Open the books

RAW FILE: notes.md, as written during the jam

Source game/devlog/notes.md in the jam repo. Last git commit of this file: 2026-11-15. About 29,786 words. Scrubbed only for private details and one line about another entrant. Mistakes left in on purpose.

Production notes: Kid Katapult: Director's Cut

Raw notes for the making-of, newest at the bottom, and every number names its source. The finished dev log is cut from these, the decision log (decisions.md), the failure log (failures.md), assets.jsonl and git.

Format, so the log reads like a production: each day opens with a call sheet (who is on, what we shoot), entries in between are dailies, and every wreck goes in the blooper reel (failures.md).

Who is who (roles to sessions, for splitting token burn later)

RoleSessionModel
Producer and jam build ownerClaude Code desktop chat, started Thu 00:01 ET, folder NEWBOYOpus 5.5, medium
Research only (no building)prep chat "Studio setup for game jam"Opus
Studio hub (cloth, garments, machines)"Game dev studio laptop setup"Opus
Rules readerHaiku scout helper, Thu 00:04Haiku 5.5

Call sheet, Day 1: Thu Nov 12

Budget: Don's weekly quota is about 5% until it resets Thu 4:00 PM ET. Until then the crew is the machines: tower 1's queue generates for free, the producer spends tokens only on reading and small writes, helpers are Haiku.

Before 4 PM: rules check, dev log open, overnight generation running, one page design doc with every open question for Don. After 4 PM: the hair bug in packaged builds, then one real take end to end (shooting script, cue gate, one character, packaged) so Don can feel a real one.

00:01 Prep wraps, production starts

The jam opened at midnight ET. What existed at that moment was tooling and decisions, nothing in the game:

What Don's own hands told us in prep, the most useful data we have: on the cue-feel harness he pressed late every time, mean +216 ms, with a tight spread. He reacts to the call rather than anticipating it. So input windows open after the cue, and the director calls a beat early. On the sampler he gave it 3.5 stars and said it was hard to judge without "a more real game". That sets Thursday night's goal.

00:05 Overnight batch 1 queued on tower 1

Four jobs went to tower 1's queue at low priority (queue seq 558 to 561): music (5 cues, ACE-Step 1.5, 3 takes each), voice casting reels (3 director and 3 star voices, 4 lines each), 10 set props (image to 3D via TRELLIS.2) and 6 tileable set textures. The specs were written in prep at 11:45 PM (commit 2255d06) as prompts only; nothing was generated before midnight. These are candidates for casting, not final assets.

Small wreck on the way: the queue refused the first submit because the handoff commit had not been pushed, and the towers fetch from origin. Resubmitted at commit 36ef39d, which holds the same specs. (failures.md)

00:38 Research wave 1

Don asked for tool research while the quota is low: X, Reddit and the web for motion, film look, audio and new tools, plus the freebies the studio already holds. Four Haiku researchers went out in parallel, one per topic, each writing to research/<topic>.md with a license source for every claim and a local test we could run free on tower 1. The cheap tier reads the web; the director reads only their summaries.

00:55 All three machines on the job (times from the clock from here on)

Don, 00:50: the laptop "is designed to be PART OF THE STUDIO. Why would we let it sit?" He's right. Overnight batch 1 finished in 45 minutes on tower 1, not the 16 hours its specs guessed: all 27 assets OK (5 music cues, 6 voice reels, 10 props, 6 textures), plus stunt motion batch 1 with 69 of 69 Kimodo takes at about 3.2 s each. Machine time is nearly free; judging is the scarce thing. So:

The queue sends jobs to tower 1 unless told otherwise. Making it spill to an idle laptop by default isn't safe yet, because tower job specs use Linux paths. For now each job names its machine, and the studio fix waits for after the jam.

01:18 Title locked, first dailies, a tool from X

01:19 Motion previews and an hourly loop

03:06 Hourly check 1

04:00 Hourly check 2

04:59 Hourly check 3

05:14 The hair bug is dead

The bug that ate the prep night (hair, brows and moustache gone in the packaged build) is fixed by packaging on the laptop as a C++ code project. One Sonnet worker did it on a copy of the test character: a game module and targets, the Editor-only flags off every MetaHuman plugin whose content gets cooked, and LiveLink, HairStrands and ChaosClothAsset enabled for the game. The packaged exe shows the long hair, moustache and goatee. The director checked the frames; the exposure is blown out, but the strands are there. About 25 minutes of active work, plus 90 minutes waiting for the laptop GPU. Two failures on the way are in the blooper reel. Still open: cloth on a real garment, and size (this test package is 1.7 GB because the copy holds several builds of the character; the cap is 950 MB).

06:03 Hourly check 4

06:07 The jam project exists

A Sonnet worker built game/KidKatapult (commit 2df8b0e): a C++ code project on 5.9 with the MetaHuman packaging setup, MCP settings, the two gate fixes and an empty L_Stage7. It compiles, packages in 110 s and runs. The catch: an empty level zips to 1,223 MB, against our 950 MB guard (itch's web upload; the rules set no size cap). The game-enabled MetaHuman plugins are the suspect. A third Sonnet worker is measuring and trimming it now (cook only referenced content, never-cook unused plugin folders, a Shipping exe), with the rule that hair must still cook.

06:20 Package trimmed to 290 MB

The trim worker measured the cooked content: 556 MB of MetaHuman Animator tracker models, 319 MB of MetaHumanCoreTech tracking models and 103 MB of denoiser models, none of them used by the game. Never-cook rules for those folders, maps-only cooking, and turning off the NNEDenoiser plugin took the empty-level zip from 1,223 MB to 290 MB (Development). Shipping is 221 MB but loses the capture hooks, so Development stays. Commit 9983d75. Hair still to be re-checked in the next packaged character build.

06:58 Hourly check 5: the build starts early

07:14 The cue gate works

The game's core system is in C++ (commit db60632): it loads a take's cues, opens a window per cue centred at +216 ms after the call (Don's measured reaction), grades Perfect, Good, Scrappy, Early, Late, Wrong and None, routes flubs (the take rolls on) and wrecks (jump to a reset section, rejoin in the same shot), and turns the grades into a take score and a 1 to 5 star rating. Window widths per mode: Rehearsal Perfect plus or minus 70 ms, The Take plus or minus 55, Director's Cut plus or minus 40. 7 of 7 automation tests pass headless. Next: a real take's cues and Level Sequence, a pawn carrying the gate, the call text and marker UI, rumble.

