diff --git a/Standalone/Research/WorldAssemblerM3/Astra4/PROMPT.md b/Standalone/Research/WorldAssemblerM3/Astra4/PROMPT.md
new file mode 100644
index 0000000..e4ae246
--- /dev/null
+++ b/Standalone/Research/WorldAssemblerM3/Astra4/PROMPT.md
@@ -0,0 +1,37 @@
+# VoxelForge — open read-only review: what should be done and focused on now?
+
+You are an outside reviewer. **Do not edit, create or delete any file, and do not build or run anything.** Read the
+repository and its reports and reason. Be thorough; mark each claim as [CODE] (with file:line), [MEASURED] (from a
+report), or [INFERENCE].
+
+## The project, in the owner's words and goals
+
+- VoxelForge is a UE5 density-field voxel plugin for a **multiplayer** game. The world is made of stacked "strates"
+  (layers) linked by tunnels. The standalone C++ generator is "Core" (`Standalone/Core`), and it is also compiled into
+  Unreal. Read `CLAUDE.md`, `CODEMAP.md`, `ARCHITECTURE.md`, `GDD.md`, `COMPOSER-NOTES.md` and `fable-idea.md` at the
+  plugin root for the game and design intent. The research reports are in
+  `Standalone/Research/WorldAssemblerM3/Linux*/REPORT.md` and the `Astra*/` reviews.
+- Owner goals stated over time:
+  - the near terrain (LOD0–2) ready in ~1 s on arrival, then far terrain keeps filling in at low priority;
+  - same seed gives the same world on every machine (byte-identity with older versions is NOT required);
+  - smooth for VR and high-speed travel;
+  - all generators (surface, floating islands, caves, dreamcore/liminal, etc.) working in Core;
+  - the system should eventually *invent* strates and worlds from tagged authored content plus a seed;
+  - keep things simple and avoid over-complication.
+- The current main line is branch `assembler/world`.
+
+## Your task
+
+Look at the whole: code, reports, design docs, and the direction of the last few dozen rounds. Tell the owner what
+should be done and focused on now, whatever it is. That may be performance, features, generation quality, gameplay
+foundations, architecture, technical debt, tooling, process, or stopping something. Don't assume the recent focus is
+the right one; question it if the evidence says so.
+
+Answer:
+1. **Where the project actually is**: what works, what's fragile, what's missing for the game to exist.
+2. **The top 3–5 things to focus on now**, ranked, each with why, a rough size, what "done" looks like, and how to
+   tell it worked.
+3. **What to stop or deprioritise**, and why.
+4. **Risks or debts** that will bite later if ignored.
+
+Write plain Markdown, at most ~1,500 words, readable by a non-specialist owner (explain jargon briefly).
diff --git a/Standalone/Research/WorldAssemblerM3/Astra4/REVIEW.md b/Standalone/Research/WorldAssemblerM3/Astra4/REVIEW.md
new file mode 100644
index 0000000..4051593
--- /dev/null
+++ b/Standalone/Research/WorldAssemblerM3/Astra4/REVIEW.md
@@ -0,0 +1,72 @@
+[INFERENCE] **Focus now on a small, persistent multiplayer descent, supported by end-to-end Unreal performance work.** VoxelForge has substantial terrain technology. The larger uncertainty is whether it delivers a reliable, interesting shared expedition. Another sequence of tunnel optimizations will not resolve that.
+
+Read-only review of `assembler/world` at `044cd6d`; no files changed and no builds, tests, or project execution. [CODE] identifies repository evidence; design-document references establish intent, not implementation. [MEASURED] means reported results, not independently reproduced measurements.
+
+**1. Where the project actually is**
+
+[CODE] **Core is real and integrated, but generator implementation remains mixed.** Unreal compiles the canonical standalone sources through validated wrappers. Core dispatches to surface, floating-island and native graph evaluators, with a legacy-generator fallback when appropriate. That is useful coverage, but does not mean every family has a complete native implementation. See [VoxelForgeCore.Build.cs:15](/home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForgeCore/VoxelForgeCore.Build.cs:15) and [ShapeWorldInternal.h:600](/home/jahni/VoxelM/Plugins/VoxelForge/Standalone/Core/src/ShapeWorldInternal.h:600).
