[AI] Castle roster bootstrap

Swordsman revision (October 2, 0.27.37)

The previous Swordsman still had a visible gap between the left hand and cuff. While walking, the sword arm pulled behind the body and the wrist folded unnaturally. The user’s follow-up exposed a gap in the earlier checks: bone connections and selected poses had been inspected, but the actual hand-to-cuff surface join had been missed.

Local version 0.27.37 re-exports all 13 groups and 76 frames. Hand attachments now follow the deformed cuff surface with a small overlap. The walking sword arm has been adjusted (historical 0.27.37 state), using the idle forearm orientation as its reference to avoid carrying forward the old motion’s excessive twist. Death, hit and defence retain their earlier body motion while receiving the same hand-attachment correction.

Two intermediate attempts were rejected. One lost the existing hand scale when replacing its transform, producing oversized hands. Another retained the walking forearm rotation and twisted the bracer into a pointed shape. Restoring scale and using the idle orientation resolved those defects before export. The existing Meshy body and rig remain in use; Astra edited the Blender motion and attachment transforms.

Packaging checks passed, all 338 installed files were verified against the delivery manifest, and a local battle confirmed loading. Portraits were updated too. Appearance still needs evaluation across full motions; format checks do not establish visual acceptance. Angel and Archangel sword effects remain in place, documented in the cavalry and angels article. Portrait background corrections are in the Castle town and interface article.

Current version and Swordsman poses (October 2)

The local Castle mod is now 0.27.36. This record has been superseded by 0.27.37. The Swordsman pose update landed in 0.27.35. Its free arm had been too straight, with much of the walking swing rotating the entire arm at the shoulder. The revision bends the elbow, adjusts the wrist with the forearm, and changes the sword-arm elbow direction.

The update replaces 53 frames across idle, movement, inspection, turns, movement transitions and all three attack directions. The creature retains its native 13 groups and 76 frames. Death, hit and defence remain from the previous version; the Crusader is unchanged. Astra edited Blender motion and hand attachments on the existing Meshy body and rig, without another model-generation job.

The 1× and 2× assets are installed, packaging passed with zero errors and warnings, and a local test battle confirmed loading. These checks do not certify the appearance of every motion. Angel and Archangel sword effects are covered in the cavalry and angels article.

The Castle roster now has independently reviewed meshes for all fourteen units: Pikeman, Halberdier, Archer, Marksman, Griffin, Royal Griffin, Swordsman, Crusader, Monk, Zealot, Cavalier, Champion, Angel and Archangel. They span long held equipment, a two-handed light crossbow, a winged quadruped, shield-and-sword combat, and an unarmed spellcaster. Each asset needs its own mesh and its own animation constraints.

All fourteen Castle units have Meshy bootstrap attempts, with modeling, motion repair and integration still in progress. The Zealot has returned to modeling because its old white-robed design did not match the original. The corrected blue-robed Zealot is now installed as a private test; the brown-robed Monk remains in offline animation work. The Monk is also installed as a private 0.6.0 candidate. The table distinguishes installed drafts from offline trials and rejected candidates.

The concept fixes the features that need to survive generation: blue-and-gold tabard, brown boots, uncovered face under a steel cap, and a halberd held vertically at the left side. It is a reference for construction rather than an in-game image.

Crusader local test package

The Crusader head proportions remain under revision. The original crest rises only slightly above the helmet; the previous model had a higher crest arch and a shorter helmet shell. This candidate lengthens the helmet and lowers the arch while preserving the trailing feathers. Steel shading and shoulder edges were also adjusted. The image below is an actual 1200 × 1400 Blender render; side and rear views were reviewed as well. Likeness remains unaccepted, the edits cover the holding branch only, and the game retains 0.13.0.

With the same fixed projection and panel crop, the original and candidate share the top and foot baseline. That verifies placement, not likeness. Earlier draft files named panel-source actually used a close-up camera. This comparison restores the fixed projection used by the game export.

The revised Crusader proportions and shield-arm motions are installed in private Castle mod 0.13.0. Thirteen groups and 76 frames use a fixed projection, with thicker forearm and leg armor and a larger shield. The front-attack and hit-reaction shield arms were solved again. The first hit frame put the blade into the right shoulder armor; a 3-degree wrist adjustment cleared it. Reopening and sampling 161 times found no blade–body, blade–shield or body–shield intersections. These three surface pairs do not establish complete motion or likeness acceptance.

Native likeness, skirt intersections, the falling pose and movement transitions remain unfinished. The separate narrower-foot holding branch was not mixed into this set, and neither normal-projection garment trial was adopted. The consistent existing stance and earlier displacement-transfer garment draft are retained. The fourteen-unit Castle roster remains in progress.

All 76 frames were exported again as native 2× images, with 1× resources and a 4× display cache resized from 2×; the cache adds no new rendered detail. Every frame fits the fixed battle crop. Format and native action-layout checks reported no errors or warnings. Installation replaced 507 Crusader resource files, preserved 3,317 other files byte-for-byte and backed up 0.12.0.

A local test battle ended, with client logs reading 41 distinct 4× body images across 7 groups. The log also contained 27 errors; this is neither an error-free run nor runtime coverage of every frame. No game screenshot was taken this round. Temporary display settings were restored from the original file. No VCMI source was changed and no new Meshy task was submitted. Models and the full mod remain private.

Griffin local battle test

The ordinary Griffin is installed in Castle mod 0.16.0 as a private test. Its fifteen-bone rig supplies thirteen clips and 85 frames, exported as 1×/2× body, shadow and required outline layers: 372 PNGs in total. Existing creature files were preserved, with backups and a validated rollback procedure. Appearance, the complete animation set and the full fourteen-creature Castle roster remain unaccepted.

The textured mesh came from an earlier Meshy task. Astra wrote the rigging, animation, rendering and packaging tools. This stage used no new Meshy request and made no VCMI source changes.

Attacks, transitions and shadows

All three attack directions now rotate the foreclaws downward during the lunge. Takeoff spreads and raises the wings before leaving the ground. Landing places the hind feet, lowers the body and returns to standing. Reopened checks at 257 times per transition found no floor penetration at those samples, with matching endpoints for the designated idle and flight poses. Landing from arbitrary flight phases still needs runtime coverage.

Several forearm repair trials were rejected. Preserve Volume skinning increased local stretching; reducing torso weights and smoothing weights along mesh edges also made neighboring edges worse. Converting the original short edge’s roughly threefold stretch to the game camera gave an upper bound of about 0.57 extra pixels at 2×. The current weights were retained while review continues on visible shape and motion.

Each shadow projects the actual posed 3D mesh in a fixed direction onto the same ground plane, followed by a uniform opacity reduction and edge blur. All 85 body-and-shadow pairs were exported and inspected. This directional projection does not use Cycles ray tracing; shadow appearance during continuous playback remains under review.

Runtime coverage

The first automated battle logged nine groups and 62 distinct 2× body frames. The second battle used for screenshots logged eight groups and 54 frames. The image below shows an attack pose and a corpse. Neither test covered all thirteen groups; turning, defence, upward attacks and the creature information panel still need separate checks.

Two test setups failed along the way. The first command omitted the map resource’s Maps/ prefix and could not find the map. Another trial used equal armies; the AI did not initiate the intended fight and the game ended through the townless defeat condition. The screenshots came from an isolated map that reliably triggered combat, with timed automatic captures.

Royal Griffin folded wings and stance

The Royal Griffin still uses the existing Meshy body and main wings, with Blender repair, rigging and animation tools written by Astra. The working set contains thirteen offline clips and 85 native frames: holding, hover, hit, defence, three attack directions, takeoff, flight, landing, two turn halves and death. It has not passed visual or in-game acceptance and is not installed. September 21: the 1×/2× draft is now installed in the private Castle mod, version 0.15.0. The client loads the mod and reaches the main menu; Royal Griffin battle playback and visual acceptance remain unverified.

The package contains 85 body frames at each scale, with shadow layers and selection outlines, for 372 PNGs. Native group counts and file validation pass without warnings, and the other 4,153 installed files remain unchanged. The shadows use a continuous-alpha projection onto a fixed image ground plane; they are not traced from the 3D scene. The sheet below is an offline composite of rendered frames, not a battle capture. Temporal shadow behavior and flight-to-landing transitions still need review.

The standing pose now tucks the tail behind the body and raises the two foreclaws at different angles. The earlier tail and forward-reaching claws made the silhouette too wide. A separate sizing error came from orphan vertices with no faces: a nominal 95-pixel height produced only about 71 visible pixels. Calibration now uses the actual surface. At matching height, the new pose is roughly 80 pixels wide against 81 for the native image, without horizontal image compression. Head placement, body proportions and feather contours still differ; matching width is not visual acceptance.

All thirteen drafts now use the revised standing pose. The correction fades away at attack peaks, during sustained flight and in the later death frames to retain those authored motions. Turns release the tail tuck gradually, preventing it from swinging out beside the body. Surface vertices remain finite across 1,181 sampled times, and 23 specified seams match within 0.001 mm. Preserved key poses also match their sources. This does not establish collision safety or seamless landing from any flight phase; the hit reaction still dips roughly 0.24 mm below the floor.

The turn now uses less shoulder sweep and a revised elbow fold. Side-colored diagnostics identified the feathers crossing behind the head as parts of the opposite wings. No non-coplanar intersections between those wings are detected at the new frontal key, but the folded stance and some intermediate poses still intersect. The spread silhouette also needs further comparison with the native pose. Both three-frame halves retain their standing endpoints, and all six frames fit the native-canvas preview.

Remaining work includes head and leg deformation, feather intersections, placement within the canvas, timing and game playback. Death has a nine-frame draft grounded from visible surfaces. Sustained flight exports four native frames, with an additional non-exported key for its loop. The record below retains earlier stills, failed trials and stage-specific conclusions; references to a “current” draft inside it refer only to that earlier stage.

Pikeman model and motion

Eleven Pikeman drafts cover thirteen native groups and 80 frame slots, including duplicate turns. Wrist, cuff and grip revisions are in the offline working set. Later work corrects the gait heading and adds supplemental start/stop clips for 3D inspection. The sections below distinguish each revision’s checks and rejected trials. This Pikeman is not installed. September 21 installation update: the upright-walk export is now in the enabled local Castle test mod, version 0.14.0 0.14.1 after the registration correction below. Native likeness, materials and runtime motion remain under review, and all fourteen Castle creatures remain in scope.

The later finger and thumb revision is in the working set; rejected trials are retained in its history section. The wrist images below retain the finger poses from that earlier stage.

Cuffs and wrist poses

Astra reused the existing Meshy body and donor hands, shortened the excess wrist stumps, revised their weights and rebuilt the connection from each sleeve opening. Earlier skin extensions and longer sleeves left the folded stump underneath unresolved.

The downward-thrust right wrist and falling left wrist also folded back too far. Comparing forearm and palm axes put the old poses at roughly 139° and 118°. The thrust now uses a different roll around the shaft with an original-length arm solve. After release during death, the palm gradually levels out. These axis angles compare model poses; they are not clinical joint measurements.

Reopening all eleven scenes and sampling 671 times found no intersections in the tested hand, arm, cuff, shaft and body surface combinations, including mutual hands. Denser checks at 401 times for thrust and 161 for death also passed those comparisons. Sampling and mesh partitions limit this evidence: complete self-collision, seam topology and ground contact remain unproven. The fingers and cuff materials still need work.

Finger grip and release motion

The later thumb revision is now part of all eleven offline action drafts. Astra smoothed weights between the palm, thumb root and distal segment, then adjusted opposition using the existing Meshy geometry. The old thumb segments formed an axis angle of about 128°. The closed poses now measure roughly 85° on the left and 65° on the right, reducing the folded-back tip. These are model-pose comparisons, not clinical joint measurements.

Animation exposed contacts that the idle pose did not. Sliding grips during front thrust, defence and downward thrust brought the shaft into the thumb base; the downward thrust also caught the right thumb tip. Closed-pose corrections cleared those cases, but denser sampling found brief contacts while releasing the shaft during hit and death. Small release detours now take the thumb around the shaft. The downward thrust adds a correction following its existing thrust envelope, returning to the same grip at both ends.

At the thumb-revision stage, reopened scenes cleared the tested hand, cuff, body and shaft surface combinations, plus ten digit regions against the rest of their respective hands, across 1,283 sampled poses. Of these, 805 cover denser hit, death and downward-thrust checks. Body and pike motion were retained. At the checked idle joins, evaluated mesh positions differ by less than 0.001 mm. Sampling and mesh partitions do not establish continuous collision freedom, complete self-collision or contact area, and velocity continuity remains unaccepted. This Pikeman is not installed. The later local test installation is recorded below. Native likeness, materials, walk heading and landings still need work. September 21 correction: the offline gait and transition revisions are described below; native likeness, materials and in-game motion remain unaccepted.

Native-frame export and walking revision

The Pikeman now has a complete 13-group, 80-frame body export on a shared 900 × 800 transparent canvas, twice the original dimensions. Its four-frame hover motion returns the pike before the engine switches to idle. Fixed-ground shadow and selection layers have also been generated. These are private test assets; the new Pikeman has not been installed. September 21: the package has now been copied into the enabled local Castle test mod. Its 11 used groups contain 76 frames per scale; the four unused turn-duplicate frames remain in the source export. All 326 body, shadow and selection PNGs passed resource validation at 1× and 2×. The installation preserved 3,825 existing mod files; only the mod metadata and new Pikeman resources changed. This establishes a local test installation, not live battle or appearance acceptance.

The local VCMI playback code advances a walking stack by one horizontal hex, 44 logical pixels, over five animation frames. A six-frame cycle therefore travels 52.8 pixels. Our earlier continuous Blender demonstration supplied its own root travel, so its planted feet did not establish that the exported frames would plant correctly in the game.

Astra tried matching that screen travel exactly. The rig needed a lower center of mass to reach the longer steps, producing a crouched advance. Comparing all six frames with the original showed that the posture had moved too far from its upright walk. That version was rejected.

The current candidate keeps the original body height, shortens the stride and reduces foot lift. Hand, cuff and pike motion remains unchanged. Forty-nine samples from the reopened scene retained loop closure. Over the same horizontal stance interval, measured foot drift fell from about 14.8 to 9.2 logical pixels. Residual drift remains, and these continuous bone measurements do not cover frame holding, diagonal routes or live gameplay.

The upper row uses the earlier frames; the lower row uses the upright revision before the registration correction below. This diagnostic composites ten sprite frames per second over continuous screen translation. It is not a game capture. Meshy supplied the underlying body; Astra authored the Blender motion, export and checks. Turning transitions, overall likeness and in-game review remain unfinished.

The local test package is now 0.14.1. One seven-logical-pixel rightward correction applies to every frame, with shadows and outlines rebuilt from the registered bodies. The earlier camera fit used the combined body-and-weapon bounds, leaving the body too far left. Matching both the idle foot region and the native turn pose brings the foot-region center within one logical pixel of the reference.

The test includes the engine’s horizontal reflection and its 58-pixel change in sprite origin when facing switches. Alpha-centroid displacement at that join fell from 27.5 to 13.5 logical pixels; the original sequence measures 11.9. These are image-registration measurements, not proof of matched anatomy or smooth gameplay. The client startup probe stalled after OpenGL renderer creation, so battle appearance is still unverified.

Thrust and death previews from the wrist revision

These eleven-frame thrust and five-frame death previews use actual renders of the current Blender scenes the scenes saved at the wrist-revision stage at 800 × 934 and 12 fps. Both retain one fixed blog-review camera without per-frame resizing. They are not gameplay footage, and the playback rate has not been calibrated to the game.

Archer local test package

The Archer’s 16 active groups and 96 frames are now packaged at 1×/2×, with a precomputed 4× cache resized from the 2× images rather than newly rendered at 4×. Installation into Castle mod 0.11.0 preserved the other creatures’ assets. Native action-layout and image-format checks returned no errors or warnings; motion quality remains under review.

The test battle reached its end. Client logs confirm loading 43 body images across six groups: holding, hit, death and all three shooting directions. Text-encoding, audio-device and one query error remain in the log, so this was not an error-free run. There is no game screenshot establishing full visual acceptance. The shooting preview uses the exported sequence; fingers and transitions still need review.

All fourteen meshes

All fourteen models were generated from separate reviewed concepts and then checked from eight Blender angles. Accepted Meshy mesh tasks consumed 420 credits; the earlier rejected four-view Halberdier request consumed another 30. The Halberdier and Pikeman are now past the first local rig stage: both have holding and walking review sequences at their original 450×400 canvas. The Archer holds its crossbow in the static and holding review, but its first raised-shooting test exposed separated sleeve topology and was rejected. The remaining units have accepted mesh reviews and move on to creature-specific local rigs. That was the initial bootstrap review; the table below records later problems and current action status.

Review gallery

These are Blender review renders from the accepted local states. They show the actual material under review, rather than concepts or a claimed game result. The gallery retains drafts, rejected candidates and historical review results, with their scope stated in the surrounding text and captions. Unless explicitly identified otherwise, these are offline Blender renders and do not establish native-game acceptance.