07:36 Casting, take 1: no print

The Sonnet casting worker built Rex and Walt on tower 2, and both came out bald. Every build failed on the same link error in MetaHuman's shared face control rig, and the scalp hair and beard materials lost their textures. Rex was also wrong for the part: slim and young, not a rugged 38. The director looked at the stills (the back views were blown out white) and agreed. Nothing goes to Don. Suspected cause: the test project already holds the prep character's shared Common assets, and a new build tries to rewire them. Hard debugging goes to an Opus helper: a fresh project, a casting lineup of every male preset face so the director picks Rex's base by eye, and one proof character with hair.

07:58 Hourly check 6

08:22 Casting, take 2

The Opus helper found the hair bug in the build script, not in MetaHuman: one script deleted a character's build folder while it was loaded, which broke the shared hair materials 13 seconds later. The scary control-rig link error is harmless and happens on every build, including the prep character that had hair. Hair still doesn't draw in its SceneCapture stills, while the packaged game drew hair this morning, so the next stills come from a game camera. It rendered all 14 male preset faces as a lineup. The director cast Rex on Victor and Walt on Walter; the lineup is on the board so Don can overrule.

08:30 Motion lands on the MetaHuman skeleton

A Sonnet worker built the retarget (commit 7f7014a): Blender turns Kimodo BVH into FBX, then Unreal's IK Retargeter maps SOMA onto the MetaHuman skeleton, scripted headless, a few seconds per batch. The joint-limit gate before and after: no new elbow, knee, hip or head problems on six takes, and the upper-arm and thigh twist survived (the lesson from the fight lab). The wrists drift (up to 25 degrees over in one chase take): next fix is a hand pose offset. The director checked the brawl comparison: the retarget tracks the source in front and side views. But that top-ranked brawl take is too tame; it mostly stands. The ranking rewards clean joints, not action, so the video critic has to score energy too.

09:00 Rex has hair; the ranking learns about energy

09:40 Take 1 (midway chase) assembled and played by bots

09:37 Take 1 plays end to end

A Sonnet worker assembled the midway chase in the real project (commit d0639f5): Rex (hair draws in the packaged game), 11 Kimodo clips retargeted onto his body, the Level Sequence built from chase.take.json by the prep director tools, the cue gate on a controller, a HUD with the director's call, a shrinking ring and stars, keyboard and pad bindings, and a bot flag that plays perfect, late or not at all. Packaged at 641 MB. Bots: perfect gets 5 PERFECT, 5 stars, a print; late gets 5 LATE at +301 to +306 ms, both wrecks fire, 1 star; none gets 1 star. The director looked at the late run: the mechanics read clearly, and the look is greybox: placeholder boxes for props, blue-washed light, the default undershirt. Not for Don's eyes yet. Fix list for a take Don would enjoy: the flub reactions need their own sequence sections (now they show only on the HUD); real props and the money man; blends between clips (they pop); light and the 1978 grade; and the windows may be too hard (the bot tool says a human 80 ms off passes first try 2.8% of the time).

09:58 Hourly check 8

10:25 Take 1 feel pass

10:19 Costume, take 1: no print

A Sonnet worker modelled Rex's 1978 costume from code (Geometry Script, sized from his joints, ambientCG CC0 leather and denim). It came out crude: a tube jacket with gaps at the shoulder seams, a blocky shirt, wide straight jeans over the boots. It also doesn't follow animation, because the built body mesh is neutral and Rex's real shape only exists at runtime. Nothing goes to Don. The method was wrong: garments come from cloth draped on the real posed body. Costume v2 after the 4 PM reset (director designs it): export Rex's posed body from a runtime frame, drape pattern garments in Blender cloth sim, transfer the body's skin weights to the garments, import on the MetaHuman skeleton, jacket hem as Chaos Cloth by the CLOTH.md method. Also from X (Don sent it): another MegaJam entrant has a film-producer character made with Tripo (hosted; commercial rights need a paid plan, so not for us). A film set alone won't stand out under "Director's Cut"; ours has to win on the edit bay, the directed fails, real stunt motion and the making-of log.

10:19 Take 1 feel, round 1

Commit e33af24, 9 of 9 tests pass. Flubs are real now: each one jumps to a parallel branch of the take where Rex blends into the fail clip and back. Same shot, no reset, and Walt says "Keep rolling, keep rolling!" Windows tuned against a simulated Don (+216 ms, sd 45): he'd print about 72% of the time in Rehearsal, 46% in The Take and 20% in Director's Cut; a sloppier player (sd 70) gets 22, 9 and 3. The morning's "2.8% pass" figure was the bot tool's old zero-centred model; with Don's real bias the old windows printed nearly every time. Blends failed: the sequencer doesn't blend overlapping clips the way the build script assumed, and the packaged log shows a 4 m root jump at 9.0 s. Next: align each clip's root to where the previous clip ended (the sequencer's match-to-previous root offset, or bake the offsets into the clips), then re-measure the per-frame pelvis delta.

10:58 Hourly check 9

11:50 Take 1: root alignment across clip boundaries

Rex's pelvis no longer pops at the clip handovers. The packaged log showed the sequencer does not blend overlapping animation sections linearly (the pelvis follows the outgoing clip through the overlap, then switches), so the fix does not lean on the blend. Each clip use is baked (prep/scripts/director/director_align.py) so its pelvis starts where the clip before it is at the handover (XY and yaw), sits on the outgoing clip's pelvis through the 0.25 s overlap, then the gap eases out at about 3 cm per frame. The root bone is identity and the motion lives in the pelvis, so the sequencer's root-motion offsets (option a) have nothing to act on; baking (option b) is deterministic and checkable offline. The fail clips in the flub branches are baked onto the vault at the branch point and eased onto the main take's next clip, so the rejoin gap is under 1 cm. Packaged pelvis log, per-frame pop at every boundary: under 4.6 cm in main and both branches (was 56 to 495 cm). Costs: the fail clip slides up to 15 cm per frame mid-fall in the vault flub (the take is 750 cm ahead by the rejoin), and split_b starts as a standing pose 160 cm up, so Rex stands through the first 1.6 s of the slide beat; that clip needs a re-cut or a regenerated start. Wreck resets still jump (they are cuts). Videos: D:\megajam\take1\v3.

