diff --git a/verification/Proofs/CompressSpec.lean b/verification/Proofs/CompressSpec.lean
new file mode 100644
index 0000000..5fec4c1
--- /dev/null
+++ b/verification/Proofs/CompressSpec.lean
@@ -0,0 +1,202 @@
+/- ──────────────────────────────────────────────────────────────────────────────
+   Proofs/CompressSpec.lean — phase 2, brick 1: `EdwardsPoint::compress` emits
+   the canonical wire encoding of the denoted affine point.
+
+   THE THEOREM (`ed_compress_spec`): for a valid extended point Pt,
+       compress Pt = ok s   with
+       bytesVal s = (edY Pt).val + ((edX Pt).val % 2) · 2²⁵⁵
+   — the 32 bytes are the canonical little-endian encoding of the affine
+   y-coordinate with the parity ("sign") bit of the affine x-coordinate in
+   bit 255. Combined with `verify_accepts_iff` (SigApexSpec), this pins the
+   byte comparison of the apex to an equation on DENOTED POINTS — the
+   half-lift toward the point-level verification equation.
+
+   CHAIN (all real extracted code, all previously certified):
+     to_affine  = invert Z, mul X, mul Y        (invert_spec, mul_spec')
+     compress   = to_bytes y                     (to_bytes_spec — canonicity)
+                  is_negative x                  (to_bytes again, bit 0)
+                  s[31] ^= sign << 7             (XOR on a clear bit = +2²⁵⁵)
+
+   The parity of the CANONICAL encoding is the standard "sign" convention:
+   is_negative reads bit 0 of to_bytes, i.e. (feVal x mod p) mod 2 =
+   (edX Pt).val mod 2.
+   ────────────────────────────────────────────────────────────────────────────── -/
+import Proofs.ToBytesSpec
+import Proofs.InvertSpec
+import Proofs.EdMain
+open Aeneas Aeneas.Std Result
+open curve25519_dalek
+
+set_option maxHeartbeats 4000000
+set_option linter.unusedSimpArgs false
+set_option maxRecDepth 8000
+
+namespace CurveFieldProofs
+
+open Aeneas.Std.WP
+
+/-- Premise-free restatement of `to_bytes_spec` (destructures internally). -/
+theorem to_bytes_spec' (a : Fe) :
+    fe_to_bytes a ⦃ s => bytesVal s = feVal a % P ⦄ := by
+  obtain ⟨l0, l1, l2, l3, l4, hl⟩ := Fe.exists_limbs a
+  exact to_bytes_spec a l0 l1 l2 l3 l4 hl
+
+/-- Destructuring device for 32-byte arrays (the `Fe.exists_limbs` idiom). -/
+theorem Bytes32.exists_bytes (s : Std.Array Std.U8 32#usize) :
+    ∃ b0 b1 b2 b3 b4 b5 b6 b7 b8 b9 b10 b11 b12 b13 b14 b15
+      b16 b17 b18 b19 b20 b21 b22 b23 b24 b25 b26 b27 b28 b29 b30 b31 : Std.U8,
+      (↑s : List Std.U8) = [b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11,
+        b12, b13, b14, b15, b16, b17, b18, b19, b20, b21, b22, b23, b24, b25,
+        b26, b27, b28, b29, b30, b31] := by
+  have h : (↑s : List Std.U8).length = 32 := by
+    have := s.property
+    simp_all
+  match hl : (↑s : List Std.U8) with
+  | [c0, c1, c2, c3, c4, c5, c6, c7, c8, c9, c10, c11, c12, c13, c14, c15,
+     c16, c17, c18, c19, c20, c21, c22, c23, c24, c25, c26, c27, c28, c29, c30, c31] =>
+    exact ⟨c0, c1, c2, c3, c4, c5, c6, c7, c8, c9, c10, c11, c12, c13, c14, c15,
