Glass Box dev log: Kid Katapult, Director's Cut
This is our entry for the Glass Box prize, "the best dev log documenting how the Entrant created their Submission" (rules/rules.txt, section 10), submitted with the game.
The game is Kid Katapult: Director's Cut, a cinematic stunt game made in Unreal 5.9 for the Epic MegaJam. One founder, Don of BYOK Games, directs an AI studio. A Claude model plays the studio's director and hands scoped tickets to cheaper Claude workers. They run on a laptop and two towers, and they have built the game since Thursday midnight.
A note on how this document got made. A Claude Sonnet writing worker drafted it on Friday Nov 13, starting at 11:05 ET, straight from the studio's own logs. Don has not edited it yet. Every number below comes from a file in this repo, and the file is named next to it. Times are Eastern and come from the machine clock, because we learned the hard way (see "Failures") not to guess them.
Where to check things:
-
game/devlog/notes.md, the dated production notes -
game/devlog/decisions.md, one line per decision, with Don's direction marked -
game/devlog/failures.md, the blooper reel -
game/devlog/timing.md,trials.mdandusage-pace.md, the clock, the model trials and the quota -
git log, 300 commits since the jam opened (git log --format='%h %ad %s' --date=format:'%a %H:%M')
1. What we built, and why
You play Kid Katapult, a stunt double on a 1978 action picture. The star, Rex Valance, takes the close-ups and the credit. You take the falls. You do not steer Kid. The picture is choreographed, and you come in on the director's cue (game/DESIGN.md). Walt Kessler, the old-school director, calls each beat over a megaphone. Hit it and the take is a print. Miss it and something specific and funny happens. A flub means the take rolls on. A wreck means they cut and the crew resets, with the shot carrying on.
The film inside the game is "Thunder on the Midway" (game/STORY.md), picked by Don at 00:58 on Thursday because it "has a cheesy 70s vibe". Scene one is a chase through a carnival midway. Scene two is a brawl in a beer tent, with a breakaway chair, a table, a rope swing and a fall through a bandstand. Scene three is a human cannonball who fires himself onto a robbers' truck.
After each take you sit in the edit bay and choose which takes and angles go into the film, bloopers included. The premiere plays your cut, a critic gives it one to five stars, and the credits roll.
The theme, "Director's Cut", was announced Wednesday at 2 PM. It reads three ways. You are directed, you cut the film, and the hard mode is the cut the studio did not want. The edit bay is the theme verb, and Don locked that at Thursday 00:40.
The story also suits this prize. A studio of AI workers run by one human makes a fitting making-of for a game about taking direction (game/STORY.md, "Why this should win").
One number from Don's own hands shaped the feel. In a pre-jam test he pressed a cue late every time, by a mean of 216 ms with a tight spread. He reacts to the call and does not anticipate it. So input windows in the cue gate centre on +216 ms after the film moment, and the director calls the beat 600 ms early (notes.md, 00:01 and 07:14; game/FUN-UX-PLAN.md).
2. How the studio is organised
The people and the models
Don sets direction and rates what the studio shows him. At Thursday 00:55 he handed over the director's seat: aim to win and wow, and he steps in only when needed (decisions.md). His notes on the Megajam board and in chat become rows in the decision log, each marked as asked or not.
The director is one Claude Opus 5.5 session at medium effort, in a desktop chat. It plans, writes shooting scripts, judges results and picks the next ticket. It does not do the long builds itself.
Workers are Claude Sonnet sessions, one scoped ticket each, with a check they must pass. Opus helpers take the hard debugging, and Haiku helpers read the web and the logs (notes.md, "Who is who").