Halberdier skinning revision

The earlier local rig split the Meshy surface into fragments and attached each rigidly to a nearby bone. That supported weapon-path studies, but continuous joint deformation needed a skin. The existing model has now gone through Meshy’s humanoid rigging API.

The first request returned HTTP 400 because the source had 741,522 faces, above the rigging limit. A remesh request targeted 50,000 quads and returned a GLB with 98,106 triangles. Rigging then succeeded; each task cost 5 credits. The source design was reused. The imported Icosphere bone-display helper was explicitly excluded before normalizing the mesh and armature together.

Knee, elbow and shoulder probes exercise individual joints. None triggers the current large-edge stretch threshold, which does not establish weapon rigidity, cloth quality or grip contact. Weapon weights and the grip need separate checks before authoring native-length clips.

The remesh tool and reproduction instructions are public. Three tests cover recovery without a new submission, duplicate-task prevention and rejection of a different source task. Models and complete mods remain local.

The full six-frame front attack exposed problems that the small joint probes had missed. The halberd head carried upper-arm weights while its butt carried pelvis weights. The complete weapon and primary grip now share a rigid hand control. The mask below shows the weapon in yellow and primary grip in red.

The first two-hand attack pulled the coat upward with the arm. Attenuating distant arm weights reduced the large displaced cloth panels, but the support hand, cuff and clothing boundary still formed stretched strips. Extending the mask using skin colour also selected some gold trim; tightening it left parts of the hand behind. None of these candidates passed review.

Reopening candidate 99’s six-frame file gives a maximum selected-weapon edge-length change below 4.6×10⁻⁷ model units. Each body frame still has 258–274 edges longer than 0.08 model units and more than three times their rest length. Weapon rigidity passes this check; body skinning fails. The next step is to resolve the support-hand/clothing geometry boundary before extending the binding and motion. Work subsequently moved to the arm-separated body below. The Halberdier remains uninstalled.

Arm-separated body candidate (September 18)

Local selection changes did not resolve the support hand pulling on the coat. A built-in imagegen edit now places the same character in an A-pose with open hands clear of the torso, removing the halberd. The blue-and-gold tabard, heraldic emblem, helmet and boots follow the previous design. The weapon will be attached separately. The subsequent six-frame combination is documented below.

Meshy 7 received this image through the API with a 40,000-quad target and 4K textures. The downloaded body contains 75,660 triangles. Generation cost 30 credits and the subsequent humanoid rig cost 5. The 1200×1400 Blender still below shows the new geometry; front and side views were inspected for hand separation and body depth.

Astra authored five larger single-joint probes and a six-frame two-arm range test. Reopening the saved six-frame scene found no edges longer than 0.08 units and more than three times their rest length. Four front/side key-pose renders show no strips pulling the coat toward the support hand. This covers the tested range only.

The raised pose still compresses the shoulder armour. The separate-plate experiment below follows up on this failure. Rigid plate controls, wrist orientation, finger grips and weapon attachment remain ahead of the native animation set. This is a working body candidate; the Halberdier remains uninstalled.

Two shoulder treatments followed. Assigning the entire mask to the upper-arm bone stretched the adjacent cloth, producing 6–112 large edges per frame; that candidate was rejected. Candidate 103 extracts two independent plates and follows part of the upper-arm rotation. Across six reopened poses, plate edge lengths vary by less than 2.92×10⁻⁷ units and the retained body stays below the large-edge threshold. Raised-plate placement and the neckline seam still need visual correction.

Grip work starts from the open hands. Curl attempts 104–106 flattened fingers or thumbs and were rejected. Candidate 108 adjusts the curl axis and separates the thumb treatment from the other fingers. A diagnostic shaft of radius 0.014 units exposes the remaining gaps: the thumb is open and finger contact is unfinished. This rod is not the final halberd.

Six-frame body and halberd combination

The original Meshy halberd is now extracted. The first extraction retained part of the old gripping hand, so the affected central band was removed and a short wooden handle segment inserted. The head, remaining shaft and butt cap reuse the existing geometry. Candidate 116 combines this weapon with the new body and independent shoulder plates in the original six-frame front-attack group.

Early combinations put the right hand forward and kept the shaft aligned with the body’s forward axis. The horizontal pose intersected the torso and arms, and the axe face lay flat. Comparing the original impact frame led to a right hand close to the body, a forward left hand, and a shaft crossing diagonally in front. Adjusting the rear elbow then cleared the buttward shaft. Candidates 111–115 remain rejected history.

After reopening the saved scene, none of the six body poses triggers the large-edge stretch threshold. Nine longitudinal rays along the shaft centre and circumference find no intersections with non-hand body triangles or either shoulder plate at those six frames. Grip-hand faces are explicitly excluded. Finite ray sampling does not cover every weapon surface, finger contact, self-intersections or intermediate poses, so this is not a complete collision guarantee.

This front-attack candidate includes a small forward step. The editable base and combined action are saved separately. Finger detail, native stance/camera/scale and the remaining animation groups still need work; the Halberdier remains uninstalled. No new Meshy requests were made in this step.

Holding, walking and native registration

At candidate 121, eight holding frames and six walking frames joined the earlier six-frame front attack: three groups and twenty native frames. Both new groups keep the original two-handed, slanted pole carriage with alternating legs. The first walk sank a sole by about 0.00391 units. Rigid sole weights blended into the ankle reduce maximum penetration across the twenty native frames to about 8×10⁻⁸ units. Some lowest points sit slightly above the floor, by at most about 0.000095 units.

Side-by-side review exposed a reversed blade side and an opposite starting leg phase. The blade was rolled over, the walk shifted by half a cycle, and camera azimuth changed to 52° while retaining 30° elevation. This remains an offline review camera.

The original canvas is 450×400. The first holding frame sets height, horizontal centre and ground registration; the same camera then renders all three groups without per-frame scaling. All twenty frames fit. Evaluated vertices match exactly between holding keys 1 and 9, and between walking keys 1 and 7. Those final closure keys are not extra exported frames.

No frame triggers the large-edge threshold, and the nine longitudinal shaft samples avoid non-hand body and shoulder-plate triangles. The sampling limitations above still apply. The comparison also shows insufficient attack reach and stance, with detailed finger contact still unreviewed. These clips have not passed complete visual acceptance. Up/down attacks, hit, defence, death, turns, hover, shadow layers and game integration remain unfinished. Up/down candidates have since been added below; the other items remain unfinished.

Attack and defence candidates

Candidate 131 retains eight holding frames, six walking frames and nineteen directional attack frames, and adds the original twelve-frame defence group: six groups and forty-five frames. Front preparation raises the grips, the upward wind-up now passes across the head, and strikes use a deeper knee bend. These are fixed-camera offline candidates.

Defence opens from the slanted holding pose, turns into a cross-body block and returns to holding. Draft 130 raised both hands above the forehead with a nearly upright stance. Candidate 131 lowers the grips and adds body rotation and knee bend. Its shaft is still flatter than the original block, whose blade points slightly upward.

Reopening the scene gives a maximum grip-target error of about 5.7×10⁻⁷ units across nineteen attack and twelve defence frames, with no edges above the existing stretch threshold. Nine shaft rays avoid non-hand body and shoulder plates in all forty-five frames; all frames fit the native canvas. Evaluated body, plate and weapon vertices at both defence endpoints match holding frame one exactly. These checks do not certify complete surfaces, finger contact or between-frame motion. The close-up still shows incomplete finger closure and distorted shoulder trim. Hit, death, turns, hover, shadows and game integration remain unfinished. A hit candidate is included below; the other items remain unfinished.

Hit reaction and hand release

Candidate 137 adds the original six-frame hit reaction, bringing the scene to seven groups and fifty-one frames. The body recoils, the stance widens, and the later frames lean forward again. The secondary hand leaves the shaft and opens. Astra turned the original Meshy open-hand geometry into an animated shape key; other groups keep it disabled.

Draft 132 drove the shaft tail through the right thigh. Adjusting the weapon cleared that issue, but candidate 135 still had a narrow stance and weak recovery. Candidate 137 revises foot placement and weight shift. After reopening, all fifty-one poses stay below the existing large-edge threshold and avoid the nine shaft rays against non-hand body and plates. Evaluated body, plate and weapon vertices in the earlier forty-five frames match candidate 131 exactly. Six new frames were rendered; unchanged images were reused after checking vertices and projected bounds.

The last hit frame still has a released hand, so its return to holding needs review. The gripping fingers, shoulder trim and pose detail remain unfinished. Death, turns and hover account for twelve missing native frames, followed by shadow/outline layers and in-game verification. Turn and hover candidates are added below; the death trial is rejected. The Halberdier is not installed.

Turns, hover and a rejected death draft

Candidate 140 adds two frames for each turn direction and four hover frames, extending the working candidate to ten groups and fifty-nine frames. VCMI plays the first turn segment, flips facing, then plays the second. Both segments follow the original frame counts and bring the pole upright; hover adds a small pole lift.

Adding the groups exposed a control-type regression: an integer assignment reduced the earlier hand-release values to fully open or closed. Float assignments restore the intermediate values. Reopening and comparing evaluated vertices confirms that the prior fifty-one body, plate and weapon poses match candidate 137 exactly. The eight new poses stay below the large-edge threshold. All fifty-nine fit the canvas and avoid the nine shaft samples against non-hand body and plates.

Four-frame death trials 141–144 were made separately. In 141, torso and arm rotations cancelled too much of the fall, leaving a seated-looking endpoint. Later drafts release the pole onto the ground and bring the arms back in. The high-resolution result exposes an unnatural fan of coat folds around the bent knees, and the corpse differs substantially from the original silhouette. The death group is rejected and excluded from the fifty-nine-frame candidate.

No edge exceeded the stretch threshold. The final body’s lowest point is about 0.008 units above the floor and the weapon’s about 0.006. Those measurements cannot establish plausible cloth folds. The next work is the coat and leg pose during collapse, followed by gripping fingers, shoulder trim, transitions and fuller collision review. Shadow/outline layers, installation and native visual acceptance remain unfinished. Shadow and outline layers are now included in offline preview 164 below; installation and native visual acceptance remain unfinished.

Trials 145–150 examined the lower coat weights. Thigh bones account for most of the selected blue-and-gold fabric weights; shin bones contribute about 0.7%. Added coat-fold controls in 146 tore the selection boundaries into strips. Expanding and smoothing that selection in 147 reduced the outer-coat damage, but the inner mail still stretched. Both were rejected.

A closer look at the original endpoint led to a side collapse with asymmetric knees and the raised arm brought back toward the body. Trial 150 retains the existing cloth weights. Ground placement now includes both shoulder plates, correcting the previous trial’s roughly 0.026-unit right-plate penetration. Evaluated body, plate and weapon vertices in all previous fifty-nine frames remain identical.

The coat is more compact than in 144. However, the original still holds its pole aloft in frame three, while this draft has released it. Corpse and weapon silhouettes also differ. Death remains an unaccepted draft outside the fifty-nine-frame candidate. All sixty-three trial poses fit the canvas and pass the existing edge threshold. Those checks do not establish cloth self-collision, complete surface clearance or plausible weight support. The next revision needs to retain the pole during the intermediate fall and improve body pose before reviewing landing contact and transitions.

Trial 157 keeps the pole in the left hand through frame three and releases it in frame four. Arm elevation and wrist orientation are controlled separately. Rotating the whole arm in 152 drove the butt into the floor; the wrist roll in 155 put the shaft through the forearm. Reversing that roll and moving the frame-two holding arm forward restores a visible diagonal blade. Nine shaft rays find no non-hand body or plate intersections across all sixty-three poses. Body, plates and weapon remain above the floor, and the previous fifty-nine poses match 140 at every evaluated vertex.

Early release and the edge-on blade have been addressed. The second pose is still less curled than the original, and the final corpse and dropped weapon silhouettes differ. Death remains outside the fifty-nine-frame working candidate. Body compression, weapon landing, contact and transitions remain to be revised; sampled clearance does not establish complete collision or appearance acceptance.

Trial 160 folds the hips and knees further in frame two and moves the holding arm forward to clear the head. The final weapon now rests with its blade flat; its previous upright blade held the shaft too high. This remains a death draft. The body is more compact, but the final pole is more prominent than in the original, and grip, cloth and transitions remain unaccepted.

A separate triangle-surface check covers the complete weapon, including blade and metal butt. Across sixty-three poses it finds no intersections with either shoulder plate or the body after excluding wholly hand-dominant triangles. The same check detects 349 and 199 intersecting triangle pairs in frames two and three of known-failing trial 155. This does not cover complete containment without a surface crossing, finger contact, body self-intersection or motion between sampled frames. All previous fifty-nine poses still match 140 at every evaluated vertex.

Layered review and geometry shadows

Offline package 164 contains eleven groups and sixty-three frames of 1×/2× bodies, shadows and the required selection outlines. Four death poses still await appearance acceptance. This package is not installed. Installation record 507 now installs it as a test candidate; the observed native coverage follows below.

The first shadow pass transformed body alpha around a fixed ground line. It lacked the vertices’ individual heights and depth positions. The replacement reads the evaluated Blender mesh, projects it along a fixed light direction onto world z=0, and renders that silhouette through the same camera. Opacity and spatial softness are shared between the comparisons; body images remain unchanged.

The sixty-three geometry shadows produce 126 files at 1×/2×. Body and selection-outline files match the preceding package byte for byte. Format validation reports zero errors, zero warnings and twenty informational notices about movement inside the canvas. No temporal averaging or frame-count change was applied. The native encounter below now covers some groups. Flicker, transitions and full appearance acceptance remain under review.

Installation record 507 adds the preceding package 164 to the local mod. It retains that package’s grip and does not include the later rejected finger experiments. The source-scene hash and format checks were revalidated. The installation adds 278 files while preserving 1,123 existing Swordsman, Crusader and Marksman resource files.

The native encounter loaded 38 distinct 2× body images in eight groups: holding, hover, movement, both turns, forward attack, hit and death. Upward/downward attacks and defence were not exercised. The capture below shows the installed draft; grip anatomy, shoulder cloth, corpse support and transitions remain open, along with a dedicated review of shadow stability over time.

Marksman layers and battle test

Private package 479 is installed locally. The original file contains two unused duplicate turn groups, 9 and 10. VCMI uses 16 groups and 97 frames from this set; the full reference comparison retains 18 groups and 101 slots. The first export incorrectly included both unused IDs in its configuration. Removing them resolved the format errors; the body animation validator now reports zero errors and zero warnings.

Shadows come from each evaluated 3D pose projected onto the ground, including visible hands, weapons and the bowstring. A fixed opacity and soft edge are applied before exporting both resolutions. Holding and hover also receive selection outlines. Those outlines support highlighting; the clothing keeps its original model colours, without player-colour recolouring.

The native battle loaded 39 distinct 2× body images across holding, hover, forward shooting, downward shooting and death, plus two directions of the separate Meshy bolt. The mod sets release to frame 4 and uses Blender-derived launch positions. Directional foreshortening and the visible handoff still need frame-by-frame review. This encounter did not exercise every group. Its log also contains text-encoding errors, a test-map Grail-placement warning and a query -1 error at termination, so this is not a clean-log claim.

Existing Swordsman and Crusader files were preserved byte for byte. The late fall, straight corpse legs, finger anatomy, sword-to-empty-hand transition and loading continuity remain unresolved. Revision 502 installs the prone, bent-leg ending; the mid-fall motion, fingers, sword transition and loading continuity still need work. Movement, turns, melee and upward shooting need dedicated native coverage. Subsequent logs cover these groups, as recorded below; appearance and transitions still need review. Earlier installation statements below describe the named historical versions.

Diagnostic run 506 temporarily sets Marksman ammunition to zero so the AI approaches and uses melee. It restores both creature configuration and user settings byte for byte afterwards. That battle read 63 body frames in 12 groups, including all three melee directions, movement, start/stop, both turns, hit, holding, hover and death. Together with the previous shooting and death runs, native logs now cover 87 distinct body images in 15 groups. Run 503 also exercised upward shooting. The ten-frame defence group still lacks native coverage.

These observations establish that the game selected and read those assets. They do not validate the visible sword draw, facing change, recovery or loading transitions.

Marksman melee and turns

The offline body draft now covers all 18 original groups and 101 frame slots. Groups 9 and 10 duplicate the two turn clips, leaving 97 distinct body images. Each directional melee clip has six frames; each turn half has two. Coverage is complete at the body-draft level. Pose fidelity, transitions, shadow and player-colour layers, projectile handling and native acceptance remain unfinished. Body, shadow, selection-outline and projectile drafts are now installed; the preceding section lists observed native coverage and remaining work.

The original keeps the crossbow in the left hand while the right uses a short sword. This pass reuses the independent sword previously generated through the Meshy API, with no additional generation charge. Astra aligns the grip and authors the wind-up, strike and recovery in Blender. The final melee frame still holds the sword; drawing and sheathing across the empty-handed holding pose remain unresolved.