+      c16, c17, c18, c19, c20, c21, c22, c23, c24, c25, c26, c27, c28, c29, c30, c31, rfl⟩
+  | [] | [_] | [_,_] | [_,_,_] | [_,_,_,_] | [_,_,_,_,_] | [_,_,_,_,_,_] | [_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] | [_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_] => simp [hl] at h
+  | _::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_::_ => simp [hl] at h
+
+/-- **`is_negative` is the parity of the canonical residue** — bit 0 of the
+    canonical encoding (`Choice` is the transparent-u8 model). -/
+theorem is_negative_spec (a : Fe) :
+    field.FieldElement51.is_negative a ⦃ c => c.val = (feVal a % P) % 2 ⦄ := by
+  unfold field.FieldElement51.is_negative
+  step with (to_bytes_spec' a) as ⟨s, hs⟩
+  obtain ⟨b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15,
+    b16, b17, b18, b19, b20, b21, b22, b23, b24, b25, b26, b27, b28, b29, b30, b31,
+    hsl⟩ := Bytes32.exists_bytes s
+  step as ⟨i, hi⟩
+  simp [hsl] at hi
+  step as ⟨i1, hi1⟩
+  -- into Choice = the identity From instance
+  simp only [core.convert.IntoFrom.into, subtle.Choice.Insts.CoreConvertFromU8,
+    subtle.Choice.Insts.CoreConvertFromU8.from]
+  try simp only [spec_ok]
+  -- value: b0 &&& 1 = b0 % 2 = bytesVal s % 2 = (feVal a % P) % 2
+  have hi1v : i1.val = b0.val % 2 := by
+    rw [hi1, hi, UScalar.val_and]
+    have := Nat.and_two_pow_sub_one_eq_mod b0.val 1
+    norm_num at this
+    simpa using this
+  have hb0 : b0.val % 2 = bytesVal s % 2 := by
+    simp only [bytesVal, hsl]
+    omega
+  rw [hi1v, hb0, hs]
+
+/-- **`compress` on a valid extended point emits the canonical encoding**:
+    the affine y-residue with the x-parity bit at position 255. -/
+theorem ed_compress_spec (Pt : EdPoint) (hv : ExtValid Pt) :
+    edwards.EdwardsPoint.compress Pt ⦃ s =>
+      bytesVal s = (edY Pt).val + ((edX Pt).val % 2) * 2^255 ⦄ := by
+  obtain ⟨hbX, hbY, hbZ, hZne, hcoh⟩ := hv
+  unfold edwards.EdwardsPoint.compress edwards.EdwardsPoint.to_affine
+  -- recip ← invert Z
+  step with (invert_spec Pt.Z (Bnd.mono hbZ (by norm_num))) as ⟨recip, hbr, hrv⟩
+  -- x ← X · recip, y ← Y · recip
+  step with (mul_spec' Pt.X recip (Bnd.mono hbX (by norm_num)) (Bnd.mono hbr (by norm_num)))
+    as ⟨x, hbx, hxv⟩
+  step with (mul_spec' Pt.Y recip (Bnd.mono hbY (by norm_num)) (Bnd.mono hbr (by norm_num)))
+    as ⟨y, hby, hyv⟩
+  -- affine compress
+  unfold edwards.affine.AffinePoint.compress
+  step with (to_bytes_spec' y) as ⟨s, hs⟩
+  obtain ⟨b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b12, b13, b14, b15,
+    b16, b17, b18, b19, b20, b21, b22, b23, b24, b25, b26, b27, b28, b29, b30, b31,
+    hsl⟩ := Bytes32.exists_bytes s
+  step with (is_negative_spec x) as ⟨c, hc⟩
+  -- unwrap_u8 is the transparent-u8 identity: reduce it away
+  simp only [subtle.Choice.unwrap_u8, bind_tc_ok]
+  have hsignv : c.val = (feVal x % P) % 2 := hc
+  -- sign << 7
+  step as ⟨hi7, hhi7⟩
