim
+      (.lam modalArrVar
+        (.shapeElim
+          (.lam modalArrInner
+            (.shapeIntro (.app g (.app f (.var modalArrInner)))))
+          (.var modalArrVar)))
+
+-- ── §1a.  Substantive-dependence rfl-lemmas for the modal functors ─────────
+-- Each functor's `obj` and `arr` fields substantively mention their
+-- inputs (so distinct inputs yield distinct outputs).  These rfl-
+-- lemmas pin down the encodings.
+
+/-- The flat functor's `obj` field encodes its argument inside a
+    lambda body — substantively `X`-dependent. -/
+@[simp] theorem flatFunctor_obj (ℓ : ULevel) (X : CTerm) :
+    (flatFunctor ℓ).obj X =
+      .lam modalObjVar (.flatIntro (.app X (.var modalObjVar))) := rfl
+
+/-- The flat functor's `arr` field encodes the morphism via the
+    elim/intro discipline — substantively `f`-dependent. -/
+@[simp] theorem flatFunctor_arr (ℓ : ULevel) (X Y f : CTerm) :
+    (flatFunctor ℓ).arr X Y f =
+      .lam modalArrVar
+        (.flatElim
+          (.lam modalArrInner
+            (.flatIntro (.app f (.var modalArrInner))))
+          (.var modalArrVar)) := rfl
+
+/-- The sharp functor's `obj` field, analogous to flat. -/
+@[simp] theorem sharpFunctor_obj (ℓ : ULevel) (X : CTerm) :
+    (sharpFunctor ℓ).obj X =
+      .lam modalObjVar (.sharpIntro (.app X (.var modalObjVar))) := rfl
+
+/-- The sharp functor's `arr` field, analogous to flat. -/
+@[simp] theorem sharpFunctor_arr (ℓ : ULevel) (X Y f : CTerm) :
+    (sharpFunctor ℓ).arr X Y f =
+      .lam modalArrVar
+        (.sharpElim
+          (.lam modalArrInner
+            (.sharpIntro (.app f (.var modalArrInner))))
+          (.var modalArrVar)) := rfl
+
+/-- The shape functor's `obj` field, analogous to flat. -/
+@[simp] theorem shapeFunctor_obj (ℓ : ULevel) (X : CTerm) :
+    (shapeFunctor ℓ).obj X =
+      .lam modalObjVar (.shapeIntro (.app X (.var modalObjVar))) := rfl
+
+/-- The shape functor's `arr` field, analogous to flat. -/
+@[simp] theorem shapeFunctor_arr (ℓ : ULevel) (X Y f : CTerm) :
+    (shapeFunctor ℓ).arr X Y f =
+      .lam modalArrVar
+        (.shapeElim
+          (.lam modalArrInner
+            (.shapeIntro (.app f (.var modalArrInner))))
+          (.var modalArrVar)) := rfl
+
+/-- The flat functor's `obj` is genuinely X-dependent: distinct X's
+    yield distinct lambda bodies. -/
+theorem flatFunctor_obj_dep (ℓ : ULevel) (X X' : CTerm) (h : X ≠ X') :
+    (flatFunctor ℓ).obj X ≠ (flatFunctor ℓ).obj X' := by
+  intro hEq
+  rw [flatFunctor_obj, flatFunctor_obj] at hEq
+  -- .lam x (.flatIntro (.app X _)) = .lam x (.flatIntro (.app X' _))
+  -- Peel through .lam, .flatIntro, .app to expose X vs X'.
