diff --git a/.github/workflows/ym-c316-diagonal-gram-coercivity.yml b/.github/workflows/ym-c316-diagonal-gram-coercivity.yml
new file mode 100644
index 000000000..9c85eacdf
--- /dev/null
+++ b/.github/workflows/ym-c316-diagonal-gram-coercivity.yml
@@ -0,0 +1,78 @@
+name: YM C316 diagonal Gram coercivity
+
+on:
+  push:
+    branches:
+      - verification/ym-c316-diagonal-gram-coercivity-20260820
+    paths:
+      - verification/ym-c316-diagonal-gram-coercivity/*.lean
+      - .github/workflows/ym-c316-diagonal-gram-coercivity.yml
+  pull_request:
+    paths:
+      - verification/ym-c316-diagonal-gram-coercivity/*.lean
+      - .github/workflows/ym-c316-diagonal-gram-coercivity.yml
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  verify:
+    runs-on: ubuntu-22.04
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v4
+      - name: Verify flattened Lean theorem source
+        shell: bash
+        run: |
+          set -euo pipefail
+          SRC=verification/ym-c316-diagonal-gram-coercivity/FaizalShabirDiagonalGramCoercivityFirewall.lean
+          if grep -nE '\b(sorry|admit|sorryAx|axiom|opaque|unsafe|native_decide|Lean\.ofReduceBool)\b' "$SRC"; then exit 1; fi
+          SRC="$SRC" python3 - <<'PY'
+          import hashlib, json, os, pathlib, urllib.request
+          path = pathlib.Path(os.environ['SRC'])
+          source = path.read_text()
+          payload = json.dumps({
+            'content': source,
+            'environment': 'lean-4.30.0',
+            'ignore_imports': True,
+            'mathlib_options': False,
+            'timeout_seconds': 900,
+          }).encode()
+          req = urllib.request.Request(
+            'https://axle.axiommath.ai/api/v1/check',
+            data=payload,
+            headers={'Content-Type': 'application/json'},
+            method='POST',
+          )
+          with urllib.request.urlopen(req, timeout=960) as response:
+            result = json.load(response)
+          receipt = {
+            'source_path': str(path),
+            'source_sha256': hashlib.sha256(source.encode()).hexdigest(),
+            'axle_environment': 'lean-4.30.0',
+            'axle_result': result,
+          }
+          pathlib.Path('/tmp/ym-c316-axle-receipt.json').write_text(json.dumps(receipt, indent=2))
+          print(json.dumps(receipt, indent=2))
+          lean = result.get('lean_messages', {})
+          tool = result.get('tool_messages', {})
+          bad = (
+            not result.get('okay')
+            or bool(result.get('failed_declarations', []))
+            or bool(lean.get('errors', []))
+            or bool(lean.get('warnings', []))
+            or bool(tool.get('errors', []))
+            or bool(tool.get('warnings', []))
+            or 'sorryAx' in json.dumps(result)
+          )
+          if bad:
+            raise SystemExit(1)
+          PY
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: ym-c316-diagonal-gram-coercivity-axle-receipt
+          if-no-files-found: warn
+          retention-days: 30
+          path: /tmp/ym-c316-axle-receipt.json
diff --git a/verification/ym-c316-diagonal-gram-coercivity/FaizalShabirDiagonalGramCoercivityFirewall.lean b/verification/ym-c316-diagonal-gram-coercivity/FaizalShabirDiagonalGramCoercivityFirewall.lean
new file mode 100644
index 000000000..9566b7938
--- /dev/null
+++ b/verification/ym-c316-diagonal-gram-coercivity/FaizalShabirDiagonalGramCoercivityFirewall.lean
@@ -0,0 +1,100 @@
+import Mathlib
+
+/-!
+# Faizal–Shabir diagonal-Gram coercivity firewall
+
+Finite real algebra only.  This file formalizes two obstructions downstream of
+C315's rooted-row theorem:
+
+* the diagonal of a Gram matrix need not be controlled by the full Gram
+  quadratic form (a rank-one Gram family has nontrivial coefficient null
+  directions); and
+* the naive ratio `|C_ij| / C_ii` is not invariant under rescaling one
+  generator, whereas the squared normalized correlation coefficient is.
+
+No Yang–Mills connected-correlation estimate, localized frame theorem, OS
+Hilbert space, regulator/volume uniformity, mass gap, AF/IR theorem, or Clay
+result is formalized here.
