diff --git a/research/hypercube-decoder-parity/README.md b/research/hypercube-decoder-parity/README.md
new file mode 100644
index 0000000000..3a13dfd89b
--- /dev/null
+++ b/research/hypercube-decoder-parity/README.md
@@ -0,0 +1,56 @@
+# Why the paired-cube decoder does not extend to 4-dimensional (hyper)cubes (no new κ)
+
+A scoping note on the port dimension of the paired-cube bit word. **No κ is claimed, and nothing existing changes.**
+
+## The identity being generalized
+
+The paired-cube bit word (`research/paired-cube-bit/LEMMA.md`) computes the mod-2 decoder identity
+
+    x_T = sum_{c in T} star(c) + sum_{|S cap T| = 1} x_S        (mod 2),
+
+over **ports**: triples taking one coordinate from each of three distinct coordinate pairs.
+- `star(c)` is the sum of `x_S` over all ports containing coordinate `c`. These are the shared center sums.
+- The overlap-1 sums are what the cube's local circuits, modules and partner pairs compute.
+
+A natural question is whether moving to k-dimensional cubes, with ports taking one coordinate from each of k pairs, gives a larger deficit. The source count grows like 2^k·C(p,k), while the center loss grows like p².
+
+## Result
+
+Consider the star family of decoders for k-ports:
+
+    x_T = sum_{0<l<k} a_l sum_{L <= T, |L| = l} star_L  +  sum_{0<j<k} b_j sum_{|S cap T| = j} x_S     (mod 2),
+
+with `star_L` the sum of `x_S` over the ports containing the l-set `L`. Coefficients `a_l, b_j` are in F2. The global sum (`l = 0`) and the disjoint class (`j = 0`) are excluded, because together they give only the trivial "everything minus the rest" decoder.
+
+**A decoder of this form exists if and only if k is not a power of two.** In particular, there is none for **k = 4**.
+
+*Proof.* The coefficient of `x_S` in the right-hand side depends only on `j = |S cap T|`. It equals `sum_l a_l C(j,l) + b_j`, where `b_j` is present only for `0 < j < k`.
+- For `0 < j < k`, choosing `b_j` makes it 0.
+- For `j = 0` it is 0 automatically.
+- For `S = T` (`j = k`) it is `sum_{0<l<k} a_l C(k,l)`. This can be 1 exactly when some `C(k,l)` with `0 < l < k` is odd.
+
+By Lucas' theorem, every such `C(k,l)` is even iff `k` is a power of two. ∎
+
+| k | Example | Decoder |
+|---|---|---|
+| 3 | paired cubes (today) | yes, e.g. level-1 stars plus overlap-1 corrections (the lemma above) |
+| **4** | **hypercubes** | **no**: `C(4,1), C(4,2), C(4,3)` are all even |
+| 5 | | yes, e.g. level-1 stars plus overlap-1 and overlap-3 corrections |
+| 6 | | yes, needs level-2 stars (`C(6,2) = 15`) |
+| 8, 16, … | | no |
+
+## Check
+
+```sh
+python3 -B research/hypercube-decoder-parity/check.py      # standard library, under one second
+```
+
+For `(k, p) = (3,6), (4,6), (5,7), (6,7)`, the script tests every coefficient vector `(a, b)` against every column of every port. It confirms that valid decoders exist exactly when `k` is not a power of two, with 2, 0, 8 and 16 decoders respectively. It also confirms that the k = 3 solutions include the paired-cube identity.
+
+## Scope
+
+- The note concerns decoders built from stars and overlap-class corrections. That is the family the paired-cube construction uses, and the one whose terms local circuits and shared center sums can compute.
+- It is not an impossibility proof for every decoder over 4-dimensional ports.
+- It says nothing yet about whether k = 5 or k = 6 would beat k = 3. That would need new local circuits, partner pairs, caps and a full construction, and those decoders also need correction terms at more overlap sizes.
+
+Credits: icekylinx (#144, paired cubes) and eumemic (paired-cube bit word and its lemma). Prepared by Joel Pulikkan (GamingPuzzled) with Anthropic Claude assistance; Apache-2.0.
diff --git a/research/hypercube-decoder-parity/check.py b/research/hypercube-decoder-parity/check.py
new file mode 100644
index 0000000000..ef4952c5d5
--- /dev/null
+++ b/research/hypercube-decoder-parity/check.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python3
+"""Exhaustive check of the k-port decoder parity obstruction (README.md). Standard library only.
+
+Ports: k-subsets taking one coordinate from each of k distinct pairs among p coordinate pairs (k = 3 is the
+paired-cube port of research/paired-cube-bit/LEMMA.md). A decoder in the star family writes, for every port T,
+
+    x_T = sum_{0<l<k} a_l sum_{L <= T, |L| = l} star_L + sum_{0<j<k} b_j sum_{|S cap T| = j} x_S   (mod 2),
+
+with star_L = sum of x_S over ports S containing L. For each (k, p) below, every coefficient vector (a, b) is
+tested against every column of every port, and the count of valid decoders is compared with the closed form:
+valid decoders exist iff k is not a power of two.
+Copyright 2026 Joel Pulikkan (GamingPuzzled). Apache-2.0. Prepared with Anthropic Claude assistance.
+"""
+import itertools
+import sys
+from math import comb
+
+if sys.flags.optimize:
+    raise SystemExit('refusing -O')
+
+
+def ports(p, k):
+    return [frozenset(2 * q + e for q, e in zip(pairs, sel))
+            for pairs in itertools.combinations(range(p), k) for sel in itertools.product(range(2), repeat=k)]
+
+
+def valid_decoders(p, k):
+    P = ports(p, k)
+    overlap = [[len(S & T) for S in P] for T in P]
+    found = []
+    for a in itertools.product(range(2), repeat=k - 1):
+        for b in itertools.product(range(2), repeat=k - 1):
+            def coeff(j):
+                return (sum(al * comb(j, l) for l, al in enumerate(a, 1)) + (b[j - 1] if 0 < j < k else 0)) % 2
+            table = {j: coeff(j) for j in range(k + 1)}
+            if all(table[overlap[t][s]] == (1 if s == t else 0) for t in range(len(P)) for s in range(len(P))):
+                found.append((a, b))
+    return len(P), found
+
+
+def power_of_two(k):
+    return k & (k - 1) == 0
+
+
+def main():
+    for k, p in ((3, 6), (4, 6), (5, 7), (6, 7)):
+        n, found = valid_decoders(p, k)
+        parity = [comb(k, l) % 2 for l in range(1, k)]
+        assert (len(found) == 0) == power_of_two(k) == (not any(parity)), (k, p, len(found), parity)
+        print('PASS k=%d p=%d: %d ports, %d valid star-family decoders; C(k,l) mod 2 for 0<l<k = %s'
+              % (k, p, n, len(found), parity))
+    n, found = valid_decoders(6, 3)
+    assert ((1, 0), (1, 0)) in found, 'the paired-cube identity (level-1 stars, overlap-1 corrections)'
+    print('PASS k=3 includes the paired-cube identity x_T = sum_{c in T} star(c) + sum_{|S cap T|=1} x_S')
+
+
+if __name__ == '__main__':
+    main()