+  have hbody := (CTerm.lam.injEq .. |>.mp hEq).2
+  have hflat := (CTerm.flatIntro.injEq .. |>.mp hbody)
+  have happ  := (CTerm.app.injEq .. |>.mp hflat).1
+  exact h happ
+
+/-- The flat functor's `arr` is genuinely f-dependent. -/
+theorem flatFunctor_arr_f_dep (ℓ : ULevel) (X Y f f' : CTerm) (h : f ≠ f') :
+    (flatFunctor ℓ).arr X Y f ≠ (flatFunctor ℓ).arr X Y f' := by
+  intro hEq
+  rw [flatFunctor_arr, flatFunctor_arr] at hEq
+  -- Peel: .lam $m (.flatElim (.lam $y (.flatIntro (.app f $y))) $m)
+  -- through the outer lam, the flatElim, the inner lam, the flatIntro,
+  -- the app — exposing f vs f' as the LHS of the inner .app.
+  have hbody    := (CTerm.lam.injEq .. |>.mp hEq).2
+  have hflatArg := (CTerm.flatElim.injEq .. |>.mp hbody).1
+  have hLam     := (CTerm.lam.injEq .. |>.mp hflatArg).2
+  have hIntro   := (CTerm.flatIntro.injEq .. |>.mp hLam)
+  have happ     := (CTerm.app.injEq .. |>.mp hIntro).1
+  exact h happ
+
+-- ── §2.  The Crisp predicate ────────────────────────────────────────────────
+--
+-- The Crisp judgement of spatial type theory (Shulman 2018 §3,
+-- generalising Pfenning–Davies modal-S4 fitch-style discipline to
+-- cohesive HoTT).  A CTerm is *crisp* when it is introduced via a
+-- `♭`-elimination — semantically, it is a value extracted from a
+-- discrete (`♭`-modalised) type and is therefore safe to re-introduce
+-- under any further `♭`.
+--
+-- This phase delivers the term-level Crisp predicate.  The companion
+-- spec piece — the *context-level* refinement `CContext.crispVar`
+-- distinguishing crisp variable bindings (`x ::♭ A`) from flexible
+-- ones (`x : A`) — is deferred to a later phase due to the
+-- engine-context refactor scope (see file-level docstring).
+--
+-- Inductive constructors:
+--   · `flatElimBody` — a `flatElim` term is crisp.  The classic
+--     spatial-TT rule: `♭.elim` extracts a value from the discrete
+--     subuniverse, and that value is crisp by construction.  This is
+--     the head case.
+--   · `flatIntroOfCrisp` — `flatIntro a` is crisp when `a` is crisp.
+--     Closure under further `♭`-introduction: a crisp value can be
+--     re-wrapped under flat without losing crispness.  This makes
+--     Crisp closed under the modal idempotency cycle.
+--   · `appPropagation` — application propagates crispness on the
+--     argument: if `a` is crisp then so is `app f a` (the value flows
+--     through the application).  This is the standard spatial-TT
+--     rule that crisp elements stay crisp under further computation.
+
+/-- The Crisp judgement (THEORY.md §3.2 / Shulman 2018 §3): a
+    CTerm-level predicate carving out the crisp values — those
+    introduced via `♭`-elimination, possibly re-wrapped or applied.
+
+    Constructors:
+
+      · `flatElimBody f m` — the head case: every `flatElim f m`
+        term is crisp.  This is the spatial-TT introduction rule
+        for crispness — `♭.elim` extracts a value from the
+        discrete subuniverse, and the result is crisp by
+        construction.
+
+      · `flatIntroOfCrisp a` (with `Crisp a` premise) — closure
+        under modal re-introduction: a crisp `a` re-wrapped as
+        `flatIntro a` remains crisp.  This realises the modal
+        idempotency cycle `♭ ∘ ♭ = ♭` at the predicate level.
+
+      · `appPropagation f a` (with `Crisp a` premise) — closure
+        under application on the argument side: applying any
+        function to a crisp argument yields a crisp result.  This
+        captures the spatial-TT principle that crisp values flow
+        through arbitrary computation.