11:26 Rex stops teleporting

Commit b1c8e55. Root offsets are now baked into aligned copies of the clips (/Game/Motion/KimodoAligned), because Rex's motion lives in the pelvis, not the root bone, so the sequencer's root offsets had nothing to move. Pop at each clip boundary, measured in the packaged game, before and after: 74 to 4.5, 139 to 2.0, 495 to 0.4, 215 to 3.3, 304 to 1.0 and 266 to 0.8 cm a frame. Both flub branches enter and rejoin under 4 cm. Still to fix: the slide clip (split_b) starts standing with the pelvis high, so Rex stands for 1.6 s where he should slide (re-cut or regenerate the start); the vault flub's fall drifts up to 15 cm a frame catching up with the take; the wreck resets are cuts by design.

11:28 Rex variants

Three variants at MEDIUM quality (commit 37c94fd): A is Victor with a short casual cut and a moustache, B is Orlando with long dark hair, C is Bruce with sandy side-swept hair (Lorenzo read about 60, so the worker swapped him out). The director's pick is C: Kid is seen from behind for the whole game, and C has the most recognisable silhouette. On the board as q16, with the sheet. Stubble grooms didn't draw on C, and MEDIUM builds render the hair as cards.

11:50 Rex is cast

Rex final (commit eedadad): the Bruce face, Hair_M_SideSweptFringe in sandy blond with darker roots, light stubble (Beard_S_Stubble drew on the first try), HIGH quality, strand hair in Movie Render Queue. The 360 body check passes. 741 MB on tower 2, so the package budget needs watching. The face and hair shots went to Don; the body waits for the costume.

11:58 Hourly check 10

12:11 Retargeter A/B: no swap

Don's X find (toyxyz Blender retargeter, GPL tool) against our UE IK retarget, on 4 takes (commit 416f41c, game/motion/RETARGET-AB.md). Foot slide per contact: challenger lower on all four (1.14 against 1.53, 0.82 against 1.56, 2.33 against 3.72, 1.55 against 2.73 cm), closer to the source. Root drift: challenger better on the long chase runs, ours better on the short takes. The joint-limit gate tied 9 to 8, and the challenger needed a hand-tuned reference pose to get there. Wrist twist tied. Verdict: keep ours. A 1 cm gain isn't worth a second toolchain mid-jam. The challenger is the proven fallback if foot slide shows on the long runs once Rex is lit.

12:17 Take 1: slide clip and vault flub drift

Both motion defects from the 11:26 entry are fixed in the packaged game (D:\megajam\take1\v4). No better slide take exists: of the 90 split_b takes in takes-5 and -6 plus the 13 gate_slide ones, none starts running and reaches a slide, and only s53 reaches the floor at all (its pelvis floats at 2.7 m for 2.2 s first). So split_b is trimmed instead: the take json gives it trim 1.4 s, rate 0.85 and z_w 0.6, so the slide starts on the c2_slide cue. The pelvis drops under 60 cm at 9.91 s (was 11.12 s), the sequencer reads the clip from 1.4 s in, and the trimmed 3.6 s clip fills 9.0 to 13.24 s at 0.85 (a section longer than its clip restarts the clip instead of holding it; logged). The vault flub fail clip now travels along the take's own smoothed path instead of sliding to catch up: the rejoin ramp is 22 cm, not 754. Per-frame step in the fall averages 7.9 cm against the take's 8.0 at the same film time; the part that is not the take's pace is 4.3 cm a frame at most when smoothed over 0.3 s (was 10.6; single frames reach 10 because the take's own gait bob is not copied). Pop at every boundary stays under 4.6 cm. Mirror flub unchanged.

12:16 Take 1 motion fixes, and a gag lost

Commit a7a284f. The slide now drops on the cue: no slide take in 103 starts running, so the head of the best one is trimmed and played 15% slower. The vault flub follows the take's path (rejoin gap 22 cm, was 754 cm), and every boundary is under 4.5 cm. But the flub now reads as Rex jogging on with flailing arms, not a tumble, so the gag is gone. The director won't keep that: fails must be funny. Batch 9 asks Kimodo for fails that travel with the take (tumble over the cart and roll back up running, bounce off the mirror and spin on, and the brawl fails) at 40 seeds each.

12:31 Costume v2: the rig works, the look doesn't yet

The Opus helper found why the first costume couldn't fit: Rex's built body mesh has no torso, neck or thighs (MetaHuman cuts them under the default outfit). export_geometry gives his whole body on the same 342 bones. Each garment starts as a patch of that body, eased out, with pattern details added: lapels, collar, back yoke, hem band, cuffs, pockets, belt, boots. Weights come from the nearest body face. In UE it replaces the default tee and shorts on his outfit component and follows his animation; the shoulders hold in a sprint frame (commit 28751e2, 50k triangles, 3 MB). Not shown to Don: the jeans poke through the jacket hem, the leather reads as glossy oxblood plastic, the collars are ragged, the flares split. A Sonnet polish pass is on it.

12:48 Voice bake-off

Six local engines on tower 1 (commit 194f37d): Chatterbox, Kokoro, Dia, Qwen3-TTS VoiceDesign (installed today with its weights), Maya1 and VoxCPM2. Three designs per character each, 2 takes a line, every take scored with Whisper word error and audiobox-aesthetics. 36 reels. Numbers favour Kokoro and Chatterbox because they hold one voice across lines; the voice-design engines drift between lines but may suit a gruff 1978 director better. Dia was weakest (its voice prompt got glued onto every take). Pitch checks slip an octave on gravelly voices, so that term is capped. Ten finalists are on the board (q18), including two per character picked by the director for character rather than score. Default if Don doesn't pick: Chatterbox for both. Maya1 output needs the "Built with Llama" credit.

12:54 Costume v3, and the next push

Costume v3 (commit 3871852): nothing pokes through in a sprint, the leather is matte, and the collar and hems are clean. The director looked at the sheet. From behind at game distance it reads 1978 at once: long leather coat, open collar, flares. But the jacket looks inflated (parka, not leather), the sleeves balloon, the denim is a garish royal blue, and the shoulder close-up is jagged. Since Kid wears it all game, an Opus helper takes pass 4 on cut and drape, and renders a gameplay-distance shot, the view players actually see. Meanwhile a Sonnet worker assembles take 2, the brawl, on the laptop with the take 1 pipeline. Meter 95% (pace 97.4).