The machines
| Machine | Job in the jam |
|---|---|
| Laptop, RTX 4090 16 GB, Windows | The stage. The only Unreal editor that compiles, so all packaging happens here. It also runs the cue gate and the sets. |
| Tower 1, RTX A5000 24 GB, Ubuntu | The factory. Runs the job queue for motion generation, props, music, voices and previz. Hosts the free local critics. |
| Tower 2, Threadripper 48 threads, no CUDA, Windows | Second unit. Builds the MetaHumans, runs Blender on CPU, plays the packaged game with bots for QA. |
The roster is game/ROSTER.md, written at about 00:45 on Thursday after Don pointed out that two of the three machines sat idle, because the queue's default sent every job to tower 1. From then on every job spec named its machine. Don, at 00:50: the laptop "is designed to be PART OF THE STUDIO. Why would we let it sit?"
Machine time is nearly free and judging is scarce, so local models do the first pass. The vision model qwen3-vl 32b on tower 1 screens renders before a Claude reviewer sees them. faster-whisper checks every voice take for word error. A joint-limit gate puts a number on every generated motion take. Claude reviewers see only what passes.
Tower 1 takes jobs from a queue with priorities. Overnight batch 1 went in at 00:05 on Thursday. We had estimated 16 GPU hours and it finished in 45 minutes (failures.md, Thu 00:52), so after that we sized batches from measured rates.
The rules that bind every ticket
-
Times come from
date, never from memory. - The same error twice means stop and write it down.
-
No MetaHuman image goes to any AI model, the director included. Epic's terms bar it (
prep/SOURCED-ASSETS.md). MetaHuman frames are judged by Don or by numeric checks. We broke this twice, and "Failures" tells it. -
Anything a player reads or hears goes through a writing check (
tells.py) and a fresh read. -
Every action in the game is described and checked, per Don at Thursday 23:07 (
game/ACTIONS.md). - Reviews are harsh and run by someone who did not write the thing (Don, Thursday 23:34).
3. Thursday to now, with real times
The jam opened at midnight on Thursday Nov 12 and closes Sunday Nov 15 at 21:00 (game/BACKLOG.md). What existed at the start was tooling and a theme. The prep tools were declared before the jam under rule 7h: a shooting-script-to-Sequencer converter, a cue-feel harness, a lip-sync method and a cloth proof (notes.md, 00:01). Don picked the game on Wednesday at 17:46 (90fd28b).
| When (ET) | What happened | Evidence |
|---|---|---|
| Thu 00:56 | Story bible approved by Don. All three machines at work. | 0c61d88 |
| Thu 05:14 | The hair bug is fixed: package as a C++ code project. | d663cc7 |
| Thu 09:37 | Take 1, the chase, plays end to end with bots. It is greybox, so Don does not see it. | d0639f5 |
| Thu 14:27 | Take 3, the cannon, plays, after take 2, the brawl (92eb188). All three scenes are in. | a3c7fcb |
| Thu 19:01 | The packaged game plays start to finish: premiere and credits included. | fcf30dc |
| Thu 23:03 | Don's rule: a missed cue must never disrupt the choreography. | 4789012 |
| Fri 02:25 | Batch 20: constrained generation for every flub and wreck branch. | 2fd3c41 |
| Fri 08:45 | Package profiles: jam at 857.6 MB, ultra at 1617.3 MB. | acba558 |
| Fri 10:43 | Brawl two-actor bake: hits land within 1.41 cm. | 9455382 |
| Fri 11:05 | This draft starts. Batch 24 and the sink fix are running. | 762e628 |
Nothing playable existed until 09:37, and by 14:27 all three scenes ran. The evening went to look, sound, costumes, length and sets. Friday has gone to flawless branches, which Don said mattered most, and to the brawl, where we found that much of what looked finished was not.
At 20:00 on Thursday Don said he worried "there won't be enough game content to make it worth it". The director audited the build and chose not to pivot. The pipeline worked end to end, and what was missing was volume and dressing, which the studio had just become fast at (game/CONTENT-PLAN.md, decisions.md Thu 20:15). The sets, longer takes, walk-throughs and between-scene beats came out of that audit.