+
+[MEASURED] **Passage correctness and speed have improved substantially.** L109 simplified tunnels and removed support-restoration machinery, reducing floating-island and tunnel-network near-generation times by approximately 31–38%. L125 reports 21,425/21,425 open route samples, a maximum sampled floor offset of 7.884 mm, and matching cross-compiler/worker hashes. These are meaningful engineering achievements. [L109](Standalone/Research/WorldAssemblerM3/Linux109/REPORT.md), [L125](Standalone/Research/WorldAssemblerM3/Linux125/REPORT.md).
+
+[MEASURED] **The one-second player experience is not achieved.** L125’s standalone near-ready medians were approximately 0.78–1.03 seconds. L127’s Unreal startup took 2.22–2.51 seconds; at 30 m/s, median near coverage fell from 43.36% to 40.22% after the smooth-tunnel merge. Both versions recovered after stopping. L130 found no measurable timing improvement and skipped Unreal movement. These are different workloads; subtracting their times would not isolate a particular overhead. [L125](Standalone/Research/WorldAssemblerM3/Linux125/REPORT.md), [L127](Standalone/Research/WorldAssemblerM3/Linux127/REPORT.md), [L130](Standalone/Research/WorldAssemblerM3/Linux130/REPORT.md).
+
+[CODE] **Multiplayer foundations exist; the multiplayer lifecycle remains unfinished here.** There are season hashes, local terrain-edit entry points, and streaming-anchor infrastructure. The architecture explicitly leaves join replication, authoritative edit messages, late-join snapshots and networked season changes open. The edit entry point directly dispatches to local modification paths. [VoxelWorld.h:574](/home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Public/VoxelWorld.h:574), [VoxelWorld.cpp:15771](/home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:15771), [ARCHITECTURE.md:2408](/home/jahni/VoxelM/Plugins/VoxelForge/ARCHITECTURE.md:2408).
+
+[CODE] **The composer is more than a proposal, but less than the full content vision.** It rolls candidates, measures them and rejects failures. The intended system additionally combines tagged geometry, materials, decorations, creatures and audio; theme coherence remains an open design item. [VoxelSeasonManifest.cpp:2523](/home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelSeasonManifest.cpp:2523), [COMPOSER-NOTES.md:1512](/home/jahni/VoxelM/Plugins/VoxelForge/COMPOSER-NOTES.md:1512).
+
+[INFERENCE] **This repository demonstrates a capable terrain platform, not yet a validated game loop.** I cannot establish what gameplay exists outside the plugin. The GDD itself identifies “Is the descent fun?” as its first unresolved question. [GDD.md:441](/home/jahni/VoxelM/Plugins/VoxelForge/GDD.md:441).
+
+**2. Ranked priorities**
+
+All sizes below are [INFERENCE]: rough focused engineering estimates, excluding substantial art production.
+
+**1 — Deliver one multiplayer expedition that survives reconnecting. Size: 3–6 weeks.**
+
+[INFERENCE] Use two contrasting strates and their connecting tunnel. Two players descend, find one destination, uncover or collect something, return, and retain their changes after restart. Include a third player joining after terrain has changed. Avoid expanding immediately into the city economy or community backend.
+
+[INFERENCE] The necessary foundation is a server-approved world identity, ordered authoritative edits, a saved snapshot plus subsequent edits, and collision around separated players. Give edits stable identifiers so retransmission cannot apply them twice. Reuse the existing local edit and anchor systems.
+
+[INFERENCE] **Done:** host/client edits agree; a late joiner reconstructs the same terrain; restart preserves it; incompatible world versions fail clearly; remote players have ground when away from the host.
+
+[INFERENCE] **Success:** a complete shared expedition without terrain disagreement, lost progress or collision failures—and players can explain what made them want to descend again. This tests the project’s central assumptions much earlier than a complete RPG would.
+
+**2 — Close the Unreal arrival and movement gap. Size: 1–3 weeks initially.**
+
+[MEASURED] L122 found idle generation capacity while admission was blocked; L123 measured approximately 19 ms from region request to game-thread observation despite sub-millisecond median compilation. Simple lookahead and admission relaxations failed their combined coverage/frame-time gates. L121’s modest extra far-apply lane did improve far coverage. [L121](Standalone/Research/WorldAssemblerM3/Linux121/REPORT.md), [L122](Standalone/Research/WorldAssemblerM3/Linux122/REPORT.md), [L123](Standalone/Research/WorldAssemblerM3/Linux123/REPORT.md).
+
+[INFERENCE] First establish one current-build, rendered Unreal baseline on intended hardware. Measure arrival through **visible terrain and usable collision**, then walking, 30 m/s travel, turns, teleporting and edits. Record near coverage, far gaps, frame-time spikes and memory separately. NullRHI measurements, which omit rendering, cannot certify VR comfort.