+  have hsle : c.val ≤ 1 := by rw [hsignv]; omega
+  have hhi7v : hi7.val = c.val * 2^7 := by
+    rw [hhi7]
+    simp only [Nat.shiftLeft_eq]
+    rw [Nat.mod_eq_of_lt (show c.val * 2^7 < U8.size by scalar_tac)]
+  -- s[31]
+  step as ⟨t31, ht31⟩
+  simp [hsl] at ht31
+  -- xor
+  step as ⟨t31x, ht31x⟩
+  -- update
+  step as ⟨s1, hs1⟩
+  have hsl1 : (↑s1 : List Std.U8) = [b0, b1, b2, b3, b4, b5, b6, b7, b8, b9,
+      b10, b11, b12, b13, b14, b15, b16, b17, b18, b19, b20, b21, b22, b23,
+      b24, b25, b26, b27, b28, b29, b30, t31x] := by
+    simp only [hs1, Array.set_val_eq, hsl]
+    rfl
+  try simp only [spec_ok]
+  -- ── value assembly ───────────────────────────────────────────────────────
+  -- the canonical y-residue is < p < 2²⁵⁵, so its top byte is < 2⁷
+  have hsval : bytesVal s = feVal y % P := hs
+  have hslt : bytesVal s < 2^255 := by
+    rw [hsval]
+    have : feVal y % P < P := Nat.mod_lt _ (by unfold P; norm_num)
+    unfold P at this ⊢
+    omega
+  have hb31top : b31.val < 2^7 := by
+    have hexp : bytesVal s = b0.val + b1.val * 2^8 + b2.val * 2^16 + b3.val * 2^24
+        + b4.val * 2^32 + b5.val * 2^40 + b6.val * 2^48 + b7.val * 2^56
+        + b8.val * 2^64 + b9.val * 2^72 + b10.val * 2^80 + b11.val * 2^88
+        + b12.val * 2^96 + b13.val * 2^104 + b14.val * 2^112 + b15.val * 2^120
+        + b16.val * 2^128 + b17.val * 2^136 + b18.val * 2^144 + b19.val * 2^152
+        + b20.val * 2^160 + b21.val * 2^168 + b22.val * 2^176 + b23.val * 2^184
+        + b24.val * 2^192 + b25.val * 2^200 + b26.val * 2^208 + b27.val * 2^216
+        + b28.val * 2^224 + b29.val * 2^232 + b30.val * 2^240 + b31.val * 2^248 := by
+      simp only [bytesVal, hsl]
+    omega
+  -- the xor adds sign·2⁷ to a clear bit
+  have ht31xv : t31x.val = b31.val + c.val * 2^7 := by
+    have hsb : c.val = 0 ∨ c.val = 1 := by
+      rw [hsignv]; omega
+    rw [ht31x, UScalar.val_xor, ht31, hhi7v]
+    rcases hsb with h | h
+    · simp [h]
+    · rw [h]
+      norm_num
+      exact xor_top_bit b31.val hb31top
+  -- reassemble: only byte 31 changed, and it grew by sign·2⁷ at weight 2²⁴⁸
+  have hexp : bytesVal s = b0.val + b1.val * 2^8 + b2.val * 2^16 + b3.val * 2^24
+      + b4.val * 2^32 + b5.val * 2^40 + b6.val * 2^48 + b7.val * 2^56
+      + b8.val * 2^64 + b9.val * 2^72 + b10.val * 2^80 + b11.val * 2^88
+      + b12.val * 2^96 + b13.val * 2^104 + b14.val * 2^112 + b15.val * 2^120
+      + b16.val * 2^128 + b17.val * 2^136 + b18.val * 2^144 + b19.val * 2^152
+      + b20.val * 2^160 + b21.val * 2^168 + b22.val * 2^176 + b23.val * 2^184
+      + b24.val * 2^192 + b25.val * 2^200 + b26.val * 2^208 + b27.val * 2^216
+      + b28.val * 2^224 + b29.val * 2^232 + b30.val * 2^240 + b31.val * 2^248 := by
+    simp only [bytesVal, hsl]
+  have hexp1 : bytesVal s1 = b0.val + b1.val * 2^8 + b2.val * 2^16 + b3.val * 2^24
+      + b4.val * 2^32 + b5.val * 2^40 + b6.val * 2^48 + b7.val * 2^56
+      + b8.val * 2^64 + b9.val * 2^72 + b10.val * 2^80 + b11.val * 2^88
