Graphics experiments

Clean assets only: every model on screen comes from a verified clean source. Each video is published with a provenance manifest listing every file the recording loaded, and the publisher refuses it if any is flagged.

What to look at: at 25% health the player falls after about 18 seconds; at 50% and above the arena is cleared in about 31 seconds. The arena itself is still a blockout (plain pillars and blocks); detailed walls and boundaries are planned.

• 12first cohort complete
• 1.90MKing source triangles
• 179,990working-mesh triangles
• 0 / 2transfer proofs passed

What passed: the immutable provider GLB remains preserved; a derived 10.3 MB working mesh completed in 66.87 seconds with 3.89 GiB peak RAM and zero swap. Fixed-view review retained the crowned silhouette, heraldry, hands and feet.

Why this version failed: semantic ownership at articulation boundaries is not yet safe. Face plates, pauldrons, gauntlet and shin rims cross joint regions, and the pelvis assignment captures only the codpiece tip. Rigging that mesh would make hard armor stretch.

Promotion rule: exact-coordinate topology cleanup comes before PBR bake, skeleton fitting and animation. The same versioned biped-mech recipe must pass both King and independently generated Nexus Prime without subject-name branches or manual mesh edits. Until then these Prism candidates remain quarantined interim builds, not shipping replacements.

• 19Faction style anchors
• 406Canonical identities queued
• 282Humanoid faction passes
• 124Mechanical identities
• 492Component sheets planned
• 331Motion sheets planned

Verdant lineage · Beacon

Biotech rescue vessel

Raider lineage · Behemoth

Brutalist assault vessel

What is public: these eight reviewed checkpoint boards only. The existing 965-artifact mechanical campaign remains private; the new 406-identity normalization ledger adds 282 humanoid and 124 mechanical subjects, with 31 cast overlaps retained as aliases rather than duplicate model jobs. Prompts, receipts, source anchors, and rejected iterations remain private until individual assets clear modeling and publication review. These are source concepts, not shipping models.

Latest head · gaps filled v11 — hollow skull cavity closed

Full-body assembly reference · 100 real component roots

Body v18 preserves the corrected complete hands from v17, replaces two floating flank vents with continuous textured side-armor plates, and adds paired diagonal rib braces. The heraldic chest, crown, rear spine, shoulders, legs and all other component records remain pinned. Actual-mesh assembly demonstration, not a completed character. 99 components separate around the stationary thorax and return to their recipe positions. Fixed front, quarter, profile and rear renders pass the static fit review; animated clearance, collision response, flexible cables, gameplay integration and final AAA acceptance remain open. The body-v14 film below predates the hand and torso corrections.

This section is deliberately static while the gameplay rig is rebuilt around a continuous deforming underbody. The 93.97 MB source-resolution assembly remains the accepted visual reference, not an animation-ready body. Download source-resolution body · 93.97 MB · Front · Three-quarter · Side · Rear

Assembly film · 98 real component roots

The authored demonstration uses 98 real component roots, including the new openwork outer crown and recessed thigh and shin hardware; 97 move around the stationary thorax. It is not collision- or world-solved and is not AAA acceptance. The palm-v7 film below is an earlier 96-part motion sample, not this assembly or the newer action rig.

Gameplay animation rebuild · one reusable skeleton, 102 armor grafts

Pose reference

Compare the target beside the actual model. Tap the image to inspect it full size.

Connected core · 4.37 MB Armored King V20f · 32 clips · 46.24 MB Equipped pistol · 42.81 MB Motion Loading motions… PlayPause Phase Loading the armored gameplay candidate…

Walk plays by default. Selecting a pistol motion loads the new V18 equipped candidate in this same viewer; selecting another family returns to V17c. Drag Phase to pause and inspect. The concept image is the target, not a rendered result.

Same skeleton, two inspectable exports. The cyan view is the single connected, watertight deformation core by itself. The armored view grafts all 102 rigid component roots onto named Rigify bones and sockets; the armor geometry is not smoothly skinned or decimated for motion. Both expose the same 32 named baseline clips. Only the separate assembly/adornment sequence may animate plate-local transforms.