13:25 Take 2 (beer tent brawl) assembled

Same pipeline as take 1. Twelve clips from tower 1 (brawl beats a/b/c s17/s1/s7, chair swing s15, haymaker_take s13, the early-sell, short-slide and chair-whiff fails, mark check, dust off, bow) went BVH to FBX to Rex's skeleton, in /Game/Motion/Kimodo. Props imported from the laptop GLBs (table 240 cm, chair 95 cm). The sequence and cue table come from game/takes/brawl.take.json through director_build.py and director_cues.py; -KKTake=brawl starts it, take 1 stays the default (the game removes the other take's actors at start, since all takes share L_Stage7). New in the builder: a second travelling actor ('followers', the heavy, a second Rex at 1.08 with dark clothes, root-baked like the kid and copied into the flub branches and resets), props attached to a hand socket with scale keys (the chair), a per-flub minimum play rate, and a per-clip 'travel' scale. Clip starts were set from the joint data so duck, uppercut, jaw hit, shove and chair smash sit near the 112 bpm beats. Packaged pelvis log: every kid clip boundary under 4.6 cm a frame (the beat_a handover is the worst, it is the clip's own speed); branch entries and rejoins under 2.6 cm; wreck resets are cuts. Bots: perfect 5 stars, late 1 star (both wrecks and both flubs fire), human 3 stars. Videos in D:\megajam\take2. Open: no clip has the heavy throw the haymaker, so the jaw hit on b3 has nobody hitting; b2 is HUD only (no branch); Rex_final is not in this project, Rex_v2 plays.

13:27 Take 2 plays

The beer tent brawl, assembled by the take 1 pipeline (commit 92eb188): 12 clips retargeted, the table and the breakaway chair imported from our own props, 2 flub branches, 2 wreck resets, -KKTake=brawl to start it. Bots: perfect gets 5 stars; late fires every fail, 1 star; human timing gets a print and 3 stars. Pop under 4.5 cm a frame. The uppercut's head snap lands within a frame of its beat. The miss: the heavy never throws his haymaker (no such clip existed), so Rex sells a punch nobody threw. Batch 11 makes the heavy's moves. He's a tinted Rex for now; a real heavy MetaHuman is next on tower 2.

13:30 Costume v4 (C8): tailored, not inflated

Rebuilt the jacket fit in game/costume/bl_20_build.py: real ease (1.5 cm chest, shoulders and sleeves, 3.5 cm at the waist, 2.5 mm leather), a straight fall from the chest with an A-line that just clears the hips, folds shaped into the mesh (back, underarm, elbow, wrist), seams as clean grooves on straight cuts, a clean collar band with a rolled edge, and a wear and occlusion mask in the vertex colours for M_CostumeV4. Denim is period indigo on a twill normal, snug to the knee. Stills in D:\megajam\cast\costume_v2\r4 (sheet.jpg, game\game.0000.png for the gameplay view). Open: a tear at the back of the right armpit in the sprint close-up, a hairline at the neck base, slightly stepped shirt collar edges.

13:31 Costume pass 4: it's a costume

The Opus helper's pass 4 (commit 35dacc2, version C8): ease cut to 1.5 cm at the chest, shoulders and sleeves, 2.5 mm leather, soft drape folds built into the mesh, a clean rolled collar and yoke seam, darker tan-brown leather with edge wear, indigo flares snug to the knee. The director looked at the turnaround and the gameplay-distance shot: it reads as a fitted 1978 leather jacket, not a parka, and Kid is readable from behind at 34% of frame height. Shown to Don. Still open: a small tear at the back of the right armpit mid-sprint (close-ups only), a faint neck hairline, stepped shirt-collar points.

13:50 Takes 1 and 2: Rex final, traveling fails, beat fix, size check

13:57 Both takes star the final Rex

Commit 7bc7bdc. Rex_final in the project (every clip retargeted to his own skeleton, about 5% taller than v2), Rex_v2 deleted, zip 649 MB. The worker caught a bad pick by the director: vault_tumble_roll_on-s35 never tumbles (the head stays above the hips the whole clip), so it ranked all 40 seeds and used s33, which flips head over heels and rolls on. The ranking still can't see whether a gag happened. Next ranking term: for a tumble, require the head to go below the hips. The uppercut's head snap is now placed on the 6.21 s beat. A new lift_cm key puts Rex on the table top in the brawl slide. The pop is under 3.5 cm a frame everywhere except the mirror rejoin (4.4 to 5.0) and one frame of 6.4 at the walk handover.

13:58 Hourly check 13

14:24 Costume C9c: flares fixed, boots not yet

Commit f9a212d. The flares are clean bells now (loft from the knee, turned hem, stitch groove), and they break over the boot and stay clear in the foot strike. The director looked at the foot close-ups: the boots still read as soft brown blobs, and a dark notch shows on the right outer shin. A boot is a hard object, which our image-to-3D is good at, so batch 12 generates three 1970s boot styles on tower 1, and the next costume pass fits the best one to the foot.

14:26 The cast has voices

Commit c27c994: 42 final lines in the voices Don cast (Walt 27, Rex 9, the AD 6), 180 takes rendered and scored, max word error 0.0, normalized to -18 LUFS, 8.2 MB. Maya1 drifts between lines (speaker similarity median 0.45 for Walt, 0.48 for Rex, 0.27 for the AD), so a consistency pass with Chatterbox reference-following may be worth it later. One-word lines ("Jump!") give noisy similarity scores. The WER check first counted "ten" against "10" as an error, now mapped.

14:25 Take 3 (the cannon) assembled

Same pipeline as takes 1 and 2, plus the first ballistic flight. Six clips from mj-kimodo-cannon-10 (wave s39, climb_in s17, launch_flight s11, tuck_land s20, fails belly_flop s25 and land_slip_hang s36) went BVH to FBX to Rex_final, and take 2's bow and dust-off close the shot. The take is game/takes/cannon.take.json (made by make_cannon_take.py, 23 s, four cues on the countdown music at 84 bpm: IN! hold, FIRE! tap on Walt's "fire", TUCK! hold, LAND! tap). The cannon is the redo GLB at 1.33 (about 4 m barrel); the truck is box placeholders and a blue crash pad on the bed. Kid stands on a 2 m platform at the muzzle, sinks into the barrel, and is fired 23 m onto the bed.