An 8 fps offline review, with four-frame pauses added at each clip boundary.

The first forward and downward paths exceeded arm reach by as much as 12 cm. Bringing the grip closer restored attachment. Reusing the crossbow finger pose also left the sword resting on an open-looking palm; the revised grip bends the four fingers farther and rotates the hand around the handle. The battle camera exposed another mismatch: the initial front strike projected too low and the downward strike too vertically. The latest paths adjust both sword direction and the bow arm’s backward travel.

Each saved scene was reopened and sampled at 321 times. Maximum right-hand drift against the sword is about 2.89 mm across the three clips, with no coarse body-edge distortion flags. These measurements do not establish finger anatomy, clearance around the guard or fidelity to the original poses.

This offline demonstration follows VCMI’s existing sequence: play the first half, flip facing, then play the second half. The engine source was read and left unchanged. Both halves share a frontal pose, but native positioning, the visible weapon-side change and footwork still require an in-game review.

Marksman reactions, hover and death drafts

The current body set has eleven draft groups and 75 frames. This records the stage before melee and turns; the current coverage is listed above. This pass revises hit and defence, adds the original nine-frame hover interaction, and introduces a six-frame death draft. Astra authored the animation in Blender around existing Meshy geometry, with no new generation charges. All four groups remain offline. They are now included in installed test package 479; death timing and the final pose still need substantial reference work.

This video retains death version 437. The revised 499 comparison, stills and native capture appear below.

The review adds four-frame pauses at each clip boundary. Its 8 fps playback is for inspection and does not establish engine timing.

The stronger recoil and higher defensive carry preserve continuous left-hand attachment. Maximum sampled drift is 0.50 mm for hit and 0.56 mm for defence. Defence starts and ends at holding; the hit also returns to the same evaluated body and weapon endpoint. Finger anatomy and cloth contact remain unaccepted.

The first hover draft put the crossbow below the left hand’s reachable range, leaving more than 5 cm of error. Raising it and moving it toward that shoulder reduced drift to 0.05 mm across 129 samples. Both endpoints match holding. The original still looks farther down toward the weapon.

Death revision 499 replaces the supine, extended-leg ending with a prone pose and asymmetrical knee bends. The fall also reaches the ground sooner in frames four and five. It reuses editable Meshy geometry and local animation without another API charge. Test package 502 installs the six revised body frames and their geometry shadows.

A ground-plane render exposed elevated hands and a floating crossbow that were less obvious against transparency. Lowering the torso, hands and weapon brought them close to the floor. Across 321 reopened-scene samples, the gripping hand drifts by at most 0.047 mm relative to the bow, with up to 0.074 mm of floor penetration and no coarse long-edge flags. This does not validate finger anatomy, cloth contact or the tumble’s timing.

The native client read all six revised 2× death bodies and shadows. Its battlefield now shows the prone corpse below. Only 24 body/shadow images across both resolutions were replaced; 1,099 other mod files retained identical bytes.

Earlier Halberdier fragment rig

The holding and walking images below document the earlier local rig. They do not establish finished motion on the new skin.

Halberdier: eight-frame holding cadence

Angel: independent replacement sword, static eight-view check

The replacement sword has passed its own static review. It is deliberately shown separately: the compatible unarmed Angel body did not pass wing review, so this is not yet a combined Angel action.

The accepted independent Angel sword is now combined with the original Angel body after selectively removing only the original central blade islands. Front, oblique and side static checks show a single intact sword at the retained hands, while armour and wings remain continuous. This is a static combination acceptance only; its flight and attack actions still need separate review.

The combined Angel now also passes its native-count seven-frame CANGEL.DEF flight review. All front frames and side frames 1, 3, 5 and 7 retain the one independent sword, the two retained hands, body and wing surfaces. This completes the sword-safe holding/flight path; attacks remain separate work.

The first six-frame independent-sword Angel front-attack probe reparented the sword around the static dual-hand guard pivot. An imported parent-inverse transform made it orbit an incorrect world point and leave the Angel, so it is rejected. Flight remains accepted; an attack needs an explicit world-space carrier rather than this inherited parent relationship.

The replacement six-frame ATTACK_FRONT uses an explicit world-space sword matrix around the visible dual-hand guard. It fixes the rejected inherited-parent pivot: all front frames and side frames 1, 3, 4 and 6 keep the one sword, torso and wings continuous. This accepts the weapon path; a later local rig pass will add larger arm travel.

The same explicit world-space constraint now passes the six-frame ATTACK_UP path: start, raised diagonal sweep, high contact and recovery retain the unique sword, hands, body and wings in front and side review.

The six-frame ATTACK_DOWN completes Angel’s three-direction independent-sword weapon paths. Its low diagonal release and recovery retain the one sword, retained hands, torso and wings in front and side review.

Angel’s native ten-frame DEFENCE now passes with a high diagonal sword guard. All front frames plus side frames 1, 4, 6 and 10 retain one sword and continuous hands, body and wings.

The first eight-frame Angel death uses a shared carrier for body, rig and independent sword, so it preserves continuity, but the rightward fall exits the front canvas and its end silhouette is not a readable ground fall. It is rejected; flight, attacks and defence remain accepted.

Angel’s independent-sword path now includes a six-frame hit recoil and its native two-frame start / three-frame end flight transitions. The recoil applies the same world-space displacement to body and sword; the transitions open and settle the wings while retaining the sword at the hands.

Archangel now uses the same separate-sword strategy: central embedded blade islands are removed, then an independently reviewed Meshy sword is placed at its retained hands. Static front, oblique and side review plus all seven flight frames and four side keys pass with one sword and continuous armour and wings.

The three native six-frame sword paths are now separately reviewed as well: a front sweep, an upward stroke, and a downward stroke. The sword is controlled by an explicit world-space matrix around the measured two-hand guard point, so it does not inherit the imported parent pivot or malformed hand-tail transforms. The contact and recovery keys below, including side checks for every direction, keep one sword visibly held through the action. These remain Blender motion reviews; DEF packing and game installation are not claimed.

A ten-frame defence, six-frame hit recoil, two-frame move start, and three-frame move end have now passed the same front-and-side review. Defence keeps the separate sword readable across the chest at its high point; hit applies the identical world-space offset to body and sword; the short transitions use Archangel’s own wing bones while retaining the single sword carrier. The review set intentionally does not claim a death pose, DEF packing, or in-game installation.

The first eight-frame Archangel death probe was rejected. Its front endpoint approached the original compact folded-wing silhouette, but its side review exposed a wing that had been flung away from the body by the current local wing-bone fold axis. The original CRANGL.DEF ends as a compact protected form, rather than the earlier rejected sideways-fall approach. A dedicated folded-wing rig is required before this group can be accepted.

Marksman body and crossbow separation (September 18)

Lowering the crossbow toward the original nine-frame hover pose exposed severe shoulder and sleeve distortion in trials 173 and 174. Earlier continuity reviews did not cover this pose. Both trials are rejected; earlier model and action files remain available.

Candidate 175 starts with an empty-handed A-pose body so the crossbow can receive separate controls. Built-in imagegen produced the reference; Meshy API model meshy-7 generated the textured body for 30 credits. The request used a 40,000-polygon target and 4K textures; Blender imports 43,343 vertices and 69,924 triangles. Front, side and rear renders show no obvious volume collapse. Posed shoulders, fingers and crossbow attachment remain unverified, and the model is not installed.

Meshy humanoid rigging subsequently completed for another 5 credits. Single-joint checks in Blender are in progress; crossbow handling, shooting and game frames have not been produced from this body. Single-joint checks and three two-arm poses—low carry, level aim and raised aim—have since been rendered. These empty-handed tests did not reproduce the earlier severe shoulder distortion. Full actions and game frames remain unfinished.

A separate crossbow was generated through the Meshy API for another 30 credits. Astra calibrated its size, top orientation and attachment coordinates in Blender. This body, humanoid rig and crossbow cost 65 credits together, excluding earlier models and trials.

Separate hand geometry

The existing donor hands still broke down when the thumb moved across the stock. Moving the palm reduced sampled skin intersections, but the isolated hand renders exposed sharp web and wrist geometry. Trial 241 reached both thumb-bone targets while visibly tearing the skin, and is rejected.

A new reference shows one open left hand with separated fingers and a short wrist. Built-in imagegen produced the reference; the Meshy API generated the textured model for 30 credits, using Meshy 7, a 20,000-quad target and 4K textures. Astra prepared the local Blender scene and inspected its turnarounds.

The downloaded mesh contains 37,376 triangles. Welding coincident UV-seam vertices at a one-micrometre tolerance leaves one connected mesh with 18,690 vertices, no open edges and no non-manifold edges. The side views retain palm thickness and the five digits are distinct. These are static checks. Finger rigging, thumb opposition, cuff fitting and posed skin contact still need to be done. Local rigging follows below; cuff fitting and usable grasp remain unfinished. The earlier body, crossbow and hand candidates remain available; nothing from this hand branch is installed.

The new hand now has three bones per digit, including a separate thumb-base joint, plus a wrist/palm root. The joint positions come from sections of the Meshy geometry. The first larger curl pinched the thumb base; moving that joint and widening its weight transition reduced the crease. Fifteen individual joint probes preserve the other connected fingertip regions within two micrometres, and rest geometry remains unchanged. This does not validate the complete grasp.

Combining the hand with the actual Meshy crossbow still produces skin intersections. The stock measures about 5.6 cm deep at the test grip. A 2.8 cm depth variant reduces some intersections; a separate trial preserves the original shape and changes placement. Neither passes, and the thinner variant has not been adopted as a fidelity correction. The hand remains separate from the character while finger contact is repaired. The subsequent body assembly is shown below. No additional API credits were used for these local rigging trials.

The next fitting pass evaluates the deformed skin while retaining the original independent crossbow dimensions. In the standalone hand test, vertices more than 0.2 mm inside the stock fell from 2,998 to 65, with maximum depth falling from roughly 8.6 to 1.6 mm. This checks every hand vertex, but does not cover full triangle intersections or finger self-collision; the grip is still unaccepted.

The left hand and a mirrored right hand have now been assembled with the Marksman body. The first assembly exposed excess wrist geometry and placed the left wrist target about 6 cm beyond the arm’s reach. Trimming the extra wrist section and raising the carry brought both wrists onto their targets. Each hand retains its own finger rig and follows the character’s wrist.

Candidate 262 record:

Moving the grips along the stock changes their cross-sections, so contact must be checked again. The assembled left hand has 29 vertices more than 0.2 mm inside the wood, reaching about 2.2 mm; the right has 176, reaching about 4.1 mm. Pointed cuff geometry is also visible in the close-up. These need repair before full motion. This is an editable body-and-hands candidate, with no new game installation or API charges.

The cuff section extending into the palm has been shortened by 3 cm. Its endpoint now follows the wrist, with weights blending back into the forearm. Body faces, UVs and 37,330 central vertices remain unchanged. Refitting each hand to its own stock section leaves eight left-hand vertices more than 0.2 mm inside the wood, reaching roughly 0.37 mm; none are detected on the right. These are vertex tests. Palm folds and thumb anatomy still need close inspection.

The assembled character now has a carry-to-aim study. With only three poses, hands followed the bow at the keys but drifted by about 2.1 cm between them. Solving 17 poses along the path reduces maximum relative drift to about 0.24 mm across 33 times checked after reopening the file. No body edge exceeded both three times its rest length and 8 cm in this coarse tear check. The nine-frame study covers raising the bow only; firing, recovery and the other native actions remain unfinished; the subsequent eight-frame body candidate is shown below, while the complete action package remains unfinished, with no game installation.

Native eight-frame shot

Reviewing CHCBOW.DEF frame by frame shows the right hand starting beside the waist, then joining the crossbow. Frames 4–6 hold a level aim, frame 7 raises the bow sharply, and frame 8 lowers it. The new body candidate follows that order, with the right fingers changing from a relaxed curl to their fitted grip. The first level aim sat at chest height; comparison led to a higher position closer to the original jaw-level aim.

The clip presents the eight body frames at eight frames per second for review, not as a claim about in-game timing. They come from a Blender 3D action, rendered on a fixed 900×800 canvas and camera at 2x. None reaches the canvas edge.

Reopening the file and checking 113 times gives a maximum hand-to-bow deviation of about 0.51 mm: the left grip is checked throughout, and the right from frame 4 onward. No coarse large-edge stretch check triggers. More solved poses and linear interpolation reduce sliding around the raised-bow reversal. All evaluated body, hand and bow vertices at the eight integer frames remain exactly unchanged by the interpolation update.

This is still a front-shot body candidate. Projectile, string and trigger behavior, hand close-ups, the return to holding and the other actions remain unfinished. Nothing new has been installed. This pass reused existing Meshy models without additional API charges.

That paragraph records body candidate 289. Subsequent string and bolt work follows below.

String motion and projectile handoff

The independent Meshy crossbow and bolt remain the source models. Astra added an articulated 3D string and limb deformation in Blender. During release, the string centre advances with the bolt tail while the limbs return toward their relaxed shape. These are 1400×1400 Blender stills.

The saved scene was reopened at 68 times, including 65 distinct draw values during release. String length stays near 0.556 metres; maximum bolt-tail deviation from its centre is below 0.001 mm. All 15,175 protected central-stock vertices remain unchanged. This verifies the authored geometric relationship. Limb material strain was not simulated, and finger, trigger and full collision checks are unfinished.

Advancing the bolt also changes its launch origin. In the current battle camera, the front-shot offset draft is (25, -56), still releasing on frame 4. The previous flying bolt was visibly oversized: recalibration reduces the horizontal alpha bounds from 54 to 23 pixels at 2x, close to the loaded bolt’s projected longitudinal span of 11.3 logical pixels. Up/down shots, native direction changes and readability still need review.

This remains an offline candidate, with no new Meshy charges or game installation. Trigger/finger interaction, recovery into holding and the other actions remain unfinished.

Index finger and trigger draft

A view beneath the crossbow exposed a missing part of the grip review: the right index finger wrapped over the stock and could not reach the trigger lever.

The first attempt moved the support hand far forward and inverted the rear grip. It failed: the support-hand target exceeded arm reach by as much as 12.4 cm, and the rear cuff turned upward. The replacement keeps the support hand in place, advances the rear grip by 2.5 cm and rotates it 45 degrees before articulating the index separately.

The new 3D action brings the index toward the underside of the lever, rotates the lever slightly about its front attachment and resets it after release. An earlier target touched the upper side while lifting the lever, an inconsistent contact direction; the target is now underneath.

The body, hand and crossbow are still Meshy assets, with local motion authored by Astra. The index solver does not yet constrain every joint degree of freedom. A small fingertip-to-target error does not establish anatomical motion or collision-free surfaces. Thumb folds, joint limits, metal contact and recovery into holding remain unfinished. Nothing was installed and no Meshy credits were spent. The preceding version 315 string views retain the earlier grip.

Holding and the final shooting pose

The native eight-frame sequence exposed another pose mismatch: its final frame returns the right hand toward the waist, whereas the previous candidate retained a two-handed grip. Candidate 353 now releases the rear grip after the raised-bow pose, including the trigger finger, without changing the frame count.

Playback is eight frames per second for offline review. The high-resolution final pose below shows the withdrawn right hand.

Holding candidate 354 has eight frames, a nearly vertical bow beside the leg, the free hand near the waist and a small carriage motion. The first free-hand pose was too open; the fingers were subsequently curled further. Holding currently uses a relaxed string and no visible loaded bolt. Loading, cocking and the start of shooting still need a consistent treatment.

Reopened checks find identical body vertices at the holding loop endpoints. The right-hand mesh also matches exactly between the final shot frame and the first holding frame. That is a partial handoff: the bow and support arm still differ, so the whole transition is not seamless. Both actions were sampled at 113 times without triggering the existing coarse body-edge stretch threshold.

The trigger remains unaccepted. Restricting the two distal index joints to their bending axes and limiting the knuckle leaves the current grip about 3.8 cm short of the selected contact target. This probe did not replace the old solver in the animation. It shows why reaching a target with unrestricted joint rotation was insufficient. Grip and contact must be revised together on the existing Meshy models. Up/down shots still need the body orientation seen in the original. Subsequent directional candidates follow below. No mod was installed and no new Meshy credits were spent.

Torso orientation for directional shots

The original up/down shots also change the shooter’s orientation. Each new candidate retains eight frames, turning the torso during aim before lifting the bow, withdrawing the rear hand and returning. Together with the front shot and holding, this body now has four offline groups totalling 32 frames. That count describes directional candidate 369; subsequent walking work follows below.

The first directional draft rotated the entire character about the ground axis. Its attachments stayed aligned, but the feet swivelled with it, so it was discarded. The replacement turns the lower spine while preserving the leg pose; hands and weapon follow the torso. The downward turn was also reduced because the initial pose faced the camera too directly. These angles are authored against the images, not recovered original 3D parameters.