New V19 equipped pistol · fixed geometry + five braced motions. Aim plus kinetic, laser, plasma and railgun fire use a fitted real pistol, thumb-up palm-seated grip and A01-referenced recoil/recovery, with the V18 braced stance kept intact (torso bladed 20° toward the aim line, off-hand braced across the lower chest). Geometry fix (V19): the two non-concept flank-armor and two rib-brace side panels were removed (they were extraneous side pieces absent from the concept), and six forged dark interior connectors — a torso understructure core, two chest-flank fillers, two hip gussets and a cranial core — are rigid-mounted to their owner bones to back the torso and head seams. The deforming underbody skin and all five clips are unchanged. This is a visual-review candidate, not physics or gameplay acceptance: exact solid contacts, full-body clearance, runtime LOD and the in-game runtime rebuild remain open. Download fixed equipped five-motion GLB · 42.81 MB · Compare the A01 motion reference.

V17c is a four-motion review candidate, not a completed action library. Walk now adds subtle chest twist and arms driven opposite the lead foot, without changing foot paths. Walk/run knees bend forward with much less lateral travel; fitted forearm roll points the thumbs forward. Only animation payloads replace the V16f clips: geometry, materials, skin and attachments remain unchanged. All 344 armored front/side frames were captured; all 58 revised walk frames were inspected against the concept, with prior run and hand evidence retained. Contact, loop, clip-switch and neutral-hand preservation checks pass. The attempted finger curl separated segments and was rejected: this export preserves the connected neutral finger pose, not a finished grasp. Greeting remains a bent-elbow wave. Stronger run flight/weight transfer, fitted finger pivots, firearm grips, remaining actions, collision response, gameplay LOD and final AAA polish remain open.

V15 remains rejected. It exposed all 32 clips but separated the high-detail body into fractured parts. V16f replaces that invalid local-channel copy with rest-relative rotations, fitted target-space attachments and rigid helper ownership.

Download connected core GLB · Download armored V20f four-motion GLB · Previous V17c 32-motion GLB · Accepted visual target · Static source-resolution assembly

Archived old-hand film · rig-v11/action-v10

Archive—not the model above. This film predates the corrected v17 hands and body-v18 torso armor.

Earlier rig-v11/action-v10 film: 32 actual exported poses, frames 0–31 at 24 fps. It is authored kinematics, not collision-solved motion.

Earlier 96-part motion sample · 0.53 MB, no model download

31 actual poses from the earlier 96-part rig, 1.29 seconds. This sample covers frames 0–30 and omits final frame 31; it is not the newer interactive rig or a complete recovery demonstration. No collision-solved motion or final animation acceptance is claimed.

Subsystem bench · actual intermediate models and closeups

Component Foundry · individual crown sources

Three independently forged source parts, ready to orbit and zoom. These pre-assembly versions remain available for close inspection. Small surface detail will use textures and shaders; silhouette and articulation stay in geometry.

Fleet component sources · A01, A02 and Aegis-7

New separately forged parts, not assembled characters. These are source-resolution inspection downloads, not gameplay LODs or admitted shared components. Concepts were AI-generated with Google Antigravity; isolated drivers and neural models are project-generated. Heavy GPU stages run sequentially within measured memory limits while concept and CPU review work overlap.

Earlier baseline evidence · not the replacement model above

Forged rear mechanism · integration iteration

First real component export · structural WIP

Five-layer technical proof · current evidence

1. Immutable gameplay WorldFrameTick, contacts, support, target, velocity, impact, recoil, power, damage, acceleration, and atmosphere enter one typed snapshot.
2. Bounded deterministic physicsRoot, right-chain, palm, cable, piston, contact, and material responses are deterministic; invalid input preserves authored timing through a zero-correction fallback.
3. Typed component and PBR contractSixteen pinned focus components plus body remainder carry sockets, ownership, collision, provenance, and required channel declarations. Four texture channels remain physically unbound.
4. Front, rear, and right-arm art directionThe committed references above are the pixel targets. The first forged rear rail/actuator is mounted; broad right-arm shells and the remaining component replacements are visible gaps, not accepted fidelity.
5. Source-aware export and runtime animationThe exact source-derived GLB reaches the real loader. A 1200 ms neutral → acquire → reach/aim → impact → recovery sequence drives the live skeleton and returns to neutral.