14:27 All three takes play

Take 3, the cannon (commit a3c7fcb): wave, climb in, countdown, FIRE, a ballistic arc (apex about 6.5 m over 23 m), tuck, land on the truck bed, bow. Four cues; a belly-flop flub that rolls on and a slip-and-hang wreck. Bots: perfect 5 stars, late fires all four fails, human 3 stars. Zip 653 MB. The worker caught its own bug: building any take that declares lights deletes the level's shared sun and sky light, so take 1 may now render dark. The flight reads as a floaty upright leap, not a projectile. Fix: pitch the actor's whole body to horizontal along the arc in the sequence, rather than waiting on a better clip.

14:57 Hourly check 14, the last before the reset

Meter 96% (pace 98.5). The last hour's two workers: the laptop fixes the takes (the build stops deleting shared lights, the cannon flight goes horizontal, all three takes switch to the cast voices); tower 2 fits the best of three generated 1970s boots (klein then TRELLIS.2) to Rex's feet, rigid on the foot bone with the toe split at the ball. Don rated the voices 4 stars and may record the lines himself for the performance; Chatterbox voice conversion (MIT) would carry his timing into the cast voices.

15:11 Takes v6: shared lights kept, cannon flight horizontal, final voices

The builder no longer deletes the level's untagged lights (a take's own lights carry DIR_<id>; only those are replaced). Cannon: the actor pitches to the arc tangent (60 to 100 deg of lean, in over 0.15 s after the muzzle, out 10.0 to 10.4 s), the flight clip plays at rate 0.2 from 1.5 s to hold its arms-forward pose, camera about 7 m behind and 3 m to the side on a 40 mm lens. Voices: all three takes use the final Maya1 files; walt_l12short.wav (cut from walt_l12) is the flub reaction; lines checked for overlap in each take. Package 653.3 MB. Bots: perfect 5 stars on all three, late on take 3 fires all four fails. Videos in D:\megajam\takes_v6\{chase,brawl,cannon}. Take 1 is lit but dim and blue (the sun at 8 lux barely warms it); the flight camera looks down steeply and the floor is black; both are next passes.

15:28 Boots, good enough at game distance

Commit 6077353. Of three generated boots, only the Chelsea boot came out clean (one was a shattered shell, one a low crepe shoe). Rex's foot measured 31 cm, not the 28 to 29 the brief guessed. The boot is split at the ball and mirrored, and the shin notch is fixed. The flares hide all but a brown toe cap, which is blobby from decimation. The director's call: fine at gameplay distance, polish later. Light and grade matter more.

Thu Nov 12 16:07, ration test closed

The weekly meter reset at 4 PM (now 0% all models, 0% Fable). Result of the test: we started at 92% at 01:42 and finished at 96% by 14:57, with no reading above the pace line all day. Budget was 7 points over 14.3 hours; we used about 4.

The Fable 5.2 trial (game/trials/fable-editbay.md) starts now with fresh quota.

Thu Nov 12 17:00, the heavy is cast

Heavy_v1 (preset Jorge, buzz cut, salt-and-pepper stubble, 192 cm) built on tower 2 and staged on the laptop (D:\megajam\heavycopy, import step in game/casting/HEAVY_IMPORT.txt). Director's look: the face is right, a tired, rugged midway heavy. The body reads soft rather than dangerous and not much bigger than Rex on screen, and the stills are in the default tee and shorts, so nothing goes to Don until he's costumed (1970s work jacket or denim vest over a tight tee, work boots) and maybe bulked through the shoulders. Size risk: Rex_final is 899 MB in the project and the heavy adds about 650 MB raw; check the cooked zip against our 950 MB guard after import. Also today: the Fable vs Opus trial (devlog/trials.md), the reviewer's nine fixes (d8d742b), gag checks in the ranker (66fb9f0). Look pass running on the laptop.

Thu Nov 12 17:00, Walt is rebuilt

Walt_v3 (Walter preset, 178 cm, paunch via chest 102 / waist 102 / hip 99, balding-stubble scalp hair, short full moustache, no beard, all strands set to grey) built on tower 2 with the heavy's scripts copied as game/casting/wt_00 to wt_06 and staged on the laptop (D:\waltcopy, import step in game/casting/WALT_IMPORT.txt). Six stills at 1280x720 are in D:\megajam\review\cast\walt\. I did not look at them (MetaHuman frames never go to an AI model), so the hair shape and grey are for the director to judge; if the scalp still reads dark, the old Walt_v2 bug (strand colour not responding) is back and hairDye plus params are the first place to look. The cap stays a prop.

Thu Nov 12 17:05, hourly check

Meter 2% against a pace of 0.8: the first hour carried the trial and four workers. Holding at two workers. No new feedback on the board; the heavy and Walt face sheets wait for Don.

Thu Nov 12 17:35, the blue light was a colour-order bug

Look pass 1 (0c9b93e): the take builder handed red, green and blue to unreal.Color in the wrong order, so every light in every take had red and blue swapped. The dusk sun came out pale blue, the brawl bulbs blue. Fixed, and the build log now prints each take's light colours. Also: warm dusk on the chase with 22 bulb practicals, the cannon camera down from 6.4 m to about 2.5 m on a 21 mm lens with the truck and the muzzle in one frame (checked by projecting the truck's bounds, not by eye), and the 1978 grade (grade-film-1978.json) on the take cameras only. The grade survives packaging (warmth ratios match within 0.01). Open: the packaged game is half as bright as the editor, the brawl tent crushes a fifth of the frame to black, and Stage 7 has no light of its own before a take. Look pass 2 (Opus) is on those now. Every number here came from frame statistics; no one looked at a MetaHuman frame.

Thu Nov 12 17:40, motion redo batch 13

Eight of nine moves that had no passing gag take now have one (gate_slide 9/40, kungfu_flurry 22, flub_late_jump 40, dust_off 40, brawl_beat_a 4, trapeze_catch_full 40, mirror_dizzy 22, truck_hang_on 4). The fix was plain: say the man starts standing on the floor, spell out the speed. cannon_launch_flight still stands in place in every take; Kimodo won't fly a man, so the cannon keeps its keyed ballistic arc with the flattened flight pose (0851fb7). Rankings with gag columns are written back to all six batch folders on tower 1 (e95e07f).

