diff --git a/Standalone/Research/WorldAssemblerM3/HANDOFF-ABYSS.md b/Standalone/Research/WorldAssemblerM3/HANDOFF-ABYSS.md
new file mode 100644
index 0000000..f302a5e
--- /dev/null
+++ b/Standalone/Research/WorldAssemblerM3/HANDOFF-ABYSS.md
@@ -0,0 +1,78 @@
+# Handoff: speed research moves to abyss (Debian 13, i9-12900HK)
+
+Written 2026-10-19 by the Windows orchestrator (Claude) for the Claude instance on abyss.
+
+## Where things stand
+- Branch `assembler/world`. Rounds S3–S13 committed. Each round's report is `Standalone/Research/WorldAssemblerM3/Round*/REPORT.md`.
+  Latest committed: S13 `87b0e5d` (Round17, a tile-cost breakdown, diagnostics only).
+- **S14 (Round18) is running on the Windows PC.** Its targets:
+  - LOD0 marching cubes costs 243 ns/sample: find out what's inside it;
+  - fix LOD4's column-cache misses (8,224 vs about 1,700);
+  - add shared column pages only if they're still worth it.
+
+  The Windows side finishes, reviews, commits and pushes it. Pull before you touch Core files.
+- **Current in-game numbers** (owner plan, BuildHost actor, 25 cm SurfaceWorld, owner's 5800X3D):
+  - full set about 4.9 s (S7: 17.4 s);
+  - walkable ring about 0.51 s;
+  - LOD0 about 1.09 s;
+  - 40 m/s ring 75/75;
+  - about 29 worker-seconds over 8 TaskGraph background workers.
+- **Open levers, in rough order:**
+  1. the S14 findings;
+  2. a frequency-aware far detail cutoff. A global octave drop failed, and the 2 cm LOD-agreement gate rejects any far
+     simplification, so a screen-space gate would need the owner's OK;
+  3. visibility-aware scheduling in caves;
+  4. RMC batching;
+  5. per-strate voxel size;
+  6. a disk terrain cache.
+
+  The disk cache is approved as a FEATURE (multiplayer-compatible, compression tested), not as a speed fix.
+
+## Owner rules (non-negotiable)
+- **Writable scope:** the VoxelForge repo only. On abyss, ask the owner before installing system packages, UE or
+  toolchains.
+- Never push `main`. Pushing `assembler/world` (and feature branches) is allowed.
+- **Byte-identical output vs earlier rounds is NEVER a gate.** The gates are:
+  - determinism of the current build across worker counts;
+  - no quality regression;
+  - bump the recipe version when output changes.
+- **Speed is judged on the cold path** (nothing cached): walking, flying and 40 m/s running must be smooth on a
+  first visit.
+- Changes go in the shared layer and must benefit every generator family.
+- Luna (Codex `gpt-6-luna`, effort max) implements; Claude orchestrates. Claude doesn't write code itself. Sol costs
+  20× and is used only for short read-only design asks, with the owner's OK.
+- **Briefs:**
+  - investigation rounds are open QUESTIONS answered with measurements, not directives;
+  - give the goal, invariants and evidence bar, not the method;
+  - be kind: credit what the last round got right.
+- Every brief says "Do not commit, push or stash". Claude reviews against the code and captures, then commits
+  specific files (never files over 300 KB, never raw PNGs).
+- **Owner explanations:**
+  - plain language, headline first;
+  - say what he'll SEE or feel in game;
+  - say what's measured vs inferred.
+- Plugin hard rules are in `CLAUDE.md`, plus `CODEMAP.md` and `ARCHITECTURE.md` (§8.10 protected invariants).
+
+## Codex launch (Linux)
+```
+codex exec -m gpt-6-luna -c model_reasoning_effort="max" -c sandbox_mode="workspace-write" -c tools.web_search=true - < brief.md > out.log 2>&1
+```
+- Launch it as a tracked background task, so you're notified when it ends.
+- **On "Selected model is at capacity":** wait 120 s, then resume with
+  `codex exec resume <session id> -m gpt-6-luna ... -` and a short "continue where you left off" on stdin.
+- Watch for runaway processes (more than 40 min) and rounds longer than 6 h.
+
+## First round on abyss: Linux baseline (proposed)
+1. **Build.** Get the Core library (`Standalone/Core`) and its headless benchmark/audit/hash tools building with Clang
+   on Linux. The Windows standalone uses MSVC plus a D3D11 viewer, so the viewer stays Windows-only.
+2. **Baseline.** Record the laptop baseline on its performance cores (plugged in, fixed power profile):
+   - per-LOD tile cost (S13 counters);
+   - standalone full clipmap;
+   - W1/W8/W12 hashes.
+3. **Determinism.** Compare the hashes against Windows (Round17/Round16 list them). A Linux/Windows mismatch is a real
+   multiplayer finding: dedicated servers usually run Linux.
+4. **Later, only if the owner agrees and provides UE 5.7 Linux plus his assets:** headless BuildHost (`-nullrhi`)
+   tests on Linux.
+
+**Not transferable:** cache-size effects (24 MB vs 96 MB L3), thread/slot tuning, smoothness, captures. Those must be
+confirmed on the owner's Windows PC before they're claimed to him.