Both clips play at eight frames per second for review. Reopening each scene at 113 times found the selected 7,704 boot vertices unchanged from the front action. Hands, bow and bolt differed from the expected torso transform by less than 0.007 mm, and the existing coarse body-edge stretch threshold did not trigger. These checks cover assembly behavior during turning; grip anatomy, cloth appearance and pose fidelity still need review.

The configuration draft now includes all three launch offsets, each calculated at release frame 4. Loaded-bolt foreshortening differs by direction, so the separate projectile’s display size still needs corresponding calibration. The known finger-joint problem, loading transition, remaining actions and native validation are unresolved. This pass reused the Meshy models, with no new API charges or VCMI source changes.

Walking and movement transitions

The original walk carries the crossbow across the waist, unlike the lowered holding pose. The new eight-frame cycle uses a two-handed carry with alternating steps. Local leg IK follows authored ankle paths, lifting the swing foot while moving the supporting foot backward. The gaze now stays forward; the first draft had inherited the shooting head tilt.

Move-start and move-end each retain two native frames, connecting the lowered bow to the waist-level carry. This preview uses those two start frames, eight walking frames and two end frames, with a brief holding pause at either end. Playback is eight frames per second for inspection, not a claim about the game’s final timing.

The first foot path sat too far behind the hips and exceeded the rear leg’s reach. A separate pelvis-update mistake first failed to apply the intended motion, then accumulated displacement and sank the body over successive poses; those drafts were discarded. Each pose now solves pelvis, ankles and grips afresh, correcting ankle height against the deformed soles. Reopening the cycle at 129 times reduces the earlier roughly 3.8 mm floor penetration to under 0.04 mm. The auxiliary cycle endpoint matches the starting body vertices exactly.

Each transition was sampled at 65 times. Auxiliary endpoints for body, hands, bow and string differ from their neighboring actions by less than 0.001 mm; the exported groups still contain two frames each. These are offline geometry and ground-clearance checks. Native movement speed and the known finger-anatomy problem remain unresolved.

This body now has seven candidate groups totalling 44 frames: holding, walking, movement start/end and three shooting directions. Hit, defence, death, melee, turns and other actions remain unfinished. Nothing was installed; the existing Meshy models were reused without additional API charges. Historical state at candidates 398/400; hit and defence drafts have since been added, as shown in the current table.

The original Griffin has eight holding frames and four walking frames; the Swordsman has eight of each. Both tests use the original 450×400 canvas. Each now has a stable local holding review: the Griffin uses continuous weights for wings, neck and tail, while the Swordsman keeps sword, shield and armour intact in the rest pose.

The Pikeman’s first ten-frame CPKMAN.DEF front-thrust trial keeps the long weapon attached but rotates it around only one hand. The shaft cuts through the torso and the free hand never supports it. A second trial added a two-bone grip constraint, but it pulled apart the rigid sleeve and tabard components without producing a natural thrust. Both are rejected; the next attempt needs a component-level two-hand hierarchy.

A second full Meshy Pikeman candidate explicitly generated a complete visible pike with two grip points and passes static review. Its automatic humanoid weights still do not provide a controllable two-hand attack chain: the first ten-frame action reads as an upward lift, and a later axis probe still cannot reach the original horizontal rightward impact. All three dynamic attempts are rejected and will not export. The next pass needs a dedicated hand-to-pike constraint reconstruction. See the current articulated-thrust work.

The first Griffin trial rigidly parented 2,829 disconnected components to bones. Its chest feathers separated in motion. Continuous surface weights repaired holding, but its gait then pulled apart layered surfaces near the legs and tail, so the four-frame walk was rejected.

A later Griffin rebuild split the leg surface across upper, lower and foot bones, then rigidly assigned low claw islands to the foot bones. Side frames still left a front claw and rear-foot fragments behind during the lift, so both versions were rejected. They are not DEF output.

A second 20k-quad Meshy Griffin candidate, generated from the same reviewed full-body concept, reduced disconnected components from 2,829 to 1,707. Its four-frame side gait still separates foreclaws and shin-feather islands from the lifted foreleg in frame 2, so this 30-credit candidate is rejected as well. It will not replace the earlier mesh or enter DEF assembly.

The Archer’s first eight-frame upward-shot draft and a four-axis pose probe are rejected. X/Y arm rotations pull the arms away from the torso; Z rotations move the crossbow but break the two-hand hold. A component-level weapon-and-hand rebind is required before an upward shot can be authored.

The Swordsman’s first seven-frame downward attack draft was rejected after comparing it with exported CSWORD.DEF references: it merely lowers the sword behind the body rather than moving from a high ready into a low impact. Axis probes are retained privately; this group needs a rebuilt shoulder–elbow–wrist chain.

The Swordsman’s eight-frame gait was also rejected: layered components around the skirt and shield side separate. Its holding and walking Actions are saved independently; the rejected walk will not enter DEF assembly.

Update: the Swordsman now uses a Meshy humanoid rig (an additional 5 credits), with Astra authoring the local holding and walking Actions. Both eight-frame clips were reviewed at 450×400. The left and right feet travel 0.105 and 0.089 model units front-to-back, while the sword hand travels 0.072. Side review keeps sword, shield and skirt continuous, so this local walk review is accepted. It is still not DEF output or a game installation.

The Swordsman now also has a seven-frame CSWORD.DEF front-attack review. The wind-up carries the sword behind the body, but front and side checks keep the blade connected to the hand and preserve the shield; up/down strikes and the remaining groups still need authoring.

The Swordsman also now has an independent seven-frame CSWORD.DEF upward attack review. Its high-ready motion resolves into an elevated forward strike while the blade, hand, shield and skirt remain continuous in front and side checks. Downward attack and remaining groups still need authoring.

Crusader and Monk each now use their own Meshy humanoid rig (an additional 5 credits each); neither reuses the Swordsman’s mesh or weights. The Crusader’s eight-frame holding and walking reviews pass intact. Its seven-frame double-strike test is rejected because part of the sword remains weighted to the body chain, and will return after a local rebind. The Monk’s six-frame holding and walking reviews pass; its ten-frame front cast is saved as a gesture calibration. The robe remains intact, but the raised-hand amplitude needs refinement, so it is not a final spell delivery. This was a continuity review of the gray-robed model, whose identity was rejected on September 18, 2026.

These trials show that a convincing Meshy surface is not automatically an animatable continuous character topology. The next pass will rebuild local weights from visible joints and connected regions. Until then, these are Blender review assets only: no creature frames, shadows, overlays, DEF output, or game installation.

An Angel two-arm sword probe shows the sword remaining on the body chain while both arms move. It is rejected before attack authoring; a local component-level sword rebind is needed.

After static review of the Angel Meshy humanoid rig, Astra did not rotate the whole wing as one object. It located wing components by position among 1,356 connected components of the shared surface. Four local root/tip bones control 10,319 wing vertices; an eight-frame restrained wing review keeps armour, two-handed sword and torso still. This is a holding-wing result only: flight and sword motion remain to be authored.

Angel now also has a seven-frame CANGEL.DEF moving review: root and tip bones drive the existing classified wing regions through a full flap, while the torso and two-handed sword remain continuous in front and side checks. It is flight motion only, still without the remaining action groups, creature layers, DEF output or game installation.

Archangel completed the same local wing process on its separate mesh: 8,888 position-classified wing vertices are assigned to four root/tip bones. The eight-frame review keeps upgraded armour, torso and two-handed sword stable. It does not reuse Angel vertex groups, and is not flight or game installation.

Archangel now has its own seven-frame CRANGL.DEF moving review, built from its separate wing mesh and weights. Front and side checks keep both wings, upgraded armour, torso and sword continuous through the flap. It does not cover its remaining action groups or game-ready layers.

Cavalier and Champion each ran a local horse-and-rider probe against their own meshes and original alpha-height anchors. Holding, walking, front lance, move start and move end pass crop checks; a three-frame side review also keeps the horse legs, rider, barding and lance connected through the gait. This is still a probe: each needs its complete 87-frame action set and per-action timing before export.

Royal Griffin uses its own mesh for four root/tip wing bones and 24,793 component-classified wing vertices. Its eight-frame holding wing review passes; its ground gait will not reuse the rejected Griffin leg/tail weights. Rechecking CRGRIF.DEF corrects the premise: its four-frame MOVING group is airborne flight, not a terrestrial gait. The first split root/tip flight attempt tears neck, chest and wing-root components, so it is rejected. A separate 30-credit Meshy candidate now passes initial eight-view static review with its airborne wings, talons, rear lion legs and tail present. Its replacement assigned complete connected wing components to one shoulder bone per side, which removed the earlier wing-root tearing; chest feather defects remained in side review. Candidate 03 was generated from a front-facing clean-plumage concept (30 credits): the chest improved, but side review exposed a near-planar wing spread, so it too is rejected. Candidate 04 was generated from a strict three-quarter-volume concept and retained a real lion torso and layered wing roots, but chest mesh holes persisted after both shallow-geometry and UV-only repair probes. It is rejected; Royal Griffin now moves to component-model reconstruction before a four-frame airborne review. No rejected candidate has DEF output or game installation.

Archer motion repairs

The replacement Meshy body, separate hands, light crossbow and dagger now have offline action drafts. Reading CLCBOW.DEF again establishes the complete scope as 16 active groups and 96 frame slots, excluding duplicate turn groups 9 and 10. A six-frame death draft also exists. Appearance and transitions remain under review, and the Archer has not been installed.

The three shooting recoveries now move the crossbow and both hands outward while retaining the original aim. Independent checks caught a regression: rewriting object motion had dropped the string_draw property curve that drives the bow’s shape keys. Restoring its six original keys and reopening each scene recovered the original string deformation. Across 225 sampled poses per direction, the tested body and crossbow surfaces do not intersect; the largest wrist-to-bow position error is about 0.164 mm. Finger contact and full in-game behavior are outside those checks.

A selection-motion trial rotated the crossbow clear of the body but made it too upright and was rejected. The current draft keeps the previous orientation and raises both hands outward during the middle of the motion, limited by arm reach. Its 257 sampled poses have no tested body/crossbow surface intersections, with a maximum relative wrist error of about 0.425 mm. Bow angles in the other phases and the free right-hand motion still need comparison with the native animation.

The melee drafts now use the idle wrist-to-bow grip with 30° of left upper-arm abduction. Untouched body and right-hand animation curves are preserved. Across 161 sampled poses per direction, the tested body/crossbow and crossbow/dagger surfaces do not intersect. Maximum left-grip position error is about 0.064 mm; the right wrist matches the original position and orientation.

The wider equipment check exposed an existing dagger-path problem: forward and downward attacks still intersect the body at 12 and 9 sampled poses respectively. Those swings need repair. Earlier translation, wrist-rotation and unmodified idle-arm trials were not adopted. Meshy supplied the models and base rig; Astra authored these Blender revisions. No additional Meshy requests or game-asset installation occurred in this round.

Monk identity correction (2026-09-18)

Current status: the Monk is installed as a private test; see installation and runtime checks. The following sections retain the offline development history, including their earlier uninstalled status.

The original CMONKK.DEF also exposes a design mismatch in the old Monk. It wears a plain brown hooded robe and a pale rope belt, with a shadowed face and both arms participating in forward and raised casts. The old model has a gray-black robe and long stole, with motion built largely around one hand. Earlier skinning and continuity checks did not establish reference fidelity; those trials remain in the history above.

Built-in imagegen produced a new reference constrained to the brown robe, short shoulder cowl, pale rope and non-glowing face. A relaxed low A-pose separates the sleeves from the torso for modeling; it is not the intended holding animation. The Meshy 7 API request used a 300,000-triangle target, 4K textures and disabled image enhancement, costing 30 credits. The returned static mesh contains 311,994 faces. Humanoid rigging cost another 5 credits.

Astra then posed both arms in gathered, forward and raised positions, checking two views of each. These probes avoid the large stretched sheets seen in the previous candidate, but finger contact and the cuff-to-cowl relationship need closer work. An initial scale check mistakenly included an untextured origin sphere supplied with the rig. Excluding that 80-face helper restored character-only height normalization; the original GLB remains intact.

The new rig now has four body-action drafts: holding (6 frames), forward casting (10), upward casting (10) and downward casting (9). Astra authored the motion in Blender using the existing Meshy geometry and rig; this stage made no new paid Meshy requests. These 35 rendered body frames cover four of the required fifteen groups. The Monk is still not installed. This was the first four-action checkpoint; the current nine-action scope is described below.

The first folded-arm pose left both palms facing outward. A hand-axis correction brought the hands across the opposite sleeves. A separate proportion pass widened the model in the horizontal plane: its projected holding width was about 28 original pixels, against 33 in the reference at the same 78-pixel height.

The forward cast now keeps the preparation longer and reaches forward on frames six and seven. Upward casting raises both hands; downward casting bends and turns the torso. The first upward trial looked high enough in a close view but read too flat through the game camera, so the revised motion raises the hands further.

All 35 body frames fit the 900×800 canvas. Forward and upward endpoints match holding pixel for pixel; downward endpoints differ at 30 pixels by at most 2 on an 8-bit channel. Reopening the saved scene and sampling the four actions at quarter-frame intervals found no adjacent keyed quaternion sign reversals. These checks cover framing and continuity, not clothing collisions or final appearance. Fingers, sleeve/cowl deformation and reference fidelity remain open; the other eleven action groups, effects, shadows and game integration still need work.

The current scene contains nine body-action drafts: holding, three casting directions, three melee directions, recoil and defence. That is 76 frame slots from 50 body renders. This records the nine-action checkpoint; five further groups follow below. Pixel comparisons against the original establish that melee reuses the casting poses: the forward sequence drops two preparation frames, upward drops one, and downward retains the same nine. The new melee actions preserve that timing in separate editable Blender actions; effects and projectiles still require separate treatment.

The shoulder-weight experiment below was rejected. Moving upper-arm influence toward the chest preserved holding within a micrometer, but produced a horizontal shelf at the shoulder when raised. The working scene retains the earlier weights. The first recoil draft also buried the lowered hand in the waist; the revised target moves it outside the torso.

Reopening the saved scene confirmed all nine actions were retained. An earlier library-import attempt discarded unused actions on save; the corrected import explicitly retains them. The fifteen new recoil/defence body images fit the canvas. Their endpoints differ from holding at five pixels by at most one 8-bit channel value. Walking, mouse-over, death, both turns and movement start remain unauthored on this model, along with cloth repair, effects, shadows and native integration. The Monk is not installed.

Five further body drafts add the six-frame walk, two-frame movement start, two frames for each turn, and ten mouse-over frames. The saved scene now retains fourteen actions, covering 98 original frame slots with 72 rendered body images. Death remains the one missing body group; clothing, effects, shadows and game integration are still unfinished.

The first walking trial raised the knees too far and pushed a large bulge through the robe. A smaller stride and lower foot lift reduce that deformation while retaining the folded arms. The two-frame turns pass through a front-facing pose, matching the original reference sequence. Mouse-over adds a restrained head movement.

All 22 new body images fit the fixed 900×800 canvas. Movement start ends on the exact first walking image. The foot check confirms alternating lifts of about one centimetre, with the support foot near the floor; a few body vertices still dip by up to one millimetre. This is an offline geometry check, and does not establish a slide-free walk at VCMI’s playback speed. The next body action is the eleven-frame collapse; its first five original frames are pixel-identical to recoil.

The eleven-frame death sequence now has a first motion trial. Its first five poses follow recoil; the remaining six spread the arms, bend the knees, fold the torso and roll onto the side. Reopening the scene confirms fifteen editable actions with the original 109 frame slots, and the eleven new images fit the canvas. The death trial has not passed visual review.

The first versions stopped in a crouch or kneel. Inspecting the actual bone hierarchy showed that this rig names its lower spine Spine02 and its upper spine Spine. Redistributing the bend toward the waist and adding a final side fall changed the pose, but exposed severe robe deformation around the folded legs.

A volume-preserving skinning trial did not fix the skirt: the overlapping strips remain. A separate check found identical weights at coincident seam vertices, so averaging seam weights would not address this defect. The next repair needs to handle the robe independently of the deeply bent legs. The separate-surface trials below follow this checkpoint. The death trial is kept separately from the preceding fourteen-action scene; no game installation was made.

A separate skirt for the Monk (2026-09-18)

Further weight and smoothing trials did not repair the folded robe. A dedicated skirt bone still left distorted boundaries; broader selections pulled the sleeves, and stronger smoothing softened the texture without restoring a usable collapse. The original Meshy model and the preceding fourteen-action scene are preserved.

A new surface follows the existing robe with 5,248 vertices and 5,120 quads, leaving the hem open. Its 2048×2048 colour texture was baked from the Meshy material. This isolated Blender render shows the transferred brown fabric and pale cord on the new surface; it is an intermediate mesh, not a finished unit.

Separating the skirt from leg skinning removes the earlier strip-like stretching in the kneeling pose. The sideways collapse still leaves a rigid tube, however. Flattening it exposes gaps between the boots and the body. The latter two candidates in this comparison were rejected.