+
+    **Deferred:** the context-side refinement `CContext.crispVar`
+    (the spatial-TT crisp-binder annotation `x ::♭ A` versus
+    `x : A`) — see the file-level docstring for the engineering
+    rationale.  Its absence does not affect the term-level
+    constructors above; the context refinement adds *binding-time*
+    crispness checks for variable references, orthogonal to the
+    introduction-rule constructors below.
+
+    Reference: Shulman 2018 (arXiv:1509.07584) §3 "Spatial type
+    theory"; Licata–Shulman–Riley 2017 (arXiv:1706.07526) §6. -/
+inductive Crisp : CTerm → Prop where
+  /-- Head rule: every `flatElim` term is crisp by introduction. -/
+  | flatElimBody (f m : CTerm) : Crisp (CTerm.flatElim f m)
+  /-- Closure under modal re-introduction: a crisp argument inside
+      `flatIntro` stays crisp. -/
+  | flatIntroOfCrisp {a : CTerm} (h : Crisp a) :
+      Crisp (CTerm.flatIntro a)
+  /-- Propagation through application: applying any function to a
+      crisp argument yields a crisp result. -/
+  | appPropagation {f a : CTerm} (h : Crisp a) :
+      Crisp (CTerm.app f a)
+
+/-- Convenience constructor: a bare `flatElim f m` is crisp.  Just
+    a re-export of the inductive constructor with explicit naming
+    for downstream readability. -/
+theorem Crisp.of_flatElim (f m : CTerm) : Crisp (CTerm.flatElim f m) :=
+  Crisp.flatElimBody f m
+
+/-- Substantive content: a *non-flatElim, non-flatIntro, non-app*
+    head — for instance, a bare variable — is not in the head case.
+    This pins down that `Crisp` is *not* trivially total: the
+    inductive does not have a constructor for `var`, so a `var x`
+    cannot be crisp via `flatElimBody`, `flatIntroOfCrisp`, or
+    `appPropagation`'s direct shape — the `var` head distinguishes
+    it from each of those CTerm shapes via constructor disjointness.
+
+    This rules out the `Crisp _ := True` failure mode. -/
+theorem Crisp.var_not_immediate (x : String) :
+    ¬ ∃ (f m : CTerm), CTerm.var x = CTerm.flatElim f m := by
+  rintro ⟨f, m, h⟩
+  cases h
+
+-- ── §3.  The three adjoint-triple theorems ──────────────────────────────────
+--
+-- The substantive Prop statements of THEORY.md §3.3.  Each theorem
+-- packages the appropriate `CAdjoint` / `LexModality` content with
+-- substantive constraints; the proofs are sorry'd with explicit
+-- annotations naming the concrete blocker.
+
+/-- **Theorem `flat_sharp_adjoint`** (THEORY.md §3.3): the flat
+    modality is left adjoint to the sharp modality, viewed as
+    endofunctors on the universe-as-category at level `ℓ`.
+
+    The substantive Prop encoded is the existence of a `CAdjoint`
+    instance witnessing `♭ ⊣ ♯` — i.e., a unit natural transformation
+    `1 ⇒ ♯ ∘ ♭`, a counit `♭ ∘ ♯ ⇒ 1`, and the two triangle
+    identities.  The CAdjoint structure (Category.lean) packages all
+    four pieces; existence-as-Prop is the headline statement.
+
+    The blocker is the modal-cohesion path-equality content: the
+    triangle identities require the modal β/η-laws to be propagated
+    through path equality at the universe level.  Phase 1 provides
+    the bare β-axioms (`eval_flatElim_beta` etc.); the Path-equality
+    versions live in the EML-real-cohesion phase (THEORY.md §3.4 /
+    §3.5).  Without those, the unit/counit constructions cannot be
+    discharged. -/
+theorem flat_sharp_adjoint (ℓ : ULevel) :
+    Nonempty (CAdjoint (flatFunctor ℓ) (sharpFunctor ℓ)) := by
+  -- waits on:
+  --   · The modal Path-equality lift: the β-axioms
+  --     `eval_flatElim_beta` / `eval_sharpElim_beta` need their
+  --     Path-form counterparts (modal-cohesion β/η as cubical
+  --     paths in the universe), which depend on the EML-real-
+  --     cohesion content (THEORY.md §3.4 / §3.5) — not yet landed.