Exact structural checkpoint: 16 nonempty focus components plus the body remainder retain stable typed identities. Recipe 12f7765b…; structural hash c24f2a87…; v4 export 459981c6…. The rear-yoke geometry is now pinned to a separately forged mechanism source, and two clean v4 builds match byte-for-byte. The exact local review GLB remains an experiment, not an AAA or playable asset.

• Real loader: partial green — exact export/source hashes accepted without fallback in the explicit proof path.
• Component assembly: partial green — all 17 exact mesh identities round-trip through deterministic dependency groups; remaining mechanisms still need forged replacements.
• Physics: partial green — live bounded balance, aim, recoil, same-tick contact, cable and rear-piston response pass; continuous-motion pixels and gameplay input latency remain pending.
• PBR: partial green — 16 complete standard PBR surfaces load and one semantic energy surface responds through 17 stable material instances; authored material refinement remains pending.
• Measurements: pending strain, foot drift, penetration, response latency, and Compatibility frame-time evidence. Acceptance receipts now reject every over-budget value.
• Two rebuilds: partial green — two clean exports are byte-identical to each other and the committed GLB; matched visual-delta evidence remains pending.

Published evidence: R2 fetch-back verification for the concept/current-rig media is in docs/android-king-aaa/evidence/publication-v1.json; exact v4 model/still hashes and the live loader receipt are in docs/android-king-aaa/evidence/forged-rear-v4.json. This article does not change the accepted static-cast card below.

Raw sourceBASELINE

12.4M tris · 42.8 MB (meshopt). Unoptimized forge output. Every surface detail present; far too heavy to ship or stream — and the meshopt/Draco packing that gets it to 42.8 MB does not load in Godot 4.7.2.

Godot runtime · LOCKEDPASSES · KING BAR

2.18M tris · KTX2 · 112 MB. The first shipping Godot-runtime build (art-gen/godot_compress.sh: gltfpack -noq -tc -si 0.0395 — error-bounded simplify, never -sa; the error bound halts decimation at ~2.18M rather than shattering the plates). Rendered in-engine at the ladder camera: plates read clean, gold heraldry, crown and filigree fully intact, no shatter. A passing tier — high geometry + KTX2 textures, not decimation — now beaten on size by the multi-material bake beside it.

Multi-material bake · LOCKEDPASSES · SMALLER

400k tris · KTX2 · 81.5 MB. The multi-material normal bake (art-gen/king_multimaterial_bake.py): per-zone tangent normals baked from the 12.4M high-poly onto a 400k low-poly, keeping each zone's own albedo + metallic-roughness. Rendered at the ladder camera: plates read clean with baked panel-line depth, the heraldic lions, crown and gold filigree stay crisp, no shatter — and the surface gate now reports a normal map on 44 materials (the locked tier has none). 27% smaller than the 112 MB locked build at the same King bar — the "smaller AND amazing" hero result. Further shrink to the ~20–40 MB target needs deduping the per-instance normal maps.

80k decimateREJECTED · SHATTERS

80k tris · 23.7 MB KTX2 / 26 MB WebP. Verified in the matched head render: the hard-surface plates shatter into crumpled chrome and the gold heraldry degrades. Cutting 12.4M→80k triangles destroys the surfaces. Fails the Android King bar.

60k decimateREJECTED · SHATTERS

60k tris · 23.6 MB KTX2 / 25.9 MB WebP. Verified in the matched head render: the hard-surface plates shatter into crumpled chrome and the gold heraldry degrades. Cutting 12.4M→60k triangles destroys the surfaces. Fails the Android King bar.

40k decimateREJECTED · SHATTERS

40k tris · 23.6 MB KTX2 / 25.8 MB WebP. Verified in the matched head render: the hard-surface plates shatter into crumpled chrome and the gold heraldry degrades. Cutting 12.4M→40k triangles destroys the surfaces. Fails the Android King bar.

Lean 60k · 1K/512REJECTED · SHATTERS

60k tris · 1K albedo / 512 maps · 11.1 MB. Same shattered 60k geometry with 1K/512 textures — small, but it fails the King bar for the same reason. Not a shipping tier.

Normal-bakedREJECTED FOR THIS ASSET

No head crop published: the single-atlas bake broke the albedo before a detail pass was worth capturing.

60k tris · single-atlas bake · 1.5 MB. Smallest by far, but the King's gold and heraldry collapse under one baked atlas. A failed approach for an albedo-heavy hero — kept here as evidence, not a shipping tier.