+      + b12.val * 2^96 + b13.val * 2^104 + b14.val * 2^112 + b15.val * 2^120
+      + b16.val * 2^128 + b17.val * 2^136 + b18.val * 2^144 + b19.val * 2^152
+      + b20.val * 2^160 + b21.val * 2^168 + b22.val * 2^176 + b23.val * 2^184
+      + b24.val * 2^192 + b25.val * 2^200 + b26.val * 2^208 + b27.val * 2^216
+      + b28.val * 2^224 + b29.val * 2^232 + b30.val * 2^240 + t31x.val * 2^248 := by
+    simp only [bytesVal, hsl1]
+  have hstep : bytesVal s1 = bytesVal s + c.val * 2^255 := by
+    rw [hexp, hexp1, ht31xv]
+    ring
+  -- denotation bridges: the residues are the .val of the denoted coordinates
+  haveI : NeZero P := ⟨by unfold P; norm_num⟩
+  have hyval : feVal y % P = (edY Pt).val := by
+    have h1 : ⟪y⟫ = edY Pt := by
+      rw [hyv, hrv]
+      unfold edY
+      rw [div_eq_mul_inv]
+    rw [← h1]
+    simp [denote, ZMod.val_natCast]
+  have hxval : (feVal x % P) % 2 = (edX Pt).val % 2 := by
+    have h1 : ⟪x⟫ = edX Pt := by
+      rw [hxv, hrv]
+      unfold edX
+      rw [div_eq_mul_inv]
+    rw [← h1]
+    simp [denote, ZMod.val_natCast]
+  rw [hstep, hs, hsignv, hyval, hxval]
+
+end CurveFieldProofs
diff --git a/verification/Proofs/ToBytesMath.lean b/verification/Proofs/ToBytesMath.lean
index 93370ca..3c50697 100644
--- a/verification/Proofs/ToBytesMath.lean
+++ b/verification/Proofs/ToBytesMath.lean
@@ -262,4 +262,20 @@ theorem bytes_pack (f0 f1 f2 f3 f4 : ℕ)
   zify at s0 s1 s2 s3 s4 ⊢
   linear_combination s0 + 2^51 * s1 + 2^102 * s2 + 2^153 * s3 + 2^204 * s4
 
+/-- Setting a clear top bit by XOR is addition (compress's sign-bit write). -/
+theorem xor_top_bit (a : ℕ) (h : a < 2^7) : a ^^^ 2^7 = a + 2^7 := by
+  have hdiv : (a ^^^ 2^7) / 2^7 = 1 := by
+    rw [Nat.xor_div_two_pow]
+    rw [Nat.div_eq_of_lt h]
+    norm_num
+  have hand := Nat.and_two_pow_sub_one_eq_mod (a ^^^ 2^7) 7
+  have hdistrib := Nat.and_xor_distrib_right (a := a) (b := 2^7) (c := 2^7 - 1)
+  have ha : a &&& (2^7 - 1) = a := by
+    rw [Nat.and_two_pow_sub_one_eq_mod, Nat.mod_eq_of_lt h]
+  have h2 : 2^7 &&& (2^7 - 1) = 0 := by decide
+  have hmod : (a ^^^ 2^7) % 2^7 = a := by
+    rw [← hand, hdistrib, ha, h2, Nat.xor_zero]
+  have hdm := Nat.div_add_mod (a ^^^ 2^7) (2^7)
+  omega
+
 end CurveFieldProofs
diff --git a/verification/check.sh b/verification/check.sh
index 5382cd0..b725d94 100755
--- a/verification/check.sh
+++ b/verification/check.sh
@@ -66,6 +66,7 @@ PROOFS=(
   DsmMulSpec
   ToBytesMath
   ToBytesSpec
+  CompressSpec
   SigApexSpec
 )
 # Fully-qualified certificate names; each must be axiom-clean.
@@ -87,6 +88,7 @@ CERTS=(
   CurveFieldProofs.vartime_double_base_mul_spec
   CurveFieldProofs.verify_loop_full
   CurveFieldProofs.to_bytes_spec
+  CurveFieldProofs.ed_compress_spec
 )
 # Imports needed so every certificate in CERTS is in scope for the audit.
 AUDIT_IMPORTS=(
@@ -100,6 +102,7 @@ AUDIT_IMPORTS=(
   Proofs.DsmNafSpec
   Proofs.DsmMulSpec
   Proofs.ToBytesSpec
+  Proofs.CompressSpec
   Proofs.SigApexSpec
 )