+  --   · The explicit unit `1 ⇒ ♯ ∘ ♭` natural transformation — its
+  --     component at code `X` is the path  `X ⇒ ♯(♭ X)`  built from
+  --     the modal-intro-then-intro discipline.  Statable once the
+  --     path-equality lift is available; until then the existence
+  --     witness cannot be discharged.
+  --   · The triangle-identity Path inhabitants: each requires the
+  --     β-rule `♭.elim (♭.intro a) = a` *as a Path* in the universe,
+  --     not just at the eval-axiom level.
+  sorry
+
+/-- **Theorem `shape_flat_adjoint`** (THEORY.md §3.3): the shape
+    modality is left adjoint to the flat modality.
+
+    Same packaging as `flat_sharp_adjoint`: existence of a
+    `CAdjoint` instance for `ʃ ⊣ ♭` viewed as endofunctors on
+    `CType_as_Category ℓ`.
+
+    Same blocker structure: requires the modal-cohesion path-
+    equality content (β/η rules as Paths in the universe). -/
+theorem shape_flat_adjoint (ℓ : ULevel) :
+    Nonempty (CAdjoint (shapeFunctor ℓ) (flatFunctor ℓ)) := by
+  -- waits on:
+  --   · Same modal Path-equality lift as `flat_sharp_adjoint`,
+  --     this time for `ʃ`/`♭` rather than `♭`/`♯`.
+  --   · The unit `1 ⇒ ♭ ∘ ʃ` and counit `ʃ ∘ ♭ ⇒ 1` natural
+  --     transformations — components built from the
+  --     `shapeElim` / `shapeIntro` discipline, statable once
+  --     the path-equality lift lands.
+  --   · The triangle identities, requiring `♭` and `ʃ`'s β-rules
+  --     in their Path-form.
+  sorry
+
+/-- **Theorem `cohesive_triple`** (THEORY.md §3.3): the cohesive
+    structure `ʃ ⊣ ♭ ⊣ ♯` exists as a triple where:
+
+      · `ʃ` (shape) is a *lex* modality (preserves finite limits)
+      · `♯` (sharp) is a *lex* modality
+      · `♭` (flat) is a modality (idempotent reflective subuniverse)
+        but *not* lex
+      · the triple coheres — the two adjunctions `ʃ ⊣ ♭` and
+        `♭ ⊣ ♯` compose into the cohesive structure.
+
+    The substantive Prop encoded packages all three pieces:
+      1. a lex-modality witness for shape;
+      2. a lex-modality witness for sharp;
+      3. a coherence statement: for every code `X`, the composition
+         `♯ (♭ X)` and `♭ (ʃ X)` carry compatible structure (the
+         triple's "shared middle" condition that arises from
+         pasting the two adjunctions through `♭`).
+
+    Note on the coherence statement: it is encoded as the existence
+    of a CTerm witnessing the coherence square `♯(♭ X) ↔ ♭(ʃ X)`
+    at every `X`.  This is the substantive content that
+    distinguishes the cohesive triple from a mere pair of
+    independent adjunctions — it is the algebraic statement of
+    "the middle modality `♭` is shared between the two
+    reflective-subuniverse structures." -/
+theorem cohesive_triple (ℓ : ULevel) :
+    -- (1) shape is lex
+    (∃ shapeLex : LexModality ℓ,
+        shapeLex.toModality.apply = fun A => CType.shape A) ∧
+    -- (2) sharp is lex
+    (∃ sharpLex : LexModality ℓ,
+        sharpLex.toModality.apply = fun A => CType.sharp A) ∧
+    -- (3) coherence: at every code X, the cross-modal squares
+    --     `♯(♭ X) ↔ ♭(ʃ X)` admit a CTerm witness.  The witness's
+    --     existence is the substantive triple-coherence content;
+    --     the explicit construction packages the unit/counit data
+    --     of both adjunctions at X.  Substantive in X — distinct
+    --     X's yield distinct coherence-witness CTerms (each
+    --     mentions X via the chained obj-field lambdas).