• 1 · Size is textures, not triangles. Geometry meshopt-compresses to under 1 MB across every tier. Dropping from 80k to 40k triangles barely moves a ~24 MB file — the texture set carries the weight, so texture resolution is the real lever (lean 60k proves it: 26 MB → 11.1 MB).
• 2 · Normal-baking helps geometry heroes, hurts albedo heroes — if it is one shared atlas. Baking high-frequency geometry into a normal map pays off for detail-driven assets, but the earlier single-atlas bake merged all 51 material zones into one image and collapsed the King's gold and heraldry — the smallest tier and the worst-looking one. The fix is a multi-material bake: normals baked per zone onto a low-poly, keeping each zone's original albedo + metallic-roughness untouched. That landed — it is the "Multi-material bake" passing tier above (81.5 MB), not the single-atlas result shown below.
• 3 · Godot loads neither meshopt nor Draco at runtime. Verified against the 4.7.2 binary: no EXT_meshopt_compression support and no Draco decoder, so any glb that requires either fails to parse (ERR_PARSE_ERROR). KTX2/basisu and WebP textures both load. Two builds per asset, then: web = meshopt + KTX2 (smallest); Godot runtime = uncompressed geometry + KTX2/WebP textures. The earlier "meshopt tier renders in Godot" claim could not have been true.
• 4 · The passing approach is locked: keep geometry high, let textures carry size. Every 40–80k decimated tier shatters the plates (verified in-engine) — hard decimation is a dead end for this hero. The LOCKED tier keeps geometry at the error-bounded ~2.18M-triangle floor with KTX2 textures and passes the King bar in-engine. And smaller-and-still-amazing is now solved: the multi-material normal-bake tier holds the same bar (clean plates with baked panel-line depth, gold + heraldry + crown intact, no shatter) at 81.5 MB — 27% smaller — a 400k low-poly whose detail comes from per-zone baked normals, not decimation.
• 5 · Honest size: two Godot builds pass; the geometry-heavy one is bigger, not smaller. Godot 4.7.2 cannot load meshopt or KHR_mesh_quantization, so the geometry-heavy LOCKED build (112 MB) is larger than the meshopt RAW forge output (42.8 MB, which will not open in Godot); the three.js WEB build (meshopt + KTX2) lands around ~75 MB. The multi-material bake is the smaller Godot-loadable pass at 81.5 MB. Getting the hero all the way to the ~20–40 MB target is the remaining stretch — it needs deduping the per-instance normal maps (the King has ~110 instanced zones, many sharing geometry) so identical normals are not stored twice.
• 6 · The fleet is already small AND amazing — the King hero is the geometry-bound outlier. Where "lighter + amazing" is fully achieved: the 23 mechs compress to 2.4–7.4 MB and the 22 finished characters to 0.5–4.9 MB, all verified non-shattered against the bar. See the compressed builds gallery →. Only the ultra-dense King hero sits at 112 MB; the volume assets are the win.
• 7 · Gold-on-dark ornate assets ship WebP albedo, not ETC1S. KTX2/ETC1S crushes gold-on-dark heraldry into dark, fractured, speckled albedo (verified on the apotheosis + cathedral capital ships). The fix is full WebP textures (GODOT_COMPRESS_ALBEDO_WEBP=1 in godot_compress.sh), which stays bright and still loads in Godot 4.7.2 via EXT_texture_webp.

Scope: a real optimization pass on one hero model, rendered at matched camera/lighting per tier. The tiers are measured build outputs, not concept art. Two tiers pass: the geometry-heavy LOCKED build and the smaller multi-material normal bake (81.5 MB); the single-atlas normal-baked entry is a rejected result kept as evidence. Open: deduping the per-instance normals to reach the ~20–40 MB stretch target, broader per-cast rollout, runtime LOD selection and in-engine frame-time evidence.

Review loop: capture the real scene → generate a reference-based concept → implement one bounded change → capture the same camera/seed → compare concept, before and after → name the gap and refine. Maximum three candidates before recording a specific non-convergence follow-up. Concept art alone is never scene-completion evidence.

Acceptance evidence: reachable startup and player interactions, navigation/collision, scene identity, visual readability, and median/p95 frame time on recorded hardware. Bridge movement requires authoritative ship-state evidence, not a moving background. No production approval is implied.