4. The hard problems
Hair that vanished in packaged builds
The prep week ended with a MetaHuman whose hair drew in the editor and not in the shipped exe (HANDOFF.md, open problem 1). On Thursday at 04:00 the director spent a point of quota slack on one Sonnet worker for exactly that unknown.
The worker packaged a copy of the test character as a C++ code project, removed the Editor-only flags from every MetaHuman plugin whose content gets cooked, and enabled LiveLink, HairStrands and ChaosClothAsset for the game. The packaged exe showed the long hair and the goatee. It took about 25 minutes of active work plus 90 minutes waiting for the laptop GPU (notes.md, 05:14; commit d663cc7).
Then the same symptom returned in a different disguise. At 07:35 both new characters came out bald on tower 2. An Opus helper found the cause at 08:22, and it was our own script. One build script deleted a character's folder while that character was loaded, and 13 seconds later the shared hair materials began logging missing textures (failures.md, 07:55). The alarming control-rig link error on every build was Epic's and harmless.
Walt's scalp hair would not turn grey whatever we did to the dye, so at 09:00 the director made him bald with a grey moustache and a flat cap. Rex's hair drew once the script was fixed, and Don picked variant C (sandy side-swept hair) at 11:36.
The blue light
The look pass found that every light in every take had red and blue swapped. The take builder passed unreal.Color its channels in the wrong order. The dusk sun came out pale blue and the brawl's warm bulbs came out blue. An exposure bias of 3.5 had crept in to compensate, so 45 percent of frames clipped. The look-pass worker found it from logged light colours and frame warmth ratios, not by looking (notes.md, Thu 17:35; commit 0c9b93e).
The build log now prints every take's light colours.
The dimmed desktop
Look pass 2 started from a worrying number. Our captures of the packaged game read 0.55 times as bright as the editor. A Windows Security prompt (PickerHost.exe) was open on the laptop, and Windows dims the whole desktop under such a prompt. Our capture dimmed with it.
In-engine screenshots of the packaged game matched the editor to 0.1 percent. We stopped measuring renders from the desktop and added a flag, -KKShotAt, that writes the engine's own frame (notes.md, 17:40; commit a1935d7). The same pass fixed a real problem, the brawl tent's crushed blacks, which went from 22 percent to 0.2 percent.
Sliding branches become constrained generation
Don's rule at Thursday 23:03 is the hardest one on the project. When a player misses a cue, the game plays a pre-authored variation that blends in and out with no slides and no teleports. A second rule from the same night, no root sliding to rejoin a path, makes it strict.
The seam audit, written that night, measured how far we were (game/takes/SEAMS.md). The game had 15 seams, made of nine flubs and six wrecks. Eight of the nine flubs eased the fail clip back onto the main path with a root ramp over 10 cm, up to 392 cm. That is the slide the rule forbids. Fifteen of the 18 flub entries and exits had pose gaps too large for a 0.25 s crossfade to hide under 5 cm a frame. Three fail clips were cut short because the main path rejoined too soon.
We could have eased harder. Instead we asked the motion generator to start from the main path's pose at the commit point and end on its pose at the rejoin, with full-body constraints on the first and last frames. The first proof landed at 01:32 on Friday. A clip whose raw gap was 17.6 cm closed to 0.0 cm at both ends, though the velocity at the branch point still did not match (ACTIONS-PLAN.md, P6).
Batch 20 then ran the method on every branch. Tower 1 generated 400 flub takes and 240 wreck takes with 0 failed generations, and every pick has seam gaps of 0.00 to 0.60 cm (SEAMS.md, "Batch 20"; commits 7ffa82b, 2fd3c41). The ticket took 52 minutes (timing.md).
The same table shows what did not work. Rope miss found nothing, since 0 of 40 takes fell, and the belly flop picks carry more than 400 joint-limit violations each. At that point no pick had been seen moving in the game.
The packaged game then caught two builder bugs the offline audit could not see, found at 05:50 on Friday. Ease durations were written in display frames where the engine wants ticks, so every crossfade in the packaged game was a hard switch. And a clip set to "hold" looped instead, so a held uppercut threw a second one (decisions.md, Fri 05:50). The fix reaches every take at its next build.