+
+[INFERENCE] Investigate region completion, admission and publication before another broad density rewrite. A game-thread continuation is a candidate experiment, not a proven solution; keep world mutation on its owning thread.
+
+[INFERENCE] **Done:** approximately one-second LOD0–2 readiness on the agreed hardware, safe collision, steadily filling far terrain, and an explicit movement-coverage target without worsening frame spikes. **Success:** repeated rendered runs meet those limits, including with remote-player collision work. Do not redefine “ready” as only the nearest handful of tiles.
+
+**3 — Turn generator breadth into playable variety. Size: 2–4 weeks for the first content slice.**
+
+[INFERENCE] Establish a compact capability matrix for each family: Core dispatch, entrances/exits, interior connectivity, edits, materials/content, far terrain and performance. Include dreamcore/liminal explicitly; motif generation alone does not establish a convincing anomalous place.
+
+[MEASURED] L58 illustrates the distinction: its lattice route passed reachability checks, but the review described a corkscrew tunnel rather than the intended open vertical shaft. Technical validity and identity are separate acceptance criteria. [L58](Standalone/Research/WorldAssemblerM3/Linux58/REPORT.md).
+
+[INFERENCE] Extend the existing offline composer with a small tagged library: a grounded cave, an island environment and one liminal identity, each with coherent materials, sound and a destination. Preserve owner review before publishing a season.
+
+[INFERENCE] **Done:** several seeds produce recognizably different, coherent places with findable progression and usable encounter spaces. **Success:** players navigate them without developer guidance and distinguish their identities without reading labels. Expand families after this works.
+
+**4 — Consolidate the evidence and remove obsolete complexity incrementally. Size: 2–4 days, then ongoing.**
+
+[INFERENCE] Maintain one short current-state page: shipped behavior, open defects, active baseline, hardware and next acceptance test. Keep historical reports as history. Remove rejected experimental paths when they no longer serve a specific comparison; extract stable responsibilities when touching oversized files.
+
+[INFERENCE] **Done:** a contributor can identify the current evaluator, runtime path, multiplayer gaps and performance limits without reconstructing 130 rounds. **Success:** new work no longer repeats rejected experiments or treats old measurements as current.
+
+**3. Stop or deprioritise**
+
+- [INFERENCE] **Pause open-ended tunnel micro-optimization.** L125/L126 missed the older speed bar; L130’s reach correction produced no measurable speedup. Keep correctness fixes, but require a current cost account and a bounded experiment before further optimization. [MEASURED: L126](Standalone/Research/WorldAssemblerM3/Linux126/REPORT.md), [L130](Standalone/Research/WorldAssemblerM3/Linux130/REPORT.md).
+- [INFERENCE] **Do not implement Astra recommendations mechanically.** Astra1’s support-heavy model predates L109’s simplification. Astra2/3 offer hypotheses, not measured wins for today’s code. Re-profile before adopting them.
+- [INFERENCE] **Defer GPU generation, a mesher replacement, elaborate discovery systems and large-scale backend work.** None currently resolves the missing shared expedition.
+- [INFERENCE] **Stop requiring old-version byte identity for deliberate generation changes.** Preserve peer agreement within a version and explicit versioned saves; permit reviewed geometry improvements.
+
+**4. Debts that will bite later**
+
+[INFERENCE] **Determinism needs a deployment contract.** Linux Clang/GCC and worker-count agreement is strong evidence, not proof across every intended machine. Include Windows/compiler/CPU combinations actually supported. World identity must cover evaluator, content, layout, scale and edit state—not merely seed. [MEASURED: L130](Standalone/Research/WorldAssemblerM3/Linux130/REPORT.md).
+
+[CODE] **Scale and fallback complexity remain active integration risks.** Core is opt-in, and runtime queries bridge different coordinate conventions. [VoxelWorld.h:300](/home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Public/VoxelWorld.h:300), [VoxelWorld.cpp:15612](/home/jahni/VoxelM/Plugins/VoxelForge/Source/VoxelForge/Private/VoxelWorld.cpp:15612). [INFERENCE] Test edits, content placement and collision across those boundaries before broadening authoring.
+
+[INFERENCE] **A passing route audit does not prove a whole strate is enjoyable, connected or suitable for six-player combat.** Retain geometric checks, but add actual traversal, destination discovery and encounter-space evaluation. Those are now more valuable than another polished tunnel benchmark.
\ No newline at end of file