Created all three required PNGs, left untracked:

• artifacts/living/model.png — 1536×1024 model sheet.
• artifacts/living/characters.png — 1920×1920 transparent atlas, 36 sprites.
• artifacts/living/world.png — 2560×1280 transparent atlas, 32 props covering four biomes.

Validated PNG integrity, transparency, gutters, anchors, and reproducible packing. Provenance, license, mappings, and limitations are in docs/living/README.md.

Remaining visual work includes directional action coverage, fishing-line cleanup, smoother animation, and seamless shoreline variants. Independent visual approval remains pending.

I've built the living-world simulation backend and checked it on this machine. All 180 backend tests pass, 13 of them new. Nothing was deployed, the live VPS was never accessed, and I made no commits; all changes are in server/, shared/, tests/ and docs/.

What residents can now do (all choices still go through the existing learned planner and mb-v1; nothing is scripted):

• Fishing: fish live on shoreline water in all four biomes and are only seen from within 3 tiles of the shore. Fishing needs a line, which residents can craft from fibre and sticks. Without one, an attempt fails and leaves the fish alone. With one, it takes 12 ticks (cast, wait, reel) and some energy, and lines wear out after 6 uses. A stock holds at most 6 fish and regains one every 450 ticks.
• Shelters: each biome has its own design and materials: timber homes, desert shades, snow refuges and coast huts. Construction stages follow the materials actually delivered, and labour can't get ahead of them. Unfinished sites and ruins give no protection. Shelters wear down over time, faster in cold and rain, and repair uses up a building material.
• Shared stores: each finished shelter stores up to 40 items. Residents only know a store's contents from their own last visit; if it's been emptied since, they find out on arrival. Taking goods out and putting them back earns no reward.
• Climate: walking in desert heat costs 50% more energy. Shelters keep out the cold and shade the heat, less well when worn. The existing heat-thirst and cold-exposure costs remain.
• Births: a birth at a shelter now needs room in that household (fewer than 8 people calling it home). This only limits conception; nobody is removed.
• What the renderer gets: every resident now reports a state, phase, progress and carried load taken straight from the engine. For example, "fish" is only reported while a resident is fishing with a line, so the art can't show casting otherwise. The new frame fields are additive and the version number is unchanged, so the current site keeps working. A real sample frame is in shared/fixtures/living-frame.json.

Existing worlds: I captured a genuine old-format (v3) save from the pinned engine as a test fixture. On a copy of it, the upgrade keeps the world ID, residents, brains and learned weights, memories and relationships unchanged; a hash check confirms it. The original file is kept byte-for-byte and survives later saves. Restarting gives the same result as running straight through, and rolling back restores the exact old bytes. The owner's upgrade tool, server/bin/expanse-upgrade.js, only checks unless given --write. No live data was available to me, so the real world itself has not been migrated or tested.

Results across seeds (full numbers in server/docs/LIVING.md and tests/results/living/):

• Six seeds at 8,000 ticks: 78–135 residents at the end, 25–63 fish caught each, 34–55 standing shelters, most of them outside the original settlement.
• Over 24,000 ticks, residents settle in all four biomes, and cold exposure becomes the main cause of death.
• Timing: at 250 residents a tick takes 22 ms at the 95th percentile, against 13.7 ms for the old engine. That is about 2% of a core at the 1-second tick.

Problems and unfinished work:

• Memory is the real risk for the small VPS. Population is limited only by food: one seed reached 602 residents and 655 MB of memory by 24,000 ticks. The old engine also grows (380 residents, 521 MB). I didn't add a hard cap because the brief forbids culling, so the owner needs to check the VPS's memory or pick a limit.
• I only measured on x86_64. The ARM VPS is unmeasured.
• Residents still fail at fishing while they learn, about 20 bare-handed attempts each, and newborns start from scratch. It was about 115 each before I made learned failure weaken the hunger drive. That change applies to every action, not just fishing.
• **The website doesn't show any of

Implemented in dist/: new Stratok art, exploration across all biomes, a real isolated engine demo, state-based animation and responsive inspectors.

Passed 28 frontend tests, 41 backend tests, 12 browser layouts and 16/16 design checks. Delivery totals 7.85 MiB. VPS untouched.

Validation, screenshots, recordings and remaining limits.

Saved .imd-findings.json with {"findings":[]}.

Reviewed all 23 exported files and corresponding source. No substantiated hosting blockers or wallet/payment flows found. Artwork provenance matched supplied inputs; demo, recorded and live-failure labels passed local browser checks. Production remained untouched.

Revision needed: three high and two medium findings. Saved with exact source citations, fixes, and reproductions in .imd-findings.json.

• High — bug: Mobile tile evidence overlaps the HUD; its close button becomes unreachable in landscape.
• High — bug: Camera limits make valid world-corner tiles impossible to view or pick.
• High — fidelity: Highlands and banks lack the requested terrain relief.
• Medium — bug: Actual repairs never select the work animation because the frontend expects the wrong phase.
• Medium — bug: The settlements navigator is headed “Resident inspector.”

Verification: 41 backend and 11 frontend tests passed. Chromium and WebKit passed resident interactions, local backend access, and connection-error recovery. Independently simulating 12,000 ticks reproduced the shipped demo checkpoint byte for byte. Implementation files remain unchanged.

The visual idea is a living isometric storybook world of golden, leaf-crowned residents, framed by a compact moss-and-gold observer interface.

The complete requirement checklist follows. Short frontend filenames refer to web/src/expanse/; engine filenames refer to server/src/expanse/; captures refer to artifacts/living-interface/.

Review complete. The findings file is written with four substantiated defects, none critical.

What I checked

• Ran the schema-4 backend suites (living and living-interface, 16 tests) and they pass, including conservation across seeds, fishing prerequisites and regrowth, store capacity, climate costs, exploration beyond the settlement, determinism/restart and the real schema-3 fixture migration with pre-migration backup and rollback.
• Read the full engine, terrain generator, migration code, view frame and the renderer's state-to-pose mapping.
• Inspected the supplied character sheet and the temperate desktop capture. The species is the golden scalloped leaf-crown design with no ears, muzzle or tail. The map shows distinct terrain, lake, forest, homes and log piles, not a repeated cabin grid. The shipped demo worker bundles the schema-4 engine, not the schema-3 one.
• Confirmed no file outside the findings file was changed.

Findings, ranked

1. Medium, engine.js line 200. Residents learn a shelter store's exact contents just by seeing the shelter within perception range, and are offered retrieve actions for it. The docs state stores are known only from a resident's own last visit. Reproduced: one call to perceive on a creature 5 tiles from a never-visited stocked shelter fills its memory with the full stock.
2. Low, appearance.ts line 19. The repair work pose requires phase "work", which the engine never emits. It sends "repair". Repairs always render as carry or idle.
3. Low, engine.js line 457. Shelters decayed to integrity 0 still accept deposits and retrievals and appear in the client's finished-shelter list. Reproduced with a direct deposit into a ruin.
4. Low, engine.js line 728. A build step chosen this tick is reported as "idle" rather than "starting", unlike every other primitive.

Not found No rabbit or rodent art, no fabricated activities in the state mapping beyond the gaps above, no save reset in the migration path, and no production writes in the delivered scripts. Economy conservation, fish reachability and tool gating hold under the shipped tests and my probes. I did not have time to replay the biome recordings frame by frame or to measure browser timings independently.