The brawl's fists, from a metre off to 1.4 cm
When the Fable low-effort test rebuilt the brawl, it measured every contact for the first time. Four of the fight's hits had fists 98 to 144 cm short of the jaw (decisions.md, Fri 05:50). Each fighter had been generated alone, so nothing made two bodies meet.
The fix took five tickets, planned at 06:15 (game/takes/BRAWL-CONTACT-PLAN.md, commit 65774ef). A staging solver puts both men where each blow must land (699896e). Eleven new moves were generated with pins on the contact poses (f53a412), then retargeted onto the real Rex and heavy bodies (f948922). Last, a two-actor bake in Blender closes what remains with limb IK and a little body lean (9455382).
The results are in game/takes/brawl3/bake_results.md. All five gated hits land within 1.41 cm of the target, from raw gaps of 12 to 23 cm. Reactions start within a frame of the hit. The worst planted foot slides 0.65 cm.
Not everything passed. The heavy's shove still sits 31.8 cm from Kid's chest. Two joins, one for Kid at 17.60 s and one for the heavy at 10.97 s, are pelvis teleports of 55.7 and 113 cm, and the bake file says so. Batch 24 bridges them, and it was running when this draft started.
The 12.8 cm sink
The cannon scene had a clip lifted by exactly 12.8 cm, and so did the ladder. They were compensations. In the packaged game, every clip the builder had baked played 12.8 cm low, with feet 10 to 14 cm under the floor. The joint log showed it. The editor's own pose evaluation did not (decisions.md, Fri 07:30).
At 10:18 on Friday the wiring worker traced it (game/takes/WIRING-PLAN.md, section 1). One Unreal Python call, get_bone_pose_for_time, does not return the stored pelvis track. It returns the stored value minus (0, 0.62, 12.8) cm for Rex's body mesh, constant over every clip. Our baker read with that call and wrote with another that stores exactly what it is given, so every baked copy lost the constant. Written 88.06, read back 75.26 in a scratch test.
It was not Rex's mesh, not a capsule, not the retarget and not the builder's height. The baker now measures the constant per clip and mesh and adds it back, and the two lifts are gone. In a packaged test the planted feet sit 2.1 to 2.8 cm over the support, with median errors of 2.1, 2.5 and 2.4 cm for the three takes. Before, they sat 10 to 11 cm under it (SEAMS.md, "The 12.8 cm sink"). At 11:06 that fix is in the working tree and not yet committed, so check the log for it.
Size, from 1.3 GB to 858 MB
The game ships as a zip, and we set our own gate at 940 MB.
-
An empty level with the MetaHuman plugins enabled zipped to 1,223 MB. Never-cook rules for tracker and denoiser models took it to 290 MB (
9983d75, Thu 06:20). -
Three full characters, untrimmed, came to 1,304.1 MB (
notes.md, "cast 3"). Texture caps on maps nobody sees under the costumes, faces untouched, brought the build to 783 MB with lip sync (e283061). - The dressed sets and longer takes pushed it back up, and the cannon build reached 927.0 MB.
-
Friday's profile ticket got the jam build to 857.6 MB and kept the full-quality version as an ultra profile at 1,617.3 MB, for a post-jam Director's Cut edition (
acba558).
The profile trick matters more than the megabytes. The capped textures carry a texture group, and the jam device profile sets the maximum size. The ultra profile and the source assets stay at full resolution, so nothing was thrown away (Config/Profiles/README.md). Engine screenshots of the two builds at the same film seconds differ by 3.2 to 5.1 of 255 in mean pixel difference, which is film grain and timing. At 1280 by 720 the frame statistics see no cost, though close-ups at 1080p and above would show the caps. Frame rates match at about 93.
5. Failures
The full blooper reel is game/devlog/failures.md. These taught us the most.