A cloth trial spreads the eleven death poses over 101 simulation frames, pins the waist and uses simplified leg, torso and floor collisions. The bake completes, but the skirt contracts into the body. Reopening the saved scene and measuring the evaluated mesh confirms the contraction; the completed bake does not establish a usable result.

The separate surface and transferred texture remain available for further repair. The simulation shrinkage and the connection around the boots need attention before reviewing the other actions with this mesh. The diagnosis and follow-up trials are recorded below. Existing Meshy outputs were reused without new paid jobs. The Monk remains uninstalled.

Incorrect collision lengths on the Monk (2026-09-18)

The contraction investigation found a concrete error in my local collision builder. It used bone tails from this scene: the left thigh’s tail was roughly 34 metres from its head, while the knee joint was about 34 centimetres away. The collision surface was therefore about a hundred times too long. A run without collisions did not produce the same contraction. Using adjacent joint heads restores a skirt that bends around the knees. The existing Meshy model is unchanged.

Removing the original robe also exposed missing internal surfaces between the body and boot openings. Dark-brown leg geometry now fills that space. A separate self-collision and forearm-collision trial still turned the hem outward and was rejected. The current candidate uses the corrected joint collisions and adds forward tilt to the last two poses, lowering the previously raised head. Waist seams, lifted boots and the final curled silhouette remain unaccepted.

Eleven body frames were exported again with the battle camera at 900×800; all remain inside the canvas. The complete comparison still shows differences from the original poses in the second half. This is an offline death trial. The preceding fourteen-action scene and the installed game resources have not been replaced.

The reusable joint collision builder now checks segment lengths across the requested frames before creating proxies. Its synthetic-rig regression needs no game assets. Models and complete mods remain local.

Monk waist attachment and raised-arm direction (2026-09-18)

The skirt’s upper 640 vertices now follow sampled points on the original body, retaining their holding offsets. Across the eleven sampled poses, the two fully pinned rings’ maximum error relative to those targets falls from about 3.9 centimetres to less than 0.001 millimetres. This verifies the attachment targets, not the appearance of the waist folds. A longer hem also reduces the exposed boot openings in the standing pose.

A colour-and-foot-weight cleanup removed 315 suspected old robe faces in a separate trial, without visibly repairing the collapse folds; it was not adopted. Comparing the full sequence exposed another error: frames seven and eight raised the opposite screen-side arm from the original. Swapping the raised arm initially put the hand against the hood. Lifting and opening the arms improves that direction, while the high-resolution view shows distorted, layered folds at the shoulder/sleeve connection.

The following pixel comparison records the arm-only edit before refreshing the cloth binding. All eleven frames were rendered again. Pixel comparison with the longer-skirt candidate confirms that this arm edit changes only frames seven and eight; the other nine match exactly. A forearm quaternion sign flip was also corrected before export. The timing of the body lowering, final collapse, shoulder connection and clothing still differ from the original. This remains a local trial, with no installed resource replacement.

The combined scene resamples the waist references to account for the few millimetres of body movement introduced by the arm changes, then rebakes the full cloth simulation. The sequence below now shows both repairs together.

Torso lean and arm spread in the Monk death sequence (2026-09-18)

Enlarging frames seven through nine showed that the waist heights in seven and eight were already fairly close to the original. The clearer differences were the missing lean toward screen right and arms drawn inward too early in frame nine. This trial keeps the waist position and revises the torso and wrists.

Six wrist reference points were selected manually from the original three frames. After rotating the torso, the first arm solve kept the previous wrist depths and could not reach several targets. Adjusting depth along the camera rays reduces their mean projected error from about 19.6 to 2.8 logical pixels while retaining bone lengths. This describes the six manually estimated wrist targets only; it is not a score for overall pose fidelity. The depth-selection helper is now public, checked with synthetic cases and these six wrist solves.

The torso rotation also requires new waist targets and an updated torso collision surface. These were refreshed before rebaking the 101-frame cloth simulation and exporting eleven 900×800 body images, all within the canvas. Reopening the saved scene reproduces frame seven pixel for pixel. The final collapse silhouette, shoulder and waist folds, palm orientation and integration with the other actions remain unfinished. Game resources have not been replaced.

Hood and inner lining revisions

The preceding figures retain the earlier hood geometry. Lowering its tip alone produced a backward-pointing spike; that candidate was rejected. Moving the upper tip toward the head as well produces the shorter hood below. Its height and match to the original silhouette still need review.

A triangular patch beside the upper leg came from the inner lining intersecting the robe. At four sampled camera rays, the lining was about 1–4 mm in front of the robe surface. The upper lining radius is now reduced by 18%, tapering smoothly to the unchanged knee. The edit covers all twelve shape keys and preserves the dimensions at the boot opening. After reopening the scene, all four patch samples and two adjacent control samples hit the robe first. The rendered triangle is gone.

All eleven exported body frames fit the canvas. The final pose still lifts the boots too far and bunches the clothing around the waist; it has not passed visual review. This describes the preceding candidate; the next section records the subsequent foot and skirt revisions. This work reuses the Meshy model, with Astra editing and rendering the Blender geometry. There were no new paid generation jobs or changes to installed game assets.

Foot placement and skirt folds during collapse

Both boots were suspended in the previous final pose. Revising lower-leg and foot rotations brings the soles close to the floor while retaining bone lengths. Editing only the last two native poses left a roughly 2 cm floor penetration during interpolation. Additional transition keys address that dip, with matching updates to the inner lining and collision surfaces.

A self-collision experiment took longer to simulate and retained tangled folds, so it was rejected. The current candidate smooths the free portion of the simulated skirt, excluding the fully pinned waist vertices. This reduces the fragmented folds near the hem; the waist seam remains rough. The following section records its subsequent repair.

Reopening the saved scene reproduces the native-camera final frame pixel for pixel. Across forty-three poses sampled at half-frame intervals over the final part of the motion, vertices with combined foot/toe weight above 0.55 remain above the floor, with about 5.1 mm minimum clearance at the current model scale. This check covers the feet, not all intersections in the character. All eleven exported frames fit the canvas. Waist seams, palms, hood proportions relative to the original and adaptation of the other fourteen actions to this skirt remain unfinished. The following section records the waist repair and adaptation of the other fourteen actions; palms, original proportions, effects and integration remain unfinished. Installed game assets are unchanged.

Waist seam and full body-action drafts

These figures document the waist revision, before the shoulder correction in the next section.

The jagged waist edge came from the junction between the retained body and the separate skirt. Deleting whole triangles had left the boundary following irregular mesh edges. The revised body is split along a defined waist plane in its standing pose, interpolating UV coordinates and bone weights at new vertices before converting them back to animated model space. Skirt attachment offsets now follow the corresponding source vertices’ skin transforms as well.

Tucking a wide band of the skirt inside the body produced a dark horizontal line while standing, so that trial was rejected. Narrowing the overlap removes the line and the jagged collapse boundary in the two views below.

The body, lining and separate skirt are now adapted to the other fourteen actions. Each action gets new waist targets, inner-leg poses, collision surfaces and its own cloth bake. Holding and walking simulate three cycles and export the third. The maximum skirt-vertex displacement across the holding wrap drops from about 3.3 cm to 1.8 mm; it is not perfectly seamless.

The original’s fifteen effective groups now have 109 body images at 900×800, excluding duplicate turn aliases from the count. Every image fits the canvas. Bone transforms at the 98 native poses in the other fourteen actions match the preceding local skeletal drafts exactly. Each of the fifteen scenes was reopened and one frame rerendered, reproducing its corresponding image pixel for pixel. These checks establish reproducibility, not fidelity to the original or visual acceptance.

This work reuses the Meshy model, with Astra authoring the Blender repair and action-adaptation scripts. Shoulder clothing, palms and proportions against the original still need review. Spell effects, shadows, transitions and native-game validation are also unfinished. Offline shadow candidates appear below; the other items remain open. The Monk has not been installed.

Shoulder volume during raised-arm poses

The junction between the mantle and upper arm becomes thin when the arms rise. Moving upper-arm weights to the torso produced a sideways flap, so that approach was rejected. Enabling volume-preserving skinning over the entire body filled out the shoulder but also moved the waist: selected waist-band vertices shifted by up to about 7 mm in the inspected upward-cast pose.

The current candidate blends the two skinning methods near the shoulders, excluding vertices dominated by head, neck, hand and forearm bones. Leaving the local blend enabled throughout still changed the collapsed shoulder. Its influence now increases smoothly with upper-arm elevation relative to the torso, returning to the original linear deformation when the arms are lowered.

All fifteen groups and 109 native body frames were rerendered. Ten frames activate the correction; the other 99 match the preceding images pixel for pixel. The 172,195 vertices outside the selected region retain their positions across all 109 poses. Each scene was also reopened and one frame rerendered, matching its saved image exactly. These checks bound the change; they do not establish visual acceptance of every mantle deformation.

The local skinning-blend helper is public. A synthetic mesh without game assets tests zero, full and partial influence, plus rejection of invalid settings. This round reuses the Meshy model with repair and verification scripts authored by Astra; no new paid model job was submitted. Mantle folds, palms, original proportions, spell effects, shadows and native-game validation remain unfinished. The next section adds offline shadow candidates; the other items remain open.

Separate shadows for all fifteen actions

The current body scenes now produce 109 geometry-projected shadow frames. Each action retains its own saved cloth cache. The exporter evaluates the deformed body, lining and skirt, projects them onto the ground along one fixed direction, and renders through the body camera. Existing body images remain unchanged.

All shadows use 50% opacity and a 1.6-pixel blur, with no per-frame resizing or recentering. Every composite fits the 900×800 canvas. Rendering the first frame of each action a second time gives identical pixels in all fifteen cases. This establishes repeatability; temporal stability during native playback still needs review.

Astra authored the projection and layer checks using the existing Meshy model. These are offline shadow candidates. Mantle folds, palms, original proportions, spells, transitions and game integration remain unfinished.

Directional casting effects

The original Monk starts its upward charge on frame 3, the horizontal charge on frame 4, and the downward charge on frame 2. White-blue pixel bounds provide a timing and size reference. Existing 3D stars and mist cores are positioned around the revised hands, with the evaluated body, sleeves and skirt supplying occlusion.

The first pass was too dense and cloud-like. Reducing the mist opacity and particle count made the second too sparse. The current trial restores some smaller particles while retaining larger stars. These are Blender effects composited over the saved body and shadow layers; the close-ups are enlarged for comparison.

Fourteen effect frames remain empty, matching the original timing, and every composite fits the canvas. Reopening each of the three scenes and rendering frame 6 reproduces its saved output exactly. The source body scenes and all 109 body images retain their hashes. Flight after release, effects required by melee or defence, appearance repairs, transitions and native validation remain unfinished. The following section adds flight, melee and defence candidates; appearance, transitions and native validation remain open. Meshy supplies the reused creature model; Astra authored the 3D effects and adaptation scripts.

Melee, defence and projectile layers

The three melee directions add 26 effect slots using the same 3D sparks as casting. Defence needs a separate shape: the original opens a curved band of light. A seven-frame star arc now renders against the current sleeves and body as occluders.

The first ward was a thin, predominantly golden line. The revision gives the particles a thicker distribution and restores white-blue stars. Brightness and the unfolding motion still need comparison with the original.

Flight uses the casting layer’s actual star meshes and materials. Several initial angles touched the 120×120 canvas edges. Reducing the overall scale leaves transparent margins at all nine angles. The larger still below renders the same 3D model; native projectile size and its release position remain unverified.

The consolidated offline inventory contains 109 body-and-shadow frames and 62 effect slots, including 27 original-timing blanks. Reopening the three melee scenes, the ward scene and the projectile scene reproduces one sampled frame from each exactly. Body source files remain unchanged. Packaging, release alignment and native checks come next; clothing, palms, original proportions and full visual acceptance remain unfinished.

Local Monk test installation

The private Castle mod is now 0.6.0, adding the brown-robed Monk at 1× and 2× with separate shadows, selection outlines and 3D spell projectiles. Its fifteen effective action groups contain 109 frames; duplicate original turn groups use the existing fallback behavior. Resource validation reports zero errors and warnings. The installation adds 491 files while preserving the hashes of 2,528 existing files. No VCMI source changes were made.

The original CRANIM.TXT fires the Monk projectile on frame 8, compared with frame 6 for the Zealot. The test keeps the Monk timing. The first release comparison exposed a size jump: flight was visibly larger than the charge. The median ratio of alpha-weighted radii across three directions reduced the flight model to about 63.3% of its previous scale, followed by nine new renders. Offsets align charge and projectile alpha centroids. This offline alignment still needs native visual review.

Native logs collectively read 109 distinct 2× body images across 15 groups, plus upward-diagonal, horizontal and downward-diagonal projectiles. The ranged test records 39 emissions. These were bounded observations, not proof of a completed battle or visual acceptance. Display settings and temporary creature parameters were restored byte for byte, and installed resources were checked again.

Defence requires a specific trigger. Choosing the defend action alone does not play the whole ward; VCMI uses that animation when a defending stack survives a melee attack. An initial test with neither side able to move or shoot did not cover it. A Marksman attack trial also failed to trigger it: ranged attacks count as indirect attacks in this code path. A subsequent setup uses a Crusader in melee. Effective groups still absent from the observed loading: none.

Window capture still reports could not create image from window, so the figures remain offline comparisons. Logs establish loading and execution; cloth, palms, proportions, transitions, ward brightness and native appearance remain under review.

Zealot identity correction (2026-09-18)

The original CZEALT.DEF wears a navy hood and robe with narrow gold trim; its face is a featureless light. The previous mesh had a white robe, green stole and visible bearded face. Earlier continuity checks did not establish a faithful design. Those motion experiments remain in the collapsed history below, but the white-robed mesh is no longer a candidate for final integration.

Built-in imagegen produced this construction reference. A direct Meshy API request then generated a Meshy 7 model with 20k-quad remeshing and 4K textures for 30 credits. The actual Blender still below shows the revised hood and palette, along with visible holes at the image-left axilla and sleeve cuffs.

The clay render confirms a geometry defect. Welding duplicate vertices and recalculating normals did not resolve it. A further 5-credit Meshy remesh to 50k triangles also retained visible holes. This candidate is not ready for animation binding.

A fresh Meshy request used the same concept with automatic remeshing disabled, retaining the dense reconstruction for another 30 credits. Its eight Blender views do not show the previous axilla or cuff holes. Because this was a fresh generation, the comparison alone cannot prove that remeshing caused all of the earlier defects.

The new mesh has 939,712 faces. The rigging API rejected it with HTTP 400 because it exceeds the 320,000-face limit. The 300,000-face result retains the closed axilla and cuffs in front/rear review, though small dark surface marks remain. The dense source is preserved for repair. Meshy completed the reduced model’s rig for 5 credits. The emissive face material, original folded-arm pose and deformation checks also remain unfinished; this is not an in-game delivery. Subsequent material and motion trials appear below; the model remains uninstalled.

Astra then authored a two-arm IK probe in Blender, checking the rest pose, a midpoint and crossed forearms from front and side. The wide sleeves do not show the earlier large tears in these three sampled poses. Both palms still turn outward, however, instead of resting inside the original folded-arm silhouette. This is an unfinished pose test, not a completed holding animation. The face remains a gold surface awaiting its emissive material.

The next local revision raises the crossed arms and turns the wrists inward. It adds six keyed holding frames and a ten-frame head-turn draft on the editable Blender rig. These use the original frame counts but enlarged offline framing, not the game canvas. The subsequent native-canvas comparison appears below. Too much of each hand remains exposed and the cuffs are more open than the reference; the pose is still under review.

The first emissive face selection missed much of the mask and accidentally lit the hood rim. Expanding that selection still leaked light. A mesh attribute now supplies a gradual emission mask, bright at the face centre and fading toward its edge, while retaining the trim’s ordinary material.

Exporting a 900×800 body frame for the original 450×400 logical canvas exposed a proportion error that the large previews had obscured. At roughly 80 pixels tall, the narrow candidate projected to only about 24 pixels wide; the original bounding box is 37 pixels wide. Widening the Meshy mesh and rig brought the body projection to about 34 pixels. The comparison holds height and placement consistent and is entirely offline.

Astra added a three-dimensional shoulder mantle, waist cord and hanging ends in Blender. A material mask fades the generated gold front seam below the chest. The additions still need repair: the mantle is too stiff and intersects the upper arm in places, while the folded hands and open cuffs remain unlike the reference. This is an unfinished shape trial.

The original fourteen-frame front cast gathers the hands, sends them forward at different heights, holds both arms raised, and recovers. The blue-robed rig now has separate 3D keys for that sequence. All fourteen frames and the forward pose’s side view were inspected. This is body motion only, without spell effects, a projectile or geometry shadows, and the cloth-intersection review remains unresolved.

Holding, fidget and front casting now have 30 RGBA body frames exported through one fixed camera at 900×800, all within the canvas. This verifies export and bounds for these three groups only. The other fifteen groups, shadows, effects and game integration remain unfinished.

