diff --git a/10_hanging_bar.ipynb b/10_hanging_bar.ipynb
index 5630bdd..f80b8c5 100644
--- a/10_hanging_bar.ipynb
+++ b/10_hanging_bar.ipynb
@@ -414,7 +414,7 @@
     "## Takeaway (the 10-minute summary)\n",
     "\n",
     "* On this test, whose **deformation has a closed form**, FEM lands within a few percent of the analytic bar — but all three Newton solvers settle markedly softer at this run's budgets (XPBD ≈ 3.7×, VBD ≈ 3.2×, explicit ≈ 2.0× the FEM tip); the **explicit** solver is the closest of the three and VBD barely improves on XPBD (§1).\n",
-    "* The reasons differ by solver: **XPBD never reaches a force balance** — it leaves a finite equilibrium residual (§3) — because it *projects* positions instead of *solving* R(u) = 0 (mechanical, not a tuning artefact). **VBD** *is* implicit and would approach FEM with enough iterations, but at the budget used its block Gauss-Seidel has not converged on this 17-layer bar, so it stays ~3.2× soft — a convergence-budget effect, not a force-balance one.\n",
+    "* The reasons differ by solver: at any fixed budget **XPBD leaves a finite equilibrium residual** (§3) because it *projects* positions rather than *solving* R(u) = 0 — more iterations do drive that residual down (see `30_convergence`), but the cost to close it grows, so at a real-time budget it reads soft. **VBD** *is* implicit and would approach FEM with enough iterations, but at the budget used its block Gauss-Seidel has not converged on this 17-layer bar, so it stays ~3.2× soft — a convergence-budget effect, not a force-balance one.\n",
     "* That is the trade you are buying: **XPBD is fast on the GPU (§6)** at the cost of an unforced equilibrium; **VBD** is implicit but, at the fast budget here, still settles far soft (closing the gap costs many more iterations — see `30_convergence`); **FEM** is the force-balanced reference. Every diagnostic is computed identically for both sides; the underlying force/energy model is **unit-tested** (`pytest tests/`)."
    ]
   }
diff --git a/30_convergence.ipynb b/30_convergence.ipynb
index 8b7b9c8..7c4ac9c 100644
--- a/30_convergence.ipynb
+++ b/30_convergence.ipynb
@@ -9,7 +9,7 @@
     "\n",
     "The Newton-vs-FEM gap has two different sources, one per solver family:\n",
     "\n",
-    "* **Newton (XPBD)** — *Extended Position-Based Dynamics* — is an iterative positional projection — its *effective stiffness* depends on the number of solver iterations and substeps. We sweep both and watch the equilibrium-residual RMS fall and the tip drop move toward (but not reach) the FEM / analytic value: more iterations lower the residual, yet the XPBD tip plateaus well above FEM (~2.4×) — more compute narrows the gap without closing it.\n",
+    "* **Newton (XPBD)** — *Extended Position-Based Dynamics* — is an iterative positional projection — its *effective stiffness* depends on the number of solver iterations and substeps. We sweep both and watch the equilibrium-residual RMS fall and the tip drop move toward the FEM / analytic value: more iterations lower the residual **monotonically** and the tip keeps descending — at the largest budget tested it is still ~2.4× FEM and falling, and the per-doubling gains shrink, so closing the residual gets progressively more expensive.\n",
     "* **FEM (FEniCSx)** converges under mesh **h-refinement**. We watch the tip drop and strain energy reach a mesh-independent limit, compare it to the analytic 1-D bar (FEM credibility), and confirm the *converged* answer is independent of the number of gravity load increments (a check on the nonlinear solve).\n",
     "\n",
     "This separation — *solver* error (XPBD) vs *discretisation* error (FEM) — is the heart of the Newton-vs-FEM difference. Produce the data with `python -m newton_run.convergence` and `python -m fenics_run.convergence`."
@@ -51,10 +51,16 @@
     "More **iterations** tighten the positional projection (substeps mainly buy stability — the\n",
     "substep sweep here is essentially flat). The **equilibrium residual** is the leftover net\n",
     "nodal force at the settled state — how far the body is from a true force balance (zero for an\n",
-    "exact solver, finite for XPBD); we report its RMS over free nodes, in newtons. It is the\n",
-    "cleanest \"have we converged\" measure: with more iterations it should fall and the tip drop\n",
-    "should move toward the FEM / analytic line — though even at the largest budget the XPBD tip\n",
-    "**plateaus well above** FEM (here ~2.4×); more compute narrows the gap without closing it."
+    "exact solver, finite for XPBD); we report its RMS over free nodes, in newtons. It is the cleanest \"have we converged\" measure: with more iterations it falls\n",
+    "**monotonically** and the tip drop moves steadily toward the FEM / analytic line. At the\n",
+    "largest budget tested (32 iters) the tip is at ~2.4× FEM and **still descending** — the\n",
+    "residual has not bottomed out. The per-doubling gains shrink, so closing the residual gets\n",
+    "progressively more expensive: that escalating cost is the price of a fast positional solver,\n",
+    "and the residual is the gauge for where a given real-time budget lands.\n",
+    "\n",
+    "**Note on the substep sweep:** the `substeps = 4` point is below the explicit stability floor\n",
+    "for this dt — its energy and residual blow up — so it is a stability artefact, not a point on\n",
+    "the convergence trend; read the substep sweep from `substeps ≥ 8`, where it is flat."
    ]
   },
   {
@@ -201,11 +207,13 @@
    "source": [
     "## Takeaway\n",
     "\n",
-    "* XPBD's answer is **iteration-budget dependent**: too few iterations leave a large equilibrium\n",
-    "  residual (it under-resolves the elastic balance). More iterations lower the residual and move\n",
-    "  the tip toward FEM, but it **plateaus well above** the FEM/analytic value (here ~2.4× even at\n",
-    "  the largest budget), and the substep sweep is essentially flat (substepping mainly buys\n",
-    "  stability). More compute narrows the gap without closing it — the price of a fast positional solver.\n",
+    "* XPBD's answer is **iteration-budget dependent**: too few iterations leave a large\n",
+    "  equilibrium residual (it under-resolves the elastic balance). More iterations lower the\n",
+    "  residual **monotonically** and move the tip toward FEM — at the largest budget tested\n",
+    "  (32 iters) it is at ~2.4× FEM and **still descending**, not at a floor; the per-doubling\n",
+    "  gains shrink, so closing the residual gets progressively more expensive. The substep sweep,\n",
+    "  by contrast, is essentially flat (substepping mainly buys stability). That escalating cost —\n",
+    "  not an intrinsic floor — is the price of a fast positional solver.\n",
     "* FEM's answer is **discretisation dependent** but converges monotonically under h-refinement to a\n",
     "  budget-independent equilibrium, and is independent of the load-step schedule. That separation —\n",
     "  *solver* error vs *discretisation* error — is the essence of the Newton-vs-FEM difference."
diff --git a/README.md b/README.md
index 7bd76ea..258b7cc 100644
--- a/README.md
+++ b/README.md
@@ -1,4 +1,5 @@
-# How accurate is NVIDIA Newton, really? — Newton vs. FEniCSx (FEM)
+<!-- headline: confirm wording -->
+# NVIDIA Newton vs. FEniCSx (FEM) — a soft-body accuracy benchmark
 