Next environments: planets; asteroid fields; stars and nebulae; mountains, ridges, rivers and lakes; trees and forest edges; rural sci-fi settlements; urban dwellings, streets and infrastructure (tasks 973–979). The opening pilots lead this initiative.

Evidence delivery: future concept/before/after images load automatically near the viewport. Published media uses hash-verified R2 assets; no placeholder artwork is presented as a finished engine scene.

Gold standard: the approved A01 dodge/slide and melee concepts below guide silhouette, joint motion and attack intent. Pistol and revised rifle sheets extend the target set. These are AI-generated concept images, not screenshots of completed animations. Compare the current playable rig trials → · Browse the scene match + concept library →

Latest equipped review: A01 conventional pistol and A01 shouldered rifle now start with firing playback. These are new partial custom-body trials; golden-body baseline versus fitted comparisons remain separate. No batch rollout has been approved.

More upright torso · partial slide: the right hand now reaches in front of the torso. At the reviewed glide phase, the torso axis is 24.78° from vertical, previously 33.33°. The low pelvis, extended leading leg and folded trailing leg remain visible; lower-body transforms shift by up to 18.5 mm. The support hand remains a closed hovering fist, with no verified palm contact. Only this version's side glide pose has been compared; entry, recovery and full playback quality remain unapproved. Open exact slide GLB · 7109119ec815

Motion PausePhase

Loads automatically near the viewport. Drag to inspect; pause and scrub to compare a pose.

Weapon Motion PausePhase

Loads automatically near the viewport. Drag to inspect; pause and scrub to compare a pose.

Weapon Motion PausePhase

Loads automatically near the viewport. Drag to inspect; pause and scrub to compare a pose.

Weapon Motion PausePhase

Loads automatically near the viewport. Drag to inspect; pause and scrub to compare a pose.

Equipped trials · partial: Choose Weapon separately from Motion to compare a sword and an axe, two pistols and two rifles across the available poses. Current rifle 5e0efeb97a6d is shouldered — the stock seats in the shoulder pocket inboard of the firing arm at cheek height, barrel horizontal-forward, both hands on the weapon; this re-bake also abducts the firing elbow off the ribs and repairs the shoulder/torso skin weights that previously tore the deltoid and chest seam apart at the aim pose. Pistol d17d7c1bdc9f keeps the unused left hand at the model’s side and drops the firing arm to a slightly-bent, shoulder-height forward aim. Sword 67d71938cc21 still has measured body/weapon intersections (a fix is in progress). New alternatives also fail clearance checks. Fingers remain fused and open; these poses do not yet reproduce the concept’s closed finger grips. Finger grip and final concept matching remain unfinished. Equipment, other models and terrain must not penetrate model surfaces before acceptance; broader runtime collision is still unverified. These independent review models are not game-runtime selections.

Aegis7 · fitted wrist and authored sword

Motion PausePhase

Current partial · b381f3d752c8: source-fitted wrist and elbow joints replace the misplaced forearm pivot. The existing Quaternius sword handle moves 18.5 cm from the cuff origin into the visible palm. Heavy contact was inspected from top, front and side: fingers remain open, the hilt is not enclosed, and arm armor shears. Six melee clips are selectable; complete phase matching and gameplay integration remain unfinished. The strict Walk reimport check still fails (0.267 mm / 0.111°). No grip acceptance or golden-body rollout.

Loads automatically near the viewport. Pause and scrub to inspect the open hand and weapon contact.

Articulated mechanical hand · pistol grip development

Pose PausePhase

15 articulated digit joints · opening and closing: GripPistol solid clearance passes. Lower fingertips now bend around the handle's far side, then straighten to about 9° of remaining bend during release, moving about 27 cm. Four clips retain rigid plates and pass sampled clearance against all eight pistol solids after GLB reimport. Thumb opening now uses the correct armature-space target; the index stays outside the trigger guard. The palm and long finger plates still look boxy rather than matching the concept, and full thumb opposition remains unresolved. This separate prototype has not replaced any character hand or received rollout approval.

Loads automatically near the viewport. Drag to inspect both sides of the grip.