Thu Nov 12 17:40, the dark package was a dimmed desktop

Look pass 2 (a1935d7): the packaged game draws exactly as bright as the editor (in-engine shots match to 0.1%). Our desktop captures read 0.55x because a Windows Security prompt (PickerHost.exe) was open on the laptop, and Windows dims the whole desktop under it. Lesson: measure renders with in-engine screenshots (-KKShotAt), never desktop capture alone. Also fixed: the brawl tent (crushed blacks 22% to 0.2% with three warm back bulbs and haze), a neutral Stage 7 work-light rig that builders never delete with each take's lights switching on only during its take, film grain that now measurably reads, and the hot placeholder gate. Package 653 MB. The prompt needs Don; we don't touch security dialogs.

Thu Nov 12, cast 3: Heavy_v1 and Walt_v3 are in the game

Thu Nov 12 17:59, the heavy's costume (HC1)

The heavy is dressed for 1978: a near-black rib-knit crew tee, an open olive duck chore coat (corduroy collar, steel buttons, two chest and two lower patch pockets), brown duck work trousers (straight leg, double knees, a belt), and Rex's generated boot, refitted to his feet and graded to a darker, scuffed oxblood. It was built in Blender on tower 2 the same way as Rex's costume, as patches of his real body offset, relaxed, cut and skinned. Every height comes from his own joints. Scripts: game/costume/hc_00 to hc_07 and bl_50_heavy.py; the build log with all the numbers is D:\megajam\cast\heavy_costume\hc1n\build_log.json on tower 2.

Thu Nov 12 18:22, the whole game plays end to end

Premiere and credits (1eb43fa, Opus worker): after scene 3 the film opens on its title card and plays the player's cut, one reel per scene, replaying each section's recorded branch and angle from the take's own sequence, with no input judged. Bloopers play as comic beats, Rex's close-up is cut in where chosen, the audience reacts between reels, a critic gives stars, then the credits (cast, BYOK Games, every tool and model, thanks). A bot played the whole loop in the editor: call sheet, three takes, three cuts, premiere (9 of 9 sections played as planned), credits, back to the sheet. 25 of 25 tests. Director's calls on credits: the heavy's actor is "Lou Brannigan"; Claude and Whisper stay in the credits (Glass Box: we credit the tools that made it). Open: no audience or fanfare audio yet (SFX batch on tower 1 now adds them), the main title and end credits music cues, about 115 s against a 90 s target, the blooper reel over the credits. Next: the heavy's costume in, the close-up's facing checked by numbers, and a packaged run of the whole loop.

Thu Nov 12 18:45, Walt's costume (WC1)

Walt_v3 is dressed as a 1978 director on set: a pale blue shirt open at the neck over a white crew undershirt, a rust wide-wale corduroy field jacket worn open (dark brown cord collar, four flapped patch pockets, horn buttons), pressed taupe-brown slacks with a front and back crease and a slight boot cut, a dark belt with a brass buckle, Rex's generated boot regraded to sand suede desert boots with crepe soles, and a brown-grey herringbone tweed flat cap. The cap is its own rigid prop: a static mesh in the head bone's frame, attached to the Body component's "head" socket with no offset. Built in Blender on tower 2 with the heavy's HC1 method (patches of his real body, offset, relaxed, cut, skinned). Scripts: game/costume/wc_00 to wc_07, bl_60_walt.py, bl_61_diag.py, wc_run.ps1, wc_pack.ps1. Build log with all numbers: D:\megajam\cast\walt_costume\wc1b\build_log.json on tower 2.

Thu Nov 12 19:01, the packaged game plays start to finish

b226773: the heavy wears his costume in the game (a new "outfit" field in the take builder, after a Blueprint edit wouldn't stick), and the premiere's Rex close-up now faces his face, measured against the head bone (dot -0.93 at the start, -0.75 at the end). The packaged build (868 MB zip) ran the whole loop in about 5 minutes: call sheet, three takes, three cuts, the premiere with all 9 sections played as planned, the critic at 4 stars, the credits, back to the sheet. Content/KKCast (2.2 GB of characters) lives only in the laptop's working copy, so it's now tarred to tower 1 ~/studio-ai/backups/megajam. Next on the laptop: the sound effects and title music wired into the takes and the premiere.

Thu Nov 12 19:20, costume fit fix FX1 (all three costumes)

One fit pass shared by the cast's costumes (C10c on Rex_final, HC1 on the heavy, WC1 on Walt), run on the finished FBX files on tower 2 so all three share one check and one set of fixes. Scripts: game/costume/bl_70_fit.py (the check), bl_71_fitfix.py (driver), bl_72_fixes.py (the fixes), bl_73 to bl_75 (boot diagnosis, garment-only look renders, round-trip check), fx1_05 to fx1_10 and fx1_pack.ps1 (Unreal and staging). Runs and logs: D:\megajam\cast\fit1 on tower 2 (fx1_rex, fx1_heavy, fx1_walt; the A/B runs beside them).

Thu Nov 12, lip sync for the cast's voice lines

Thu Nov 12 19:30, the game has sound

511c2fc: one sound map (game/audio/sound_map.json) feeds the takes, Stage 7 and the premiere. Each take has its ambience loop, footsteps, beat and fail sounds, its own sounds on flub branches and wreck resets, and cheers only on clean sections. Stage 7 hums between takes, the clapperboard snaps at each take, Walt's megaphone squelches before his calls. The premiere opens on a murmur and a projector clunk, the main title plays under the card, the projector runs low under the reels, the audience laughs at the bloopers you cut in, and the credits get their own theme. Levels: voice 1.0, music 0.55, effects 0.6, crowd 0.45, ambience 0.16. Packaged loop ran clean, 897 MB zip. Nobody has listened yet: the sounds are on the board for Don's ears. Also tonight: Walt's director moves (batch 14), the costume fit pass (boots fixed on all three; upper-arm poke-through accepted for now), lip sync starting on tower 2.