 A quantitative, apples-to-apples comparison of one **deformable soft body** simulated two ways:
 
diff --git a/docs/METHOD.md b/docs/METHOD.md
index f61673c..19bb966 100644
--- a/docs/METHOD.md
+++ b/docs/METHOD.md
@@ -90,7 +90,7 @@ is the solver. Two mechanisms enforce that:
   the collapsed/degenerate-hexahedron tets with reduced integration + hourglass control
   that explicit codes (e.g. LS-DYNA) often use for "tets". Real P1 simplex tets are known
   to be **over-stiff in bending and near-incompressibility (locking)** because a
-  constant-strain element is too poor to represent those modes, so this is the honest
+  constant-strain element is too stiff to represent those modes, so this is the honest
   like-for-like element, not the most accurate one. → `data/fem_result.npz`
 - **hex** — an *independent* structured hexahedral mesh of the same block geometry
   (`create_box`, trilinear Hex8, full 2×2×2 integration — no reduced-integration
diff --git a/docs/STATUS.md b/docs/STATUS.md
index 42b80b4..df30e56 100644
--- a/docs/STATUS.md
+++ b/docs/STATUS.md
@@ -104,7 +104,7 @@ What is **still numerically open** (quality, not wiring):
 
 | item | what is unsettled |
 |---|---|
-| **SemiImplicit drop** | numerically unstable at this dt/substep budget — the energy blows up (peak strain energy 7536 J vs FEM 3.4 J); the only clearly-broken numeric this run |
+| **SemiImplicit drop** | numerically unstable at this dt/substep budget — the energy blows up (peak strain energy 7536 J vs FEM 3.4 J); the only clearly-unstable numeric this run |
 | **FEM drop** (`fenics_run/run_drop.py`) | dt / damping are not tuned to convergence; the transient numbers are observations, not a settled benchmark |
 | **VBD / SemiImplicit indentation contact** | the `soft_contact` penalty is too soft — the sphere sinks ~33 mm through the 40 mm indent (strain energy ~0.1 J vs XPBD's ~13 J), so **XPBD** is the only Newton solver that geometrically resolves the indentation; the soft-contact stiffness for the implicit/explicit solvers is unsettled |
 | **θ\* fit** (`newton_run/diffsim.py`) | one converged fit characterising the **SemiImplicit** solver vs FEM (θ\* = 1.79); not cross-validated across budgets/scenarios |