diff --git a/whitepaper/MD5SUMS b/whitepaper/MD5SUMS
index d8a57cb17..60256e35d 100644
--- a/whitepaper/MD5SUMS
+++ b/whitepaper/MD5SUMS
@@ -6,3 +6,4 @@ f076f22e9e70a94f51884562aad6fdc5  undefect-cwe407-2026-03-25.pdf
 ff52abf9f47a7e6bb25e4519b1325090  undefect-minecraft-enterprise-java-2026-03-24.pdf
 c7fe499eb004271b384a31ac01b38852  undefect-minecraft-enterprise-java-2026-03-25.pdf
 818d29731df88333d29cfdd3eefeb3a2  undefect-minecraft-enterprise-java-2026-03-26.pdf
+a10aed2e5e846290cbe61120d8ef7520  undefect-cwe407-2026-03-27.pdf
diff --git a/whitepaper/full-paper.md b/whitepaper/full-paper.md
index 2747b2375..e6156a416 100644
--- a/whitepaper/full-paper.md
+++ b/whitepaper/full-paper.md
@@ -2149,6 +2149,85 @@ this entirely via `Float32Array` — no GC-eligible objects in the render path.
 to Swing requires making the same guarantee explicitly: pre-allocate all scratch buffers
 as fields, resize only on structural change, never allocate inside the frame loop.
 
+### 19.4 Browser SEW: Two-Manifold Live Demo and Agent Science
+
+The browser workbench (`~/git/cupPCB`) was extended with a split-viewport experiment
+that runs the MOAD and its patch side by side in the same session. The left manifold
+runs the unpatched heat model; the right runs the patched model. Both share the same
+Sym²(X) geometry. Agents (friends) walk both manifolds simultaneously.
+
+#### Two-Manifold Heat Model
+
+| Parameter | Left (MOAD) | Right (patched) |
+|-----------|-------------|-----------------|
+| Injections per frame | 20 × 0.5 | 1 × 0.4 |
+| Diffusion decay | 0.975 | 0.90 |
+| Equilibrium heat | ~4.0 | ~0.3 |
+| z-displacement scale | 120 | 40 |
+| Wireframe color | red | green |
+
+The left manifold reaches ~4.0 mean heat at equilibrium; the right stays near 0.3.
+The z-displacement (vertex distortion) is proportional to local heat. The left manifold
+deforms dramatically; the right stays close to the rest shape. This is the defect made
+geometric: O(n²) heat accumulation vs. O(1) constant throughput.
+
+#### Clock Drift Observation
+
+The two renderers run in **separate `requestAnimationFrame` loops**: the kernel's loop
+drives the left renderer and increments the global `tick` counter; `two-manifolds.js`
+runs its own loop for the right renderer. A HUD overlay shows both frame counters live.
+In practice, the two loops run within 1–2 frames of each other on a single-core browser
+tab (they share the same event loop and are both rAF-scheduled). Drift appears when
+the left manifold's heat diffusion pass (O(n) over all vertices) takes long enough to
+push past the 16ms frame budget — the kernel loop falls behind the twin loop by 1 frame
+per heavy frame. This is a direct measurement of the MOAD's compute tax in the renderer.
+
+#### Friend Temperature Differential
+
+Each agent (friend) has a current vertex index `vIdx`. The HUD reads `heat1[vIdx]`
+(MOAD) and `heat2[vIdx]` (patched) for every live agent and displays both
+simultaneously. At equilibrium, MOAD-side temperatures per agent are 10–15× higher
+than patched-side temperatures at the same vertex. This is the individual-agent view
+of the defect: an agent traversing the MOAD manifold accumulates heat both because
+the manifold itself is hotter and because the agent's own `injectGrowth()` call
+compounds the chaos (+1.0 to `heat[v]` per visit on the MOAD side vs. visit-count-only
+on the patched side).
+
+#### kcjones Agent — Comparative Traversal Science
+
+A special agent, kcjones, was deployed on both manifolds simultaneously with identical
+navigation logic. Its `chooseNext()` scores neighbors by three terms:
+
+```
+score = guide(friends) + heatScore(heat[v] × 2.0) + novelty(unvisited ? 3.0 : 0)
+```
+
+On the MOAD manifold, heat is high everywhere after ~200 frames. The heat term
+dominates; kcjones clusters in already-hot zones, reinforcing them, reducing coverage.
+On the patched manifold, heat is near zero; novelty and friend proximity dominate;
+kcjones spreads broadly, covering new vertices each step.
+
+The `kcjones.locker` command reports the divergence live:
+- `visited` set size: patched side accumulates unique vertices faster
+- `heatLedger`: MOAD side shows top nodes visited hundreds of times (clustering)
+- `heatLedger2`: patched side shows flat visit distribution (broad coverage)
+- `discoveries`: events where kcjones first reached a vertex above heat threshold 2.5 —
+  on the MOAD side these are rare (high threshold, clustered), on the patched side they
+  don't fire at all (heat never reaches 2.5)
+
+The science summary: **the MOAD makes agents cluster where heat already exists, creating
+a positive feedback loop. The patch breaks the feedback: agents explore freely, heat
+dissipates, the manifold stays navigable.**
+
+#### PCB Language — KNOT Container
+
+The PCB NON LINEAR LANGUAGE was extended with a `KNOT`/`TONK` container backed by
+`Set` instead of `Array`. All `contains`/`sniatnoc` operations are O(1) `Set.has()`
+instead of O(n) `Array.includes()`. This fixes the MOAD at the language level: any
+PCB program using a visited-set should use `KNOT`, not `POCKET`. The container
+fix is a one-line substitution — the same one-line substitution documented across
+every ecosystem in this paper.
+
 ---
 
 ## 20. MOADS: The Universal Bottleneck Across the Complete Manifold
diff --git a/whitepaper/undefect-cwe407-2026-03-27.pdf b/whitepaper/undefect-cwe407-2026-03-27.pdf
new file mode 100644
index 000000000..ec88960ff
Binary files /dev/null and b/whitepaper/undefect-cwe407-2026-03-27.pdf differ