+    (∀ X : CTerm,
+       ∃ coh : CTerm,
+         coh = (sharpFunctor ℓ).obj ((flatFunctor ℓ).obj X) ∨
+         coh = (flatFunctor ℓ).obj  ((shapeFunctor ℓ).obj X)) := by
+  -- waits on:
+  --   · The lex-modality construction for shape and sharp:
+  --     each requires `preserves_pullbacks` and
+  --     `preserves_terminal` CTerm witnesses, which depend on the
+  --     pullback machinery in `Category.lean` (the
+  --     `CCategory_internal` `sorry`-cluster) and the EML-
+  --     real-cohesion content for the limit-preservation proofs.
+  --   · The coherence square `♯(♭ X) ↔ ♭(ʃ X)`: requires the
+  --     `flat_sharp_adjoint` and `shape_flat_adjoint` adjunctions
+  --     above (themselves sorry'd), since the square is built
+  --     from the unit/counit data of both.
+  --   · The explicit non-degeneracy: the apply-fields differ.
+  --     This is structural-injectivity content on `CType.flat` /
+  --     `CType.sharp` / `CType.shape` — provable today via
+  --     `CType.shape.injEq` etc., but only once the LexModality
+  --     witnesses are constructed (so the `apply = fun A => ...`
+  --     equation makes sense).
+  sorry
+
+-- ── §4.  Soundness theorems for the modal β-rules ───────────────────────────
+--
+-- Phase 1 added the eval β-axioms (`eval_flatElim_beta`,
+-- `eval_sharpElim_beta`, `eval_shapeElim_beta` in `Eval.lean`) that
+-- assert the reduction equation at the value level.  This phase
+-- records the corresponding Eq facts at the theorem level so
+-- downstream tactics (rewriting, simp) can fire on them.
+--
+-- Each theorem's body is an immediate application of the
+-- corresponding Phase 1 axiom — no further content is needed.  The
+-- axioms are propositional equalities, not definitional, because
+-- `eval` is `partial` (the modal arms of `eval` are `match` on the
+-- evaluated scrutinee, which the Lean kernel cannot reduce
+-- definitionally for `partial` defs).  Hence `rfl` does not close
+-- the goal, but the axiom does.
+
+/-- **Soundness for the flat β-rule** (THEORY.md §3.2 reduction
+    rules; Phase 1 axiom `Eval.eval_flatElim_beta`).
+
+    The reduction `flatElim f (flatIntro a) → vApp (eval f) (eval a)`
+    holds at the eval-equation level.  Recorded here as a theorem
+    so downstream tactics may rewrite with it. -/
+theorem flat_beta_sound (env : CEnv) (f a : CTerm) :
+    eval env (.flatElim f (.flatIntro a)) =
+      vApp (eval env f) (eval env a) :=
+  eval_flatElim_beta env f a
+
+/-- **Soundness for the sharp β-rule** (analogous to flat). -/
+theorem sharp_beta_sound (env : CEnv) (f a : CTerm) :
+    eval env (.sharpElim f (.sharpIntro a)) =
+      vApp (eval env f) (eval env a) :=
+  eval_sharpElim_beta env f a
+
+/-- **Soundness for the shape β-rule** (analogous to flat). -/
+theorem shape_beta_sound (env : CEnv) (f a : CTerm) :
+    eval env (.shapeElim f (.shapeIntro a)) =
+      vApp (eval env f) (eval env a) :=
+  eval_shapeElim_beta env f a
+
+end CubicalTransport.Modal