Two mantle constructions were tested next. Copying and thickening the Meshy surface amplified fragmented seams and irregular boundaries. Welding, normal repair and smoothing did not produce a usable garment; the failed trial is preserved below.

The replacement uses a continuous low-resolution mesh shaped against the shoulder surface, with skin weights transferred from the body. It now deforms with the shoulders during the forward cast without the fragmented surface seen above. The lifted lower edge, fit from other angles and hand contact still need review.

The new six-frame walk exposed another binding error: each leg pulled part of the long robe, producing a trouser-like silhouette. A separate garment bone removed most of that deformation. Remaining failures came from missed hem vertices and coincident UV-seam vertices with different weights. An attempted cut treated the mixed-weight region as a boot seam and created holes; that edit was discarded. The current mesh is intact, with garment selection based on texture colour, position and bone membership, and matching weights at coincident vertices.

Walking is still not accepted: some hem edges stretch in the six-frame sequence. Floating soles were corrected separately by evaluating the sole vertices and adjusting the character root. The largest previous minimum-sole height was about 2.86 centimetres; the corrected minimum reaches the ground. That measurement establishes lowest-sole contact only, not complete gait or cloth acceptance.

Two-frame move-start, left-turn and right-turn drafts were also added. They remain offline; the transition through the engine’s facing flip has not been verified in game. No MOVE_END group was invented for a source that does not contain one.

The seven current groups now have 42 body frames re-exported through a fixed camera at 900×800, all within the canvas. Eleven original groups, shadows, effects and integration remain unfinished, and walking is still an unaccepted draft. This records candidate 545.

A follow-up on September 18 identified 54 missed garment vertices through mesh connectivity. Rebinding these islands to the robe bone preserved both the boot weights and the uncut mesh. The low-hem diagnostic counts edges longer than 3 cm and four times their rest length. Counts across the six walking frames fell from 111,134, 94,110,167 and 94 to 0, 22, 0, 0, 54 and 0. Two frames still contain roughly 3.6 cm stretches; the walk remains unaccepted.

The original hit sequence unfolds the arms, dips the head and raises one hand during recoil before returning to the folded pose. Defence separates the hands vertically in front of an arcing magical barrier. Those references now have eight-frame hit and seven-frame defence body drafts. The barrier has not been authored, and the raised arm still exposes angular mantle deformation and intersections in these actual Blender renders.

All fifteen new frames fit the fixed 900×800 export canvas. Each clip returns to holding with a maximum body-vertex difference of about 0.00037 mm; this checks endpoint continuity, not cloth intersections. The current inventory is nine draft groups and 57 body frames. Nine original groups with 93 frames, shadows, effects and game integration remain unfinished. The earlier 42 rendered images predate this hem repair and must be regenerated before packaging. This records candidate 551. Existing Meshy geometry and rigging were reused without new paid requests.

Upward and downward casting now each have thirteen body frames, with direction-specific facing, arm height and recovery. The upward pose turns farther from the camera; the downward pose bends toward it. Original melee counts are nine frames upward, eight forward and nine downward. Upper and lower melee recover sooner than their ranged counterparts, while the forward attack pushes both hands out horizontally. These actual Blender stills retain unresolved mantle intersections, cuffs and hand shapes.

All fourteen groups have now been re-rendered from the repaired scene: 109 body frames at 900×800, all within the canvas, with the complete contact sheets inspected. The five new clips add 52 body frames. Their endpoints match holding within 0.0004 mm at the body vertices. A separate RGBA comparison confirmed that all ten original special-action frames are identical across the three direction groups, allowing one shared body clip. Eleven death frames, thirty frames across the special groups, cloth repairs, shadows, effects and game integration remain unfinished. This records candidate 559. A mantle-weight smoothing trial preserved the holding shape and softened angular edges, but raised-arm intersections remained; it was not adopted as the current version.

A shared ten-frame special body clip now serves all three original direction groups. There are 119 distinct body images covering 139 frame slots across 17 native groups. The new clip fits the canvas and returns to holding; the eleven-frame death remains unauthored. This records candidate 563.

Spell effects are also built from three-dimensional Blender meshes. The first 49 glowing particles looked like white chunks and were rejected. A second trial used 181 crossed-axis sparks, but remained too small and sparse at native scale. The third expands the burst, adds five emissive core lobes and leaves the released tail in the air. It still lacks the original’s dense light and soft transition; mantle intersections also persist. These are unaccepted trials.

Body and effects live in separate editable scenes so particles cannot accidentally appear in other clips. The current exports are body images and combined review renders; game-ready effect layers and occlusion checks remain unfinished. This records candidate 566; the later isolated-layer trial is described below.

The eleven-frame death draft starts with the hit poses, opens the arms, loses balance and falls forward. Bent legs and a shortened robe bone gather the lower garment in the later frames. Its first render exposed an object relationship error: the independent mantle deformed with the bones but did not follow the whole-character root during the fall, leaving it suspended above the body.

Parenting the mantle to the armature object preserved the holding shape and made it follow the fall. Minimum-height checks across visible geometry found no substantial floor penetration in the eleven keyed frames; this does not establish correct hand or cloth contact. The corpse remains too extended for visual acceptance.

The first timing pass also reached the floor by frame nine, earlier than the original; a subsequent revision delays the main fall to the last frames. A further revision adjusts the falling root’s ground-plane travel with the camera held fixed. The original ending occupies 44×28 logical pixels; the draft is about 44.5×40.5 and remains too tall. These dimensions describe candidate 574. Body drafts now cover all eighteen native groups: 130 distinct images mapped to 150 active frame slots. Cloth, hands, effects, shadows and game integration are still unfinished. This records the outstanding work at candidate 574; shadow drafts follow below. Earlier exports and separate effect scenes have not yet been regenerated with the mantle root-follow repair. This records the export state at candidate 574.

Further trials showed that turning the corpse sideways or laying it flatter still left it too extended; one side-facing trial reached 90 logical pixels in width and was rejected. Folding the torso over bent knees produced a more compact pose. That stronger bend exposed another attachment error: the waist cord followed the chest and floated above the lower back.

The cord and both hanging ends now follow the pelvis while retaining their holding-pose positions. The revised pose has been incorporated into the eleven-frame death clip. Its ending occupies 44×33 logical pixels, closer to the original 44×28 but still too tall; hand and cloth shapes remain unaccepted.

An additional sweep checked 41 poses at quarter-frame intervals. It found roughly 1.6 mm of body penetration between keyed frames. The continuous motion therefore remains unaccepted even where the native keyed frames meet the floor.

All 130 body frame files have now been re-exported from one scene containing both attachment repairs. Every frame fits the fixed 900×800 canvas, and the full contact sheets were inspected. The mapping was checked against every original group: 150 active slots across eighteen groups, plus the duplicate turn slots referencing the same files. Separate effect scenes still need these attachment repairs; game integration remains unfinished. This records the synchronization state at candidate 581.

Two cloth simulations were then tested on the mantle. The front cast was stretched over 85 simulation frames, using a reduced body mesh for collisions. Pinning only the collar region let the first trial bunch into a scarf-like shape even at rest, so it was rejected.

The second trial pins more of the shoulder, increases stiffness and retains 15% of the body mesh for collision instead of 5%. It preserves more of the resting silhouette, but raised-arm intersections and unnatural edges remain. It has not replaced the current mantle.

Both trials are saved separately; current body exports still use candidate 581. The separate effect scene now includes the mantle-root and waist-to-pelvis attachment repairs. All ten combined frames were re-rendered, checked and remain inside the canvas. Effect appearance is still unaccepted.

The special action also has ten transparent effect-layer renders. The body acts as an occluder without appearing in that layer. Frames 1, 9 and 10 are empty; the remaining effects can be composited over the existing body frames. Recomposition still differs from the integrated scene render, with a peak-frame mean RGBA channel error of about 6/255 over visible pixels. These layers are not yet accepted for the game.

All 130 body frame files now have separate shadows. Each shadow projects the evaluated body, mantle and waist attachments onto the ground in a fixed direction, using the same camera as the body render. Denoising and animated random seeds are disabled. The masks then receive uniform 50% opacity and a 1.6-pixel blur on the 2× canvas.

Every shadow fits the canvas. Repeating the same holding pose produced identical RGBA pixels, and ten action endings matched the holding shadow exactly. These checks exclude random rendering differences and endpoint jumps in the tested poses; they do not establish flicker-free playback for the whole animation in game.

Defence also has a seven-frame curved magical ward built from independent three-dimensional star meshes. The first 151-spark trial was visible in a large render but too thin and faint at native size. The revision uses 241 larger sparks, leaves a released tail in frame six and has empty effect layers at both endpoints. All seven body, shadow and effect recompositions fit the canvas.

Only special actions and defence currently have isolated effect trials. Other casting and melee effects, the projectile, mantle intersections, hands, death pose and final layered appearance remain unfinished. The Zealot package is not installed in the game. This records candidate 592.

On September 18, candidate 595 added 3D effects for front, upward and downward ranged casts and melee actions. The six clips occupy 66 original frame slots: 26 contain effects and the rest are empty. Each slot has a full-scene render and a transparent effect pass with body occlusion, composited offline over the existing body and shadow. All effects fit the canvas. Visible-pixel mean RGBA error between recomposition and full-scene rendering reaches 4.79/255; these are still trial layers.

The comparison exposes smaller bursts than the original, while the high-resolution stars look too hard. Upward casts also place the hands and burst too low. Sleeves still intersect the mantle. Both images retain this unaccepted attempt. The original projectile has been extracted in nine directions for reference; its 3D replacement remains unmade, and the Zealot package is not installed. This records candidate 595. A projectile draft follows below; game integration remains unfinished.

The next revision raises both hands in upward ranged casting and melee, with 22 body frames re-exported. The first target shift exceeded arm reach and left the solver near a straight elbow. Candidate 599 caps targets at 94% of the combined arm-segment length. Keyed target error is below 0.001 mm and all 22 frames fit the canvas. This verifies target placement only; sleeve and mantle intersections remain.

Larger particles and a wider cloud then produced an overexposed white core in candidate 600. Candidate 601 reduces brightness and opacity, restores blue and gold, and rotates the elongated core toward the casting direction. It is still too blue and its high-resolution stars remain hard. Only frames four and five of the front and upward casts have been rendered for this comparison; the complete 66-slot export above has not been replaced, and revised body shadows have not been regenerated. This records candidate 601; candidate 611 below updates the complete export. These are actual Blender renders; the unit remains uninstalled.

The projectile now has a separate 3D model with 900 crossed stars and 15 transparent cores. The original CPRZEAX.DEF contains nine directions, and the Zealot row in CRANIM.TXT releases it on frame six. The angles are 90°, 72°, 45°, 27°, 0° and the corresponding downward angles. Candidate 606 renders all nine onto a 120×120 canvas for a logical 60×60 image, with a separate 1× export. All nine fit the canvas. Resources and candidate launch offsets are stored locally, not installed.

The first cloud was small and showed distinct core boundaries. The larger revision also exposed dark patches where transparent cores overlapped. Raising Cycles transparent bounces to 64, with geometry and materials unchanged, removed those patches. The comparison preserves the original, faulty 604 and revised 606. Spherical core boundaries and hard star shapes still need work at high resolution.

An offline handoff uses the fifth-frame charge centre to place the projectile on frame six, then illustrates travel by 16 and 32 logical pixels. Candidate offsets are (22,-54) forward, (13,-74) upward and (9,-34) downward. These come from the current camera and right-facing geometry, with no native validation yet. The downward charge still uses the earlier small effect and visibly jumps in size. The other directions also change colour. Charge and flight appearance need to be unified before native handoff testing.

Candidate 609 directly reuses projectile 606 geometry and materials, scaling by the pixel ratio between the orthographic cameras and rotating toward each casting direction. Effect layers share the projectile colour transform and retain body occlusion; existing body renders keep their own colour treatment. The downward size jump shown above is reduced in the new offline comparison. Fidelity and native handoff remain unverified.

All 66 directional effect slots have been re-exported, and every expected empty frame is empty. The raised-arm clips also have 22 regenerated shadows; all four start/end shadows match holding pixel for pixel. Candidate 611 consolidates 130 body frames, 130 shadows and 83 effect slots. Seventeen defence and special-action slots still use older trials. All composites fit the canvas. Cloth, hands, death and walking detail remain unfinished, and the unit is not installed.

A separate mantle trial adds pose corrective shape keys. Candidate 613 moves nearby vertices outward when their nearest body-surface normal indicates penetration, then inverts skin deformation to store editable coordinates. The diagnostic count falls from 96, 86 and 116 to 3, 10 and 5 across three poses. This heuristic does not prove collision-free geometry; front and side renders still show underarm folds.

Candidate 615 extends the correction to all 14 front-cast frames and fades it at the endpoints. Corrective weights sum to one at all 53 quarter-frame samples, and all native rendered frames fit the canvas. These checks cover weight interpolation and bounds only. Hiding the added mantle leaves folds in the original underarm garment, so sleeve deformation also needs inspection. Neither trial replaces candidate 611 or enters the game.

Further sleeve trials reject two shortcuts: preserve-volume skinning bulges the torso, while local smoothing creates broken underarm edges. Candidate 621 first welds coincident vertices, then applies the same smoothing to shoulder weight transitions. Vertex count falls from 247,757 to 132,649; polygon count stays 265,495 and the UV layer remains. The new raised-arm render avoids those broken edges, but folds still need work. Holding and front/upward casting total 33 checked native-canvas frames. This trial does not replace the main resource set.

The Zealot test package was installed locally on September 18, 2026. It uses candidate 611 layers and excludes the sleeve experiment above. The package covers 18 active groups and 150 slots, reusing 130 body images, at two resolutions, with geometry shadows, holding/hover outlines and nine projectile directions. All 1,401 existing files for the other four units are preserved.

The first test (624) reads 62 distinct 2× body images across holding, hover, hit, death, front shooting and upward shooting, plus two projectile directions. The actual screenshot below shows the new unit, corpse and flying cloud. It does not prove playback of the other twelve groups. This describes the first test; formation tests below extend its coverage. Cloth, hands, death shape, full action coverage and release handoff still need review. The client exited and original settings were restored. Logs also contain text-encoding errors and an exit-stage query error, so this is not an error-free run.

Tests 626 and 627 temporarily removed ammunition to exercise movement, starting, turns and all three melee directions. A seven-stack ranged battle in test 628 also loaded downward shooting. Across the four runs, logs identify 113 distinct 2× body images in 14 groups, plus the 27°, 0° and −27° projectile images. Defence and the three special groups still lack native loading evidence. Loading records alone do not establish correct display of every frame or acceptable motion.

The image below is an actual game-window capture from test 627. Test 628 produced logs but no successful window capture. Tests 626–628 required forced termination after their time limits; their original game settings and creature configuration were restored byte for byte. The installed unit retains its ranged attack. Cloth intersections, hands, corpse shape and release timing still need review.

The next modeling trial changes how the mantle follows the arms. Candidate 630 retains 100%, 35% or none of the upper-arm weights, transfers the remainder to the chest bone and solves vertex positions back to the same holding shape. In the existing binding, an edge measuring 4.24 mm in holding stretches to about 27 cm on upward-cast frame six. Removing upper-arm influence lowers the worst edge-length ratio in that clip from 63.9 to 2.05 and reduces the pointed folds in the gold hem. This measures mesh stretch relative to holding; it is not a cloth-collision or physical-simulation test.

Candidate 631 also tried projecting the mantle away from the body at front-cast, upward-cast and defence poses. Contact still looks unresolved, so those corrective shapes were not adopted. Candidate 632 uses the binding with no upper-arm influence and exports all 130 body images inside the original canvas. Complete action comparisons were reviewed, with another 472 quarter-frame samples per binding measuring mantle deformation. At candidate 632: This revised body is offline; the installed game still uses 611. Installation 638 follows below. Sleeves, hands and corpse shape need further review. Shadows and effect occlusion have not been regenerated for the new geometry. This records the layer state at candidate 632.

The mantle revision is now installed locally as test 638, in mod version 0.5.1. Its new body has 130 regenerated geometry-shadow frames and 83 effect layers rendered with the revised mantle as an occluder. All 45 expected blank effect frames remain empty; repeating the holding shadow produces identical pixels. Every composite fits its canvas. The 1×/2× body, separate shadows and outlines pass resource checks against the original action inventory with zero errors and warnings. Appearance remains under review.

Installation adds 552 files, switches the two CZEALT indexes and updates the mod version, preserving the other 1,974 existing files. Ranged and melee tests 639–640 read 113 body images from the new czealt636 resource directory across 14 groups, plus three projectile directions. Both clients required forced termination after their test limits; original settings and ammunition configuration were restored. Window capture failed in these runs, so the image below is an offline layer comparison. The earlier native screenshot from test 627 shows the previous version. Defence and the three special groups still lack native loading evidence; sleeves, hands, death shape and effects remain unfinished.