Thu Nov 12 19:55, costume FX1 and Walt's moves in the game

etarget\work\walt14.log). Per take: idle_director loop (period 6.967 s), clap_once with the hands meeting on the word "action" (chase after Rolling!, brawl in the line, cannon on "fire!"), shout_cut one second before each "Cut!" line, facepalm at 0.3 s into each wreck reset, applaud_slow in a new sequence-only applause section at 100 s (kid bows or dusts off, crowd cheer). KKPlayerController::HandleTakeFinished jumps there on 5 stars, plays 5.2 s, then pauses as before. The bot saw it on all three scenes. Not placed: arms_folded_watch, point_action. Clips cut hard into each other (a non-travel actor has no blend), so a small pose pop at the handovers is likely; nobody has looked.

Thu Nov 12 20:00, everyone dressed; a near miss on Rex

5eb7c18: Rex, the heavy and Walt wear their fixed FX1 costumes in every take, Walt has his flat cap, and Walt finally acts like a director: an idle loop, a clap on "action", a shout before each "Cut!", a facepalm on every wreck, and slow applause when a take ends at 5 stars. The near miss: the packaged log showed Rex's costume material failing to compile in the cook, so his whole costume was drawing the engine's default material. A texture audit had passed; only the cooked log caught it. Fixed with default textures on the master. Lesson: grep the packaged game's own log every build, not just the editor's. Size: 955 MB, near the jam's 1 GB rule with lip sync still to add. Next ticket trims to under 900 MB with a gate in the package script, and wires in lip sync.

Thu Nov 12 20:22, 783 MB with lip sync

e283061, 440bc24: the package dropped from 955 MB to 783 MB by capping body and default-clothing textures nobody sees under the costumes (faces untouched), and pkg_smoke.ps1 now fails any build over 940 MB. Lip sync is in: 21 face tracks across the three takes, 40 ms behind each line, with baked eye darts and blinks under every gap so Walt and Rex never stare. Speakers fixed (Walt's lines no longer say "kid"). The packaged loop ran with 0 errors. 217 MB of headroom is the budget for the sets and longer takes. Next on the laptop: the takes grow to about 50 s each (CONTENT-PLAN section 2).

Thu Nov 12 21:48, takes at 45 s; a full run is 8 min 47

3dfe3f1: every take now has five sections. Chase adds the mirror maze and the leap back onto the truck; brawl adds the rope swing and the uppercut into the bandstand fall; cannon adds a run-in up the platform, a wave to the crowd and a high five with Walt before the launch. 8, 8 and 6 cues. Bots: perfect 5 stars on all three, late and no-input 1 star with every branch firing, no errors. The packaged full loop took 527 s (8 min 47 s) with 15 of 15 sections played as planned. Known: flub branches ease Kid back onto the main path by up to 3.7 m (likely visible sliding), a few new calls are text only (six Walt calls being voiced now with the walk-throughs), the rope is a stand-in. Next on the laptop: the four dressed sets with the batch 15 props.

Fri Nov 13 01:47, fun and UX tickets 1 to 8 (laptop, Sonnet worker; started 00:12)

Commits 58298c3 (the gate: tickets 3 to 7), ca6d634 (the HUD, touch bot, UI scale, menu input), 11a185b (the three takes rebuilt after the story fixes). Packaged into D:\megajam\builds\fun1 (zip 850.7 MB, gate 940 MB), the final package at HEAD of 11a185b.

Fri 03:36, crowd faces (three extra faces, tower 2)

Extra_A (Sook-ja preset, a woman, pixie), Extra_B (Omari, curly fade and horseshoe moustache) and Extra_C (Cameron, clean cut) built at quality LOW with the heavy and Walt scripts as game/casting/ex_00 to ex_12 (ex_run.sh runs one in the tower 2 editor over ssh and MCP; ex_vars.py holds the three casts). Real start 03:05, end about 03:40; builds took 1 to 2 min each, no slow-task waits, editor PID 11508 untouched and left on L_HeavyLineup.

Fri 08:45, package profiles jam and ultra (laptop, Sonnet)

Real start 07:41, end 08:46. pkg_smoke.ps1 -Profile jam|ultra (Config/Profiles/README.md has the how-to).

Fri 12:05: hourly

Ticket 7 stopped at 11:28 on a real blocker: the bake ran on bodies 17% too small, so the hits miss by 20 to 32 cm in the game. The rework exported all 22 clips on the real bodies and re-baked on tower 2; it is waiting for the laptop to run the in-game gate. Fun tickets 10 and 15 (hit feedback, barks, streak, subtitles) hold the laptop. Don said he was blind to progress; the board now leads with the newest post, and log-0031 carries chase and brawl video and stills from the packaged build (checked by numbers only: brightness, motion, about 28 unique frames a second). Usage 18%, line 15.6.

Fri 12:27, fun tickets 10 and 15: hit feedback, barks, streak, subtitles (laptop, Sonnet worker; 11:41 to 12:27)

Fri 13:05: hourly

The brawl rebake on the real bodies passed in the game: every hit within 0.6 cm (b440c9f). Ticket 7b builds and packages it. Fun tickets 10 and 15 landed in 46 min (670c0ce). At 12:50 Don proposed pivoting to an on-rails food body-horror shooter. My answer: yes, if it's tied to the theme (SPOILED: The Director's Cut, a banned 1979 creature feature, the track as the camera dolly, forks as cuts), kept Teen, with a tight scope, and with Kid Katapult frozen as a submittable fallback and a go/no-go gate on Saturday at 09:00. Waiting on his call.

Fri 13:25, fun14: three Walt barks, call word without its subtitle, pause (ticket 14) (laptop and tower 1, Sonnet worker; 12:29 to 13:25)

Fri 14:05: hourly

Pivot hour one. Design plan done (spoiled/DESIGN.md, 80d277f): an 8.5 minute run in four reels, ending with The Wurst; replay by a 30 Hz state recorder; five questions on the board. Running: concept art (queued behind voice on tower 1), The Wurst proof (EEVEE renders on tower 2), audio, the prototype, and a bake-off of Claude-built creatures against FLUX plus TRELLIS for the methods doc. Tech scout found nothing to try. Usage 20%, line 17.2: above the line, but Don lifted the cap for the pivot; holding at five and adding none until the first results land.

Fri 15:06: hourly

Don reviewed the first 22 option cards. Creatures read too normal (Sir Loin like a known game demon), The Wurst should be a formless swarm, the voices need to be monstrous and better made. An art director and an audio pass are working on it, and the 3D Wurst is being rebuilt around the swarm. The bake-off verdict is in METHODS: FLUX for concepts, Claude in Blender for rigs, a hybrid for hero creatures.