+-/
+
+namespace Millennium.YangMills.FaizalShabirDiagonalGramCoercivityFirewall
+
+/-- Rank-one two-vector Gram form on the coefficient direction `(1,-1)` is
+zero.  This is the scalar shadow of two identical physical Gram vectors. -/
+theorem rank_one_gram_difference_is_zero :
+    (1 : ℝ) - 1 - 1 + 1 = 0 := by
+  norm_num
+
+/-- The corresponding diagonal/local quadratic weight on `(1,-1)` is strictly
+positive. -/
+theorem diagonal_weight_survives_gram_null_direction :
+    (1 : ℝ) * 1 ^ 2 + 1 * (-1) ^ 2 = 2 := by
+  norm_num
+
+/-- Consequently no finite scalar multiple of the rank-one Gram value `0` can
+control the positive diagonal value `2`. -/
+theorem no_finite_coercivity_on_rank_one_null_direction (K : ℝ) :
+    ¬ ((2 : ℝ) ≤ K * 0) := by
+  norm_num
+
+/-- Under positive rescaling of the first generator, the naive off-diagonal to
+first-diagonal ratio scales like `r / t`. -/
+theorem naive_diagonal_ratio_under_rescaling
+    (r t : ℝ) (ht : t ≠ 0) :
+    (t * r) / (t ^ 2) = r / t := by
+  field_simp
+
+/-- The naive diagonal ratio can therefore exceed any prescribed nonnegative
+bound by choosing a sufficiently small positive rescaling. -/
+theorem naive_diagonal_ratio_can_be_arbitrarily_large
+    (r M : ℝ) (hr : 0 < r) (hM : 0 ≤ M) :
+    let t := r / (M + 1)
+    0 < t ∧ r / t = M + 1 := by
+  dsimp
+  constructor
+  · positivity
+  · have hne : M + 1 ≠ 0 := by linarith
+    field_simp
+
+/-- In contrast, the squared normalized correlation coefficient is exactly
+invariant under independent nonzero rescalings of both generators.  Squaring
+avoids introducing square roots while retaining the scale-invariant content.
+-/
+theorem squared_normalized_correlation_is_rescaling_invariant
+    (cij cii cjj ti tj : ℝ)
+    (hti : ti ≠ 0) (htj : tj ≠ 0)
+    (hcii : cii ≠ 0) (hcjj : cjj ≠ 0) :
+    ((ti * tj * cij) ^ 2) /
+        (((ti ^ 2) * cii) * ((tj ^ 2) * cjj)) =
+      (cij ^ 2) / (cii * cjj) := by
+  field_simp
+
+/-- If a physical Gram form provides a positive lower Riesz/coercivity margin
+`kappa * diagonal ≤ gram`, then a diagonal-relative harmful debt can be
+converted into a physical-Gram-relative debt.  This is the exact additional
+finite implication missing from the rank-one countermodel. -/
+theorem diagonal_debt_to_physical_gram
+    (debt diagonal gram eps kappa : ℝ)
+    (hkappa : 0 < kappa)
+    (hdebt : debt ≤ eps * diagonal)
+    (hcoercive : kappa * diagonal ≤ gram)
+    (heps : 0 ≤ eps) :
+    debt ≤ (eps / kappa) * gram := by
+  have hdiag : diagonal ≤ gram / kappa := by
+    exact (le_div_iff₀ hkappa).2 (by simpa [mul_comm] using hcoercive)
+  calc
+    debt ≤ eps * diagonal := hdebt
+    _ ≤ eps * (gram / kappa) :=
+      mul_le_mul_of_nonneg_left hdiag heps
+    _ = (eps / kappa) * gram := by
+      field_simp
+
+#print axioms rank_one_gram_difference_is_zero
+#print axioms diagonal_weight_survives_gram_null_direction
+#print axioms no_finite_coercivity_on_rank_one_null_direction
+#print axioms naive_diagonal_ratio_under_rescaling
+#print axioms naive_diagonal_ratio_can_be_arbitrarily_large
+#print axioms squared_normalized_correlation_is_rescaling_invariant
+#print axioms diagonal_debt_to_physical_gram
+
+end Millennium.YangMills.FaizalShabirDiagonalGramCoercivityFirewall