1 · Pin identity and intent. Record the source model hash, textured reference, generation prompt, camera and target hash. Generate anticipation, peak/contact and recovery for each motion; reject changed anatomy or impossible poses.

2 · Pose, export, compare. Fit the selected Rigify skeleton, author controls and render the actual exported GLB at matching phases and camera. Compare silhouette, elbow bend, palms, weapon grip, foot support and armor seams against these targets.

3 · Repair and repeat. Name the visible gap, change controls or skin binding, then re-export and re-render. Rigid plates rotate and overlap at joints; they must not stretch, inflate or tear. Cap a candidate at three iterations and record non-converging defects.

4 · Verify motion, not only stills. Check transitions, timing, foot plants, attack windows and recovery in playback and the game runtime. A matching pose is not animation acceptance. Current rigs remain partial trials; future locomotion, chain and weapon variants need their own evidence.

Baseline, not a new animation candidate: the model below is the current golden-body reference with ten existing clips. The concept montage defines the intended improvements. Compare human balance, torso rotation, elbows, wrists and planted feet; retain character-specific proportions and customizations.

Motion PausePhase

Loads automatically near the viewport. Drag to inspect; pause and scrub to compare a pose.

Same pistol · baseline motion versus fitted trial

Both panels below load the same independent review GLB, 55c05deec72c, with the same procedural pistol. Left plays preserved source aim/fire curves; right defaults to the new two-handed support trial and also retains the one-handed trial clips. The original equipment-free reference above remains unchanged. Camera and playback controls are independent.

Trial gaps: the support palm now reaches beside the handle, but fingers intersect the grip/trigger frame. Thumb opposition and trigger clearance are not established; the muzzle retains upward pitch. This is a rejected grip experiment, not rollout approval. The original 13 clips match an unedited export/import control at all sampled frames. Open exact equipped review GLB.

Conventional rifle · baseline motion versus fitted trial

Rifle trial gaps: both panels use the same conventional stock/fore-end fixture and independent review GLB 1e4c558c6497. The barrel points forward, but the left hand is open and the stock is not seated to the shoulder. This replaces the loop-shaped fixture for this golden-body experiment; it does not establish a valid rifle hold. Open exact rifle review GLB.

Golden shortsword · weapon bone-mounted on the stock animation

Review GLB golden_equipped_blade_v9 with the project-authored shortsword at 0.60 review scale. The custom per-finger grip solve was abandoned: the weapon is bone-mounted to the right hand (grip seated in the palm to sub-millimetre) and rides the body's stock slash/melee curves as authored — the offhand keeps its source swing, no arm re-pose. The golden-rig forearm-twist fix is applied, so the elbow no longer candy-wrappers. Left plays the base slash/melee; right plays the equipped SlashGrip/MeleeGrip.

Blade notes: the finger-closure grip solve was abandoned — the weapon is bone-mounted to the right hand and aligned in the palm (sub-millimetre), riding the stock slash/melee arm curves; the offhand keeps its source swing and the forearm-twist fix removes the elbow candy-wrapper. The swing itself is the golden body's stock slash/melee motion (not a bespoke sword-attack clip). Open exact equipped blade review GLB.

Known coverage gaps: the preserved baseline has generic melee/slash and aim/fire, not distinct stab, heavy attacks or chain steps. Its dodge is a roll rather than the target dash; slide is absent. Equipped trials now include one-handed shortsword slash/melee plus pistol and rifle aim/fire/recovery. Two-handed melee, distinct heavy attacks, chains, shoulder/chest deformation and full-motion contact still need work.

No new animations are ready for approval yet. No batch updates to golden-body character files until you visually verify and explicitly approve the new animations in an equipped baseline-versus-improved comparison. Approval will identify the reviewed model/version and clips. Keep independent character GLBs and per-character overrides; do not migrate to one shared runtime body.

The hash-verified latest static cast is reviewed below. Every subject is present in the page and loads automatically near the viewport.

All 24 models are reachable by scrolling, load automatically near the viewport, and start facing front. The four stills below show Mech745 specifically.

Provenance: 7549011a886cc84bb1fe6cd57bd3a70b456fd7e63499af1639ffa19226bf3b60; 59,999 triangles; three boundary edges on one isolated triangle; zero connected-surface boundaries. This static proof does not accept custom-rig or Rigify animation, game integration, or motion or game-ready status for the other static subjects.