Fri 15:47: the SPOILED prototype plays

Plain-geometry prototype done in 2 h 32 min (estimate 6 h 30): reels 0 to 4 as data, three weapons with freeze, shatter and burn, six enemy types, Sir Loin, The Wurst with mutating crawlers, the fork both ways, cover rules, shootables, 30 Hz run recording. Bot clears both routes in about 7.5 minutes with no continues; no input dies at 0:21; 162 fps; 362 MB package. A fresh lead engineer takes round 2: feel, the swarm Wurst Don described, track v2. Kid Katapult stays frozen as the fallback.

Fri 16:06: hourly

Don's second round of notes: the jerky dog is out, the mites are a maybe, The Wurst has to look properly formed with no hole for a mouth, hyperreal is the look, the audio goes Hybrid. The biggest note is on the prototype: the action only happened at the stops. Round 2 puts enemies on the moving ride too, with cinematic beats between.

Fri 17:05: hourly

Don rated round 2 audio within minutes. Kept: the Maya1 PA, Sir Loin's Maw roar, the Hybrid mix. Cut: Pru's voice; the player is silent. Redo: The Wurst (too fast, too scratchy) and the score (too happy). Short audio round 3 on just those two.

Fri 18:05: hourly

Don picked Sir Loin as the beef centipede on the meat rail, liked the factory frames, left the key art to me, picked the darkest fight loop, and wants The Wurst offended: "Cut me?!" The prototype's second round plays with action on the move; round 3 puts the first real look on it.

Fri 19:05: hourly

Don kept all three weapons at 4 stars and asked for 5; a polish pass runs on tower 2. The first in-engine look (round 3) is on its last bot runs; its clip gets a review before Don sees gameplay.

Fri 20:05: hourly

Don pushed on intensity, scares, the full factory and movement beyond the cart, then asked for a backstory, an ending, and how we judge fun without him. Answer in flight: a director's pass, rigged creatures, a factory kit, round 4 movement modes and a scare system, a story writer, and a measurable craft yardstick. Round 3's in-engine look was held back (sphere creatures). No new board ratings this hour.

Fri 21:05: hourly

Don locked ending B, kept the A6 loop and the round 4 Wurst voice, gave the factory kit 4 stars and handed us the fifth enemy: the Hanger, a carcass that wakes when shot. The plan grew weapon reactions and one-take camera rules. Light-gun-ready aim mode goes in the build; the phone-as-gun is a post-jam build. Wurst finals land about 22:40.

Fri 22:05: hourly

Quiet hour on the board. The additive got its three names (Compound 7, label L-7, "the wiggle"), the Compound 7 sheen and veins became every creature's signature, and the beef sides and dairy blob moved to the image-to-3D route. Tower 1 has an overnight line: Wurst finals, scare sounds, story lines, beef and blob images, creep sounds.

Fri 23:07: hourly

The Wurst swarm finals landed and went to Don: reveal, rise, forms and the leaper. Sonniss libraries are downloading for layering. Creatures: Sir Loin has menace now; the spider and Hanger wait on better meat and casings. Tower 1 runs scare sounds, story lines, creep sounds and the beef and blob image jobs overnight.

Sat 00:05: hourly

Late Friday changed the story's shape: no movie framing, a MetaHuman partner (Lenny, re-dressed from Kid Katapult) who briefs us, gets dragged into The Wurst, and is saved when it falls. Don's Wurst notes: a slithering hooded brute that throws volleys of franks, not a column or a wave. Hands move to Unreal's first-person template arms. Five audio cards and the Wurst finals are on the board.

Sat 01:05: hourly

Don set the bar before sleeping: a plant that makes sense to the last detail, VR-grade presence, 4K60, Teen-safe. Written into STANDARDS. Real beef carcasses and real-size franks pass; the brute shell failed (the Wurst will be 25k to 40k instanced franks). Round 4 has ten movement modes and sixteen scares. Lenny's voice is recorded; his MetaHuman, the area layouts and the Hanger rebuild run on tower 2.

Sat 02:05 hourly

For the 08:02 morning post (added Sat 02:09)

Sat 03:05 hourly

Sat 04:05 hourly

Sat 05:05 hourly

Sat 06:04 hourly

Sat 07:05 hourly

Sat 08:10 morning post

Sat 08:11 hourly

Sat 09:08 hourly

Sat 10:05 hourly

Sat 11:05 hourly

Sat 12:05 hourly

Sat 13:05 hourly

Sat 14:05 hourly, and state for the director after any compaction

Sat 15:05 hourly

Sat Nov 14, 18:52 ET (director, after a compaction)

Sat Nov 14, 19:05 ET (hourly check)

Sat Nov 14, 20:05 ET (hourly check)

Sat Nov 14, 21:05 ET (director, hourly)

Sat Nov 14, 22:05 ET (director, hourly)

Sat Nov 14, 23:05 ET (director, hourly)

Sun Nov 15, 00:05 ET (director, hourly)

Sun Nov 15, 01:05 ET (director, hourly)

Sun Nov 15, 02:05 ET (director, hourly)

Sun Nov 15, 03:05 ET (director, hourly)

Sun Nov 15, 04:05 ET (director, hourly)

Sun Nov 15, 05:05 ET (director, hourly)

Sun Nov 15, 06:05 ET (director, hourly)

Sun Nov 15, 07:05 ET (director, hourly)

Sun Nov 15, 08:05 ET (director, hourly)

Sun Nov 15, 09:05 ET (director, hourly)

Sun Nov 15, 10:05 ET (director, hourly)

Sun Nov 15, 11:05 ET (director, hourly)

Sun Nov 15, 12:05 ET (director, hourly)

Sun Nov 15, 13:05 ET (director, hourly)

Sun Nov 15, 14:05 ET (director, hourly)

Sun Nov 15, 15:06 ET (director, hourly)

Sun Nov 15, 16:05 ET (director, hourly)

Sun Nov 15, 17:04 ET (director, hourly)

Sun Nov 15, 18:04 ET (director, hourly)

Sun Nov 15, 19:04 ET (director, hourly)

Sun Nov 15, 20:05 ET (director, hourly, last of the jam)

Sun Nov 15, 21:05 ET (director, final)