The original action inventory contains 18 active groups and 150 frames after excluding two duplicate turn slots. It has a two-frame MOVE_START and no MOVE_END; the earlier description incorrectly called our additional settling clip a native group.

The Marksman now has its own equivalent review on its separate Meshy rig: its CHCBOW.DEF front-shot group also has eight frames, and the raise, aim, recoil and return keep the light crossbow, sleeves and hands continuous. It has no bolt release or up/down firing groups yet.

Remote candidate 02 subsequently gained the native eight-frame up/down shooting body groups; bolt release still needs its own three-dimensional layer.

The Crusader’s first mesh attempted a component-level sword rebind: a screen-space selection included blade strips together with forearm and tabard fragments, creating separated ribbons during the seven-frame double strike. It remains rejected. A second Meshy candidate was generated from the same reviewed concept as a 20k-quad, 4K-texture mesh with image enhancement disabled (30 credits), then received a 5-credit humanoid rig. Its eight-view static mesh and rig-rest reviews keep sword, shield and cape distinct, but the first seven-frame double-strike still bends the sword’s layered surfaces through Meshy smooth weights. Removing its non-character Cube and Icosphere placeholders fixes an import artifact, not the blade deformation; a targeted sword rebind is still required. Neither candidate is DEF output or an in-game installation. An eighth local repair removed the known deforming blade sheets and added a rigid four-piece 3D sword parented to the right hand. Front frames kept the new sword intact, but oblique review still exposed two residual original blade strips below the knee. The attempt is rejected; those remaining source components must be removed before any double-strike acceptance. The ninth repair removes that narrow front-knee source region while retaining boots, cape and armour. The rigid four-piece sword now passes all seven front frames and all seven oblique frames: sword, right hand, shield and legs stay continuous through the double strike. Defence review subsequently exposed grip remnants and flexible shield deformation; those attack checks did not establish that the rig supports the full action set. This is Blender review only, without DEF output or game installation.

Crusader defence review — work in progress

The original CCRUSD defence group contains eleven frames: raise the shield, lower the blade, and recover. The imported skin weights fold the shield as the forearm rises. A dedicated shield bone now drives 1,453 vertices from the existing Meshy mesh, including the small trim pieces. Across eleven frames, 3,648 internal shield edges change length by at most approximately 0.003%. This measures shield rigidity only.

The front and side renders still show grey remnants on the shield and leftover geometry around the sword grip. Delivery remains pending. Rendering the shield alone traced the grey patches to the hand penetrating its surface. Two sampled rays measured roughly 1–1.5 cm of penetration; adjusting the holding offset clears it. A grip close-up also showed that the original Meshy handle was intact. The revision retains that handle and the hand, removes the duplicate procedural handle, and replaces only the damaged guard and blade. A narrower selection left small trim fragments floating outside the shield and was rejected; the retained candidate includes those pieces. The revised candidate has eleven front and eleven side renders. Sampled start, raise, block and recovery poses no longer show the identified floating blade remnants. Other actions still need the same attachment fixes; there is no DEF export or game installation. The failure images remain below, followed by the revised Blender renders.

Holding, recoil and walking studies

The repaired equipment now carries eight holding frames, six recoil frames and eight walking frames. An early recoil draft lifted the leg forward. Reviewing the original sequence corrected that to a forward torso lean, rear leg lift and extension, with the shield arm thrown backwards. The larger motion still exposes an unresolved glove/shield-edge contact problem.

The walk uses two-segment leg solves and changes the sword from raised to lowered carriage across the cycle. Each wrist travels opposite its same-side ankle. An extra ninth key closes the eight-frame loop; it is not an exported native frame. Endpoint vertex positions match, and sampled stance soles stay within roughly 1.2 mm of the ground. These checks do not establish native gait fidelity. Final frame-by-frame equipment review, movement transitions, attacks and death remain pending, with no game installation.

Attack timing correction

The old seven-frame draft contained two swings. Inspecting the original ATTACK_FRONT sequence showed one raise, slash and recovery, so the earlier double-strike acceptance is withdrawn. The replacement keeps seven frames and authors a single slash. Raising the sword overhead also exposed source handle pieces following forearm weights; binding them rigidly to the wrist removed the separated pommel in the inspected front and side wind-up views. The lower-body lunge, full sword path and the other attack directions remain unfinished. Three directional candidates now provide seven frames each on the repaired grip model, with a wider planted lunge. All 21 poses were inspected in front and side contact sheets; sampled low-foot vertices remain roughly -0.8 to +0.9 mm from the floor. The original anticipatory leg lift, torso turn and shield-arm coordination still need closer matching. These remain production candidates without game integration.

Full-set draft review

The Crusader now has drafts for thirteen standard groups, totaling 76 native frames, with an explicit action list for one shared model. Turning follows VCMI’s two TURN_L frames, facing flip, and two TURN_R frames. The middle pose brings the equipment closer to the centerline; hover has eight separate frames. The six-frame death study falls backwards and ends face-up with lowered arms and equipment beside the body.

This completes draft coverage only. Glove/shield-edge contact, corpse support, final camera alignment and transitions still need full-set review. Later hand-weight edits also affect earlier actions, so their individual review results cannot establish that the combined set is ready. There is no DEF export or game installation.

September 17 review: after adjusting the shield, the selected hand/forearm and shield triangles show no crossings across the 76 frames. This excludes other body regions and does not establish a natural grip. A ground-plane render also exposed a floating corpse: placing the lowest equipment or hand point at ground level left the body elevated. The body has been lowered and the arms repositioned.

All 76 body frames have been rendered on transparent 900×800 canvases, with holding height and foot placement registered to the original. No geometry exceeds the canvas. Comparing them with the original exposed another problem: the camera was too frontal, making the front strike point toward the viewer. The revised camera below gives a more side-on view. Each pair shows the original on the left and the new model on the right. Turns still need adjustment for that camera; shadow and outline layers and in-game validation remain unfinished. The Crusader is not installed in the game. Turns have since been adjusted for the camera and the full set rendered again, with shadows projected against a fixed ground line and precomputed hover outlines.

The Crusader test package is now installed locally as a separate castle-creature-animations mod. It contains 76 frames at each of 1× and 2× resolution. All 340 installed files match the package, and validation against the original animation groups reports no errors or warnings. A fresh client confirms that the mod loads. Battle playback remains unverified, including movement-to-stop transitions, repeated attacks, shadows and the creature information panel. A subsequent isolated test map completed a battle and the client read 41 distinct 2× body frames with their effects, covering holding, hover, moving, move start/end, downward attack and death. System capture returned a black image and window capture failed. This verifies battle execution and resource loading only; visual behavior, other directions and the creature information panel remain unreviewed. Other Castle creatures still use their original artwork in the game.

One export also caught a Blender persistence failure: the revised turn played in the authoring session but its action disappeared when the file reopened. The actions are now explicitly retained, and the complete export ran from the reopened scene.

The next two images preserve the earlier turn and death candidates, before the contact and support corrections described above.

The next three images preserve the earlier double-swing draft; that timing has been discarded.

An eighth local repair removed the known deforming blade sheets and added a rigid four-piece 3D sword parented to the right hand. Front frames kept the new sword intact, but oblique review still exposed two residual original blade strips below the knee. The attempt is rejected; those remaining source components must be removed before any double-strike acceptance.

The Swordsman now has a native six-frame HITTED review: a visible rearward recoil and raised shield keep the original defensive silhouette readable. Front full-sequence and side start, peak and recovery checks keep sword, shield, hands, armour and tabard continuous. It remains a Blender review, without DEF assembly or game installation.

The Swordsman also has a native six-frame DEATH review. It physically collapses diagonally into a low horizontal final silhouette rather than fading out; front full sequence and side start, mid and end checks keep sword, shield, hands, armour and tabard continuous within the canvas. It remains a Blender review, without DEF assembly or game installation.

The Swordsman’s first eleven-frame DEFENCE trial raised the sword sideways; a second reused the upward-strike axis but left it behind the shoulder. Neither matched the original high guard. The accepted native-count third review raises the complete sword diagonally before the head while the shield closes over the torso. Front full-sequence and side start, peak and recovery checks keep sword, shield, hands, armour and tabard continuous. It remains a Blender review, without DEF assembly or game installation.

The Swordsman’s second seven-frame ATTACK_DOWN rebuild added the correct high-ready pose, but its low impact still carried the sword behind the torso instead of the original forward/downward cut. It is rejected alongside the earlier single-axis attempt; the next version requires a local forward hand rebind. A third through sixth attempt removed the low-impact lateral rotations and probed connected-bone translation plus imported hand offsets. They keep sword and hand connected, but side review still places the chain behind the torso; the old mesh/rig cannot supply the original forward/downward cut and will not be used for this group. A separate Meshy candidate with a cleanly isolated sword, hand and shield is required. Candidate 02 now passes static eight-view and Meshy humanoid-rig rest review: its blade, right hand, left-hand shield and torso remain distinct from every side. It used a 30-credit 20k-quad/4K Meshy mesh task and a separate 5-credit humanoid rig. A rigid rebinding of the two sword components prevents them from blending into the torso and gives a correct forward low-impact probe, but the automatic shoulder/elbow axes cannot also form the high-ready pose. A local sword-control bone derived from the right hand is now required; no new attack group is accepted yet. The first rebuilt motion still placed the blade behind the body; a local SwordControl bone then isolated the components but inherited the malformed imported hand-scale axes and could not make a reliable high-ready pose. The accepted repair separates the two Meshy sword components as an independent 3D object and rotates it at the measured grip pivot. Its native seven-frame ATTACK_DOWN now has a high-ready, forward/downward impact and recovery. Full front frames plus side start, high-ready, impact, recovery and end keep sword, right hand, shield and torso continuous. This remains Blender review only, with no DEF output or game installation. Candidate 02’s independent seven-frame ATTACK_UP subsequently passes as well: its low carry rises to the elevated strike and returns while the measured grip pivot keeps sword, right hand, shield and torso continuous in front and side key frames. It still has no DEF output or game installation. Candidate 02’s seven-frame ATTACK_FRONT now also passes: a high-ready resolves into the original forward horizontal strike and recovery. The default oblique and side views clearly show the blade travelling before the shield; the camera aligned with the sword axis naturally reads it as a thin line, while sword, right hand and body remain continuous. This remains Blender review only, with no DEF output or game installation. Candidate 02 now also passes an independent native eight-frame HOLDING review: restrained body breath and the measured grip-pivot sword preserve sword, hands, shield and armour in all front/side keys. No DEF output or game installation has been made.

Swordsman import and binding correction

The September 17 recoil review stretched the raised thigh into a narrow strip. A first attempt to redistribute leg weights did not fix it. The import audit exposed an earlier error: an Icosphere placeholder was included in the model bounds, and normalization transformed the meshes without transforming the armature. The rest render concealed the mismatch; bending the joints exposed it.

Reversing that recorded mesh transform reduces severely stretched edges at the same recoil peak from nine to zero, using a threshold of 0.08 model units and three times the rest length. The new leg keeps its volume. The shield still bends and remains unfinished. The shield has since been bound rigidly to the left hand and keeps its shape through all six recoil frames. Earlier attack scripts merely rotated the independent sword while the body stayed still; those acceptance claims remain withdrawn. New drafts use the shoulder, elbow, wrist and torso to drive the attack.

The new set contains 40 frames: eight holding, eleven defence, and seven attacks in each direction. They share a raised-sword guard, with the arm carrying the blade through the parry and swing. Reopening the saved scene preserves the selected shield surface and returns defence and attack endpoints to the holding pose. Leg lunges, full contact checks and the remaining groups are unfinished. The Swordsman is not installed in the game.

The current Swordsman test candidate has thirteen groups and 76 frames and is installed in the local Castle mod. The preceding nine-group, 58-frame version added lunges and walking. Each attack direction adds a foot lift, forward step and recovery. The six-frame recoil returns to the same guard. The eight-frame walk alternates the legs and swings the sword hand opposite its same-side foot, raising the blade as the arm moves forward and lowering it on the backswing. Move start and move end each have two frames.

After reopening, the lowest body point across all 58 frames stays within about 0.0016 model units of the ground, and the walk closure key reproduces the first-frame mesh. These checks do not establish the absence of foot sliding, stop-phase jumps or intersections in the game. Turns, hover and death remain unfinished, and the Swordsman is not installed. Both turns contain two frames and mouse hover contains eight. Death and installation are covered below; sliding and stops from arbitrary walk phases still need native review.

The turn rotates the body and equipment together while bringing the sword hand toward the chest and the blade upright. Mouse hover raises the sword, opens its angle, then returns to the common guard. Reopening and sampling all 70 frames gives a maximum selected-shield edge-length change of about 4.2×10⁻⁷ model units. This checks rigidity; it does not establish clearance between the shield and arms.

The turn and hover images above retain their earlier diagnostic camera. The current test package registers the camera to the original holding height, horizontal centre and ground position.

The six-frame death leans back, lifts the legs and turns toward the camera as the body lands. The first draft fell too far sideways and left the lowest sword and shield points about 0.099 and 0.142 model units above the floor. That version was not used.

The correction measures body support separately from equipment clearance and lowers the arm targets. The diagnostic floor below is excluded from exported game frames.

The collision check had its own failure: the initial weight threshold selected three hand vertices and no triangles, so its zero-intersection result was invalid. An expanded mask covers 2,862 arm vertices and 3,968 triangles. It exposed crossings in defence frames 4 and 8 and death frames 2–6. Adjusting the elbow direction while preserving the wrist and shield poses removes those crossings in a reopened, 76-frame check. Coplanar contact and body regions outside the mask remain outside its scope.

The local Castle mod is now version 0.2.0, adding Swordsman bodies, shadows and selection outlines at 1× and 2×. Existing Crusader files retain their hashes. All 76 frames fit the canvas, and format validation reports no errors or warnings. A subsequent native test battle reached its result and loaded 45 distinct 2× body frames across nine groups, plus shadows and outlines, for 106 distinct 2× resources. The observed groups were holding, mouse hover, movement, movement start/end, both turns, downward attack and death. This does not establish runtime coverage of all thirteen groups. Screen capture remained black, leaving visual acceptance and transition review open. Display settings were restored after the test.

Reproducing the client’s 100×130 panel and crop origin places the holding silhouette at x=51/100. None of the six sampled poses clips the body. This is an offline placement check using the original Castle background, not a native window capture.

The following revision 60 guard and attack images are retained as history; the attacks did not yet have leg lunges.

The import-tool fix transforms the mesh and armature together and supports explicit placeholder exclusions. The Blender regression covers parented and unparented skinned meshes in three poses, including save/reopen preservation. Models, artwork and complete mods remain private.

The following stills document the actual candidate rather than only its concept: mesh review, holding, the three attack directions, and the current walk key pose. They are Blender review renders, not DEF frames or an in-game installation. The following six images are historical records and no longer establish complete rig or attack acceptance.

The Swordsman’s native two-frame MOVE_START and MOVE_END reviews now explicitly bridge holding to the first verified gait pose and back rather than cutting the walk loop. Front and side frames keep sword, shield, hands, armour, tabard and boots continuous. They remain Blender reviews, without DEF assembly or game installation.

The first six-frame Halberdier front-attack probe tried to turn the imported loose pole pieces as one left-hand carrier. The visual check exposed unmoved head and lower-shaft fragments alongside the rotated shaft, so it is rejected. The accepted holding and walking states were restored; attack work now needs a clean full-weapon component extraction before another probe.

The second six-frame front-attack reconstruction replaced the pole with one rigid authored shaft, spear, axe head and butt cap. Its broad mask correctly removed the Meshy weapon fragments but also removed the holding hand; keeping the hand’s local region retained a vertical shaft remnant. Both sub-passes are rejected. The next attempt must extract a complete Meshy hand-and-weapon component set before action authoring.

The next six-frame ATTACK_FRONT review accepts the original Meshy long weapon rather than a substitute. All forward pole/head components are regrouped around the measured visible grip as one carrier; the lower cap previously misclassified as a left-foot piece is included as well. The full rest frame, contact and side key retain the single long halberd, its original axe silhouette, and the holding hand without residual static fragments. This is a Blender weapon-path review only; it does not claim DEF assembly or game installation.

The same complete-component carrier now passes both remaining native directions: six-frame ATTACK_UP rises through a diagonal high contact, and seven-frame ATTACK_DOWN completes its descending diagonal contact and recovery. Front and side keys show the original axe head, shaft, butt cap and visible grip as one coherent object in each direction.

A six-frame HITTED recoil now translates both the body and the complete original-weapon carrier together. Twelve-frame DEFENCE then raises the same original axe head and shaft through a diagonal chest guard and recovery. The selected front and side keys keep the grip, shaft, axe head and butt cap continuous; they remain Blender reviews without DEF assembly or installation.

The first four-frame Halberdier death probe shared one carrier between the body and complete original weapon, so it retained continuity, but it became an airborne sideways fall. The original CHALBD.DEF ends as a compact ground silhouette. This probe is rejected; a future death pass must author ground contact and a folded final pose rather than simply rotate the full character.

A second probe shrank the shared carrier toward the original compact endpoint, but it only made the same airborne horizontal fall smaller. It is also rejected; scale and whole-object rotation cannot substitute for a ground-contact death pose.

Meshy mesh

The first Meshy request used the complete four-view sheet as one input image. It spent 30 credits and returned several copies of the halberd without a person, so it was rejected before rigging or animation.

The second request used the single three-quarter panel. It also used 30 credits and returned a textured human mesh with the tabard, armour and halberd intact. Its source object has 404,705 vertices. Meshy’s automatic rig endpoint returned HTTP 400 for this asset and did not create a rig task or consume rigging credits.

Local rig and motion review

Astra authored the Blender-side recovery instead of retrying the automatic rig. The Meshy export is 972 disconnected mesh components, so conventional smooth weighting would pull its clothing and weapon apart. The local rig assigns each component rigidly to an anatomical bone; 51 vertically aligned halberd components are attached to the left-hand weapon bone. The rest pose remains intact, and the weapon stays connected while the arm moves.

The first holding loop has eight authored frames. The six-frame walking review keeps the halberd upright and gives the holding arm a restrained swing alongside the opposite arm and legs. It is deliberately a review render: no creature frames, shadows, overlays, DEF packing, or game installation have been produced from it yet.

The next pass is to inspect the walk against the original group frame by frame, then author the remaining CHALBD.DEF groups before any local mod is assembled. Meshy supplies the textured geometry; Astra supplies the local rig, motion, render and validation tooling.

The third four-frame Halberdier DEATH pass replaces the failed scale trick with a ground-contact pivot at the forward foot and local limb curl. In the battle-facing review it reaches a readable low, horizontal end silhouette while the entire original halberd stays connected to the left hand. This is accepted for the front sprite path only: the side view confirms that the current model is authored for the battle camera, not as a free-rotation character performance. It remains a Blender review asset, with no DEF packing or game installation claimed.

Royal Griffin: Meshy API bootstrap

The first two Royal Griffin text-to-3D previews were rejected. One became an upright bird; the other replaced the eagle head with a lion. Neither preserves the CRGRIF.DEF silhouette, so neither will enter Blender animation work.

The next bootstrap uses an enlarged holding frame from the original CRGRIF.DEF through Meshy’s Image-to-3D API. It restores the eagle head, folded wings and tawny lower body visible in the original. Its crown has been misread as tall horns, and the gold forward-limb ornament still needs local cleanup, so this is a review-stage mesh only—not a rig, DEF export, or in-game asset.

Four local crown-placement probes and one complete primitive head-replacement probe were rejected after front/side review. The source mesh does not expose the malformed horns as separable components; overlays either obscured the face, sat behind the wing, or left the artifact visible. Replacing the whole head with primitives created a visible neck seam and the wrong proportions. The next repair must reshape continuous local topology from the Image-to-3D mesh rather than stack cover geometry on top.

Cavalier: Image-to-3D static bootstrap

A holding frame from CCAVLR.DEF now seeds a Meshy Image-to-3D Cavalier. Front, side and oblique Blender reviews retain one mounted rider, brown horse, blue saddle cloth and a full red-and-white lance. This accepts the static bootstrap only: the horse-leg topology still needs separation and a dedicated quadruped rig before any holding, movement, or charge group can be reviewed.

Champion: corrected Image-to-3D static bootstrap

The first CCHAMP.DEF Image-to-3D candidate was rejected after side review because it added a second red-and-white spike from the horse’s forehead. The corrected second candidate has one lance only and keeps the gold-helmeted rider, horse armour and blue-and-white barding intact in front, side and oblique review. It is accepted as a static bootstrap; quadruped separation and rigging are still required before animation.

A crown-reduced Image-to-3D retry also failed: it reduced the artifact to one horn but enlarged the gold forelimbs and lost the original CRGRIF.DEF silhouette. Prompt-only retries are now closed; future work must use the original frames as constraints while rebuilding continuous local topology.

The first Cavalier quadruped-rig probe was also rejected. Heat weights assigned no vertices on the overlapping rider-and-horse mesh; deterministic nearest-bone weights covered every vertex but coupled a forehoof to tack geometry, breaking the side-view joint line during a small leg probe. A usable mounted rig needs explicit component separation before weighting.

The Cavalier local leg-replacement probe was also rejected: removing the old legs disrupted the belly connection, while procedural upper/lower legs and hooves broke the source mesh’s style and proportions. It cannot substitute for continuous mesh reconstruction.

Griffin: concept-guided Meshy bootstrap

The raw CGRIFF.DEF Image-to-3D attempt was rejected because it exaggerated the ear tufts and foreclaws and lost the resting wing pose. A clean side concept derived from the same frame gives Meshy an unambiguous anatomy reference. Its resulting mesh passes front, side and oblique static review with a continuous eagle head, folded wing pair, lion hindquarters and four limbs. The ear tufts remain longer than the original, so this accepts static bootstrap only; wing rigging remains future work.

Griffin: wing-root rig probe

The first six-frame wing probe was rejected because the wing root separated from the torso in oblique review. The revised grouping includes the high side feathers at the wing root, preserving continuity through a small six-frame fold cadence in both side and oblique review. This accepts a restrained idle-wing path only; flight and combat actions still need dedicated choreography.

Griffin: flight-specific Meshy bootstrap

The first six-frame flight probe reused the folded-wing mesh. At flight amplitude its oblique review opened a gap at the wing root, so it is rejected and will not be exported.

A new Meshy Image-to-3D request used a deliberately spread-wing, airborne Griffin reference instead of attempting to stretch the folded-wing asset. Its front, side and oblique Blender reviews retain two continuous wing surfaces, eagle forequarters, lion hindquarters and tucked airborne limbs. A conservative four-frame reciprocal wing cycle also remains continuous in side and oblique review. This accepts a flight-idle foundation only: it is not yet the native CGRIFF.DEF moving, attack, hit, defence or death set, and it has no DEF output or game installation.

Griffin: native four-frame moving review

The original CGRIFF.DEF moving group is airborne rather than a ground gait. Its four reference frames read as closed wings, a low sweep, closed wings, then a high sweep. The flight-specific Meshy asset now has a matching four-frame Blender review: conservative local wing weights carry the two wings through that cadence while a shared carrier supplies the small airborne lift. Side and oblique checks retain both wing roots, torso, eagle forequarters, lion hindquarters and tucked limbs at every key. This accepts the timing and pose basis for MOVING only; attack, hit, defence, death, move transitions, output layers, DEF packing and game installation are still outstanding.

Griffin: front claw attack review

The first nine-frame front-attack probe used only a shared dive carrier. Although its wing roots stayed intact, its foreclaws never made an independent reach, so it does not satisfy the original attack silhouette and is rejected.

The replacement uses separate left/right foreclaw regions alongside the flight-wing rig. Across the native nine-frame preparation, wing lift, forward/downward claw reach and recovery, the side and oblique contact reviews keep wing roots, chest feathers, both foreclaws, torso and airborne hind legs continuous. This accepts ATTACK_FRONT as a private Blender motion review only; the upward/downward strikes, hit, defence, death, transitions, DEF output and game installation remain outstanding.

Griffin: upward claw attack review

The first upward nine-frame pass did not lift its carrier or foreclaws far enough to read separately from flight idle, so it is rejected. The revised ATTACK_UP pass increases the upward carrier arc and local claw lift while preserving the Meshy flight mesh’s wing roots and feather surface. Side and oblique peak frames keep both wings, chest, foreclaws, lion hind legs and tail continuous. It is a private Blender review only; downward attack and the remaining groups are still unmade.

Griffin: downward claw attack review

The native CGRIFF.DEF downward attack resolves from a wing lift into a descending pounce and lowered claw contact. The nine-frame Blender pass now follows that arc with its own downward carrier trajectory while retaining local wings and foreclaws. Side and oblique contact frames preserve wing roots, chest feather surface, foreclaws, torso, lion hind legs and tail. This accepts ATTACK_DOWN as a private motion review; hit, defence, death, move transitions, export layers, DEF packing and game installation remain outstanding.

Griffin: hit recoil review

The first seven-frame hit probe displaced the whole body too little to read beyond holding, so it is rejected. The revised HITTED pass uses a larger rearward carrier recoil and recovery while keeping the local wings and foreclaws intact. Side and oblique peak frames retain wing roots, chest feathers, foreclaws, torso, lion hind legs and tail. Defence, death, transitions and export remain outstanding.

Griffin: defence review

The eight-frame DEFENCE pass follows the native wing-guard rise, held protection and recovery. It uses restrained local wing motion and tucked foreclaws instead of the attack carrier. Side and oblique peak frames retain the wing roots, chest feather surface, foreclaws, torso, lion hind legs and tail. Death, move transitions, export layers, DEF packing and game installation remain outstanding.

Griffin death constraint; Royal Griffin crown repair

The nine-frame Griffin death probe rolled the complete flight mesh toward the ground. Its rigid wing remained upright at the terminal frame instead of forming the original compact grounded corpse, so it is rejected and will not export. Death requires a local wing-fold and grounded-limb reconstruction.

Royal Griffin now has a new Meshy static bootstrap from an explicit three-point crown concept. The full-body side and oblique Blender reviews retain a small three-point gold crown band rather than horns, along with two folded wings, eagle forequarters, lion hindquarters and four limbs. Its bright front plumage still needs palette comparison against CRGRIF.DEF; no rig, animation, DEF output or game installation is claimed.

Royal Griffin: original-palette correction

A CRGRIF.DEF holding comparison confirms that the bright front plumage is faithful, while the first Meshy wings were too near-black. A textured Blender wing-region correction now shifts them to silver gray without replacing the feather detail. Side and oblique review retains the crown, front plumage, wing surfaces, lion hindquarters and limbs. This is the static baseline only; rigging and all action groups remain unmade.

Royal Griffin: independent idle wing review

The palette-corrected Royal Griffin now has its own eight-frame holding-wing review, based on the native CRGRIF.DEF holding count. It uses separate local wing-root groups rather than the standard Griffin’s weights. Side and oblique checks preserve the crown, wing roots, silver-gray texture, white front plumage, lion hindquarters and limbs throughout the restrained fold cadence. This accepts holding only; flight and all combat groups remain unmade.

Royal Griffin: flight mesh and four-frame moving review

A separate Meshy airborne Royal Griffin now supplies actual spread-wing geometry for the native flying MOVING group. The first palette pass recolored crown polygons; the next crown-safe region fixed that, but its first four-frame wing rig still admitted tail vertices, producing a pale tail band at the high-wing frame. Both are rejected. The revised local wing grouping excludes the tail. Its four-frame CRGRIF.DEF cadence—folded, low sweep, folded, high sweep—keeps crown, wing roots, tail, torso, lion hindquarters and airborne limbs intact in side and oblique review. This covers MOVING only, with no DEF output or game installation.

Royal Griffin: front claw attack review

The flight-specific Royal Griffin now has a separate nine-frame ATTACK_FRONT review. Crown-safe palette regions and tail-excluded wing groups are retained while left/right foreclaws make the forward/downward reach. Side and oblique contact checks preserve the crown, wing roots, tail, chest feathers, foreclaws, torso and airborne hind legs. Up/down attacks, hit, defence, death, transitions, DEF output and game installation remain outstanding.

Royal Griffin: hit recoil review

Royal Griffin now has its own seven-frame HITTED recoil and recovery. Its peak side and oblique frames preserve the crown, wing roots, tail, chest feather surface, foreclaws, torso and airborne hind legs. This remains a private Blender review; up/down attacks, defence, death, transitions, DEF output and game installation remain unmade.

Royal Griffin: defence review

Royal Griffin’s independent eight-frame DEFENCE uses a wing-guard rise, tucked foreclaws and recovery, rather than its attack carrier. Side and oblique peak frames retain crown, wing roots, tail, chest feathers, foreclaws, torso and airborne hind legs. Up/down attacks, death, transitions, DEF output and game installation remain outstanding.

Cavalier: split Meshy horse and rider bootstrap

The earlier one-piece Cavalier mesh interleaved rider, horse and tack, so automatic quadruped weighting, nearest-bone weights and procedural leg replacement all failed. It cannot be repaired by another whole-mesh leg pass.

The replacement begins with two Meshy assets: an empty-saddle Cavalier horse and a separate armoured rider with the red-white lance. The horse passes front, side and oblique static review with four readable legs, bridle, tail, blue-white barding and saddle. The rider preserves the lance grip, armour, tabard and limbs, then passes Meshy’s standard humanoid rigging task. The next step is a Blender riding pose and saddle interface; there is no mounted animation, DEF output or game installation yet.

Cavalier: mounted static assembly

The first two assembly passes are rejected: Meshy’s rigged GLB included an unbound Icosphere proxy, then an incorrect reparenting step broke the char1 → Armature skin hierarchy and enlarged the rider mesh. A later alignment pass left the rider above the saddle or facing opposite the horse.

The accepted assembly excludes the proxy, preserves the native humanoid skin parent, aligns the rider root to the saddle and horse forward axis, and folds the rider’s legs through its own rig. Side and oblique reviews retain the horse, four legs, bridle, blue barding, saddle, seated rider, red-white lance and coherent direction. This is the mounted static foundation only: quadruped rigging, saddle/stirrup constraints and all creature action groups are still unmade.

Cavalier: mounted eight-frame moving review

The horse-only Meshy mesh now passes an eight-frame quadruped walk: four conservative lower-leg regions alternate while saddle, tack, barding and torso remain on the root. The rider is then carried by the horse body root rather than placed in leg weights, with a small matching body rise. Side and oblique key frames retain four legs, bridle, saddle, blue barding, seated rider, red-white lance, skin hierarchy and direction. This accepts MOVING at the original eight-frame count only; attack, hit, defence, death, transitions, exact per-frame pose calibration, DEF output and game installation remain outstanding.

Cavalier: independent lance and front-thrust review

The first split rider still embedded its lance in the character mesh. An axis probe showed that the embedded lance was not weighted to RightHand, so it stayed still when the arm moved and could not support a charge. The corrected assembly uses an unarmed Meshy rider with Meshy’s humanoid armature and a separate Meshy lance. Blender samples the animated right-hand position into the independent weapon on every frame. The ten-frame ATTACK_FRONT review now raises, levels, thrusts and recovers the lance toward the horse’s forward direction without leaving a static duplicate behind. This accepts the weapon/rider motion foundation only: horse attack motion, native pose calibration, other groups, DEF output and game installation remain outstanding.

Champion: component-based mounted static bootstrap

The original one-piece Champion is useful as a visual reference, but its horse, rider, tack and weapon are one mixed mesh and cannot be used as a dependable animation base. A new component workflow supplies a dark Champion horse with four visible legs, blue-white striped barding, face armour and an empty saddle; a separate unarmed, gold-crested Champion rider then passes Meshy’s standard humanoid rig. The independent lance is reused through the proven weapon path. Side and oblique Blender review retain the complete mount, seated rider, crest and lance. This is a private static foundation only; quadruped gait, attacks, reactions, death, transitions, DEF output and game installation are still unmade.

Review correction: mounted alignment is not accepted yet

The two acceptance claims above were premature: Cavalier 12 faces away from the horse, Champion 05 sits sideways, and the lance origin was mistaken for its grip. Those claims are withdrawn; the old images remain as failure records. Cavalier 13 and Champion 06 align the riders with their horses and anchor the leather grip instead. The images below are work in progress. Champion legs still intersect the barding; stirrup contact, wrist orientation and the complete attack cadence remain unresolved. These are not finished animations or game integrations.

Champion: seated pose and leg weights (work in progress)

The rider now faces forward with both legs outside the horse. A proportion adjustment preserves the pelvis position; a two-segment leg solve positions the ankles and turns the boots forward, bringing the soles closer to the stirrups. Targets were located visually, so exact tread contact remains unverified.

The initial four-leg weights are ready for a complete walk. Inspection found tail vertices in a hind-leg group; narrowing the region then missed parts of a hoof. Nearest-hoof-center assignment and blended weights reduce vertices lacking full leg weights in the specified lower-hoof check region from 43 to zero. The tail check region shows no appreciable leg-driven deformation across eight frames. This is still a small-angle deformation probe; full stepping, hoof contact, original timing and game integration remain unfinished.

Champion: segmented legs and ground checks (not accepted yet)

The Blender rig over the Meshy horse now has segmented legs, a pastern segment and independent hoof orientation. Body bob and pitch carry the rider, saddle and lance; hoof targets compensate for those motions. Sampled support-hoof minimum heights stay close to the ground across eight frames. This verifies only those sample heights, not complete contact, anatomical shape or fidelity to the original animation.

Rejected experiments include seam welding with surface smoothing, which altered shading elsewhere, and texture-color-based root weights, which introduced local stretching even after weight smoothing. Those candidates remain failure records and do not replace the baseline. The additional segment reduces some distortion, but the lifted-leg silhouette below remains unnatural. Movement is still unaccepted, with no DEF export or game integration.