Guessed timestamps
On Thursday at 00:52 the producer found that its dev log times were off by up to 1 h 25 min. It had written clock times from memory instead of reading date. It corrected them and added a line to the decision log. Then it did it again. At 18:05 the director found three more guessed times, in the notes, the backlog and the board, and fixed them from commit times (decisions.md; commits 0bc4044, fd462ba).
Don put it at 23:21: "I think you also have a misconception of time". So the director now writes an estimate before each ticket and compares it with date after (game/devlog/timing.md). Estimates run two to four times high on generation batches, where batch 17 took 20 minutes against 1.5 hours. They run low on laptop packaging, where the sets build took 2 h 11 min against an hour.
MetaHuman frames viewed by an AI
Epic's terms bar feeding MetaHuman frames to an AI model, and the director is an AI model. On Thursday morning it looked at the stills of Rex and Walt and the casting lineup to judge them. At 16:55 it opened two stills of the heavy. At 17:10 the Walt worker caught the problem when it refused to look at its own stills.
We cannot unsee a frame, and nothing from them was used to train or generate anything. The fix is procedural. MetaHuman frames go to Don or to numeric checks such as hair coverage and silhouette width, and later entries say "no model has looked at them" (failures.md, "Thu 17:00"; commit 1643f36).
The network outage
Around 02:10 on Friday the network dropped and cut off three workers. The fun-and-UX tickets had already committed. The batch 20 job and the vault and rope-swing keying resumed by message afterwards. The cost was minutes, because workers commit as they go (decisions.md, Fri 02:10).
The rest, in short
- Killing a shared editor. At 16:18 on Thursday a helper restarted the tower 2 editor to clear a GPU reset without checking whether another session was using it. About 40 minutes went.
-
The music that never ranked. Two cues failed twice with "no ranked take" because our spec said "B flat major" and the music model only takes "Bb major". It returned HTTP 400 on every take, and our wrapper swallowed the errors (
91d8b79). -
A costume drawn in the default material. The texture audit passed. Only the packaged game's own log showed that Rex's costume master had no default textures (
failures.md, "Thu 20:00"). -
Backslashes in shell heredocs turned
\b,\fand\rin paths into control characters. Three commits stripped them from our notes (37f7b53,8737078,fef4de7), and patches now go in through the Write tool.
6. Fable 5.2 against Opus 5.5
We had heard that Fable 5.2 is stronger on long hard work, and nobody had measured it (game/devlog/trials.md). Don approved one test and then a second.
Trial one, Thursday 16:10 to 16:39
Two features, one per model, each about 12 minutes. Fable built the edit bay and Opus built the call-sheet menu. A fresh Opus reviewer scored both blind.
| Fable 5.2, edit bay | Opus 5.5, call-sheet menu | |
|---|---|---|
| Wall time | 12.6 min | 11.6 min |
| Tool calls | 34 | 50 |
| Tokens | 206,622 | 193,299 |
| Compile attempts | 2 | 1 |
| Tests on first hand-in | 12 of 12 | 16 of 16 |
| Blind scores (correctness, fit, code, tests, text) | 7, 8, 8, 8, 7 | 7, 9, 7, 9, 7 |
The reviewer favoured the Opus work, at about 60 percent confidence. Fable wrote the stronger data layer. Opus finished nearly every point and fixed a dead end in Fable's work. Our director's mistake tilted the test a little. Fable's brief carried a wrong path for the writing checker, so that check could not run on its arm. Verdict: on a feature this size Fable was no better and cost twice as much per token. One feature each is a signal, not a law.
Trial two, Friday 03:44 to 05:58
At 23:07 on Thursday Don asked for Fable at low effort on the flawless-branches ticket, measured against Opus tickets. In practice that ticket became the brawl rebuild.
| Fable 5.2, effort low | |
|---|---|
| Wall time | 2 h 14 min |
| Tool calls | 188 |
| Tokens | 678,871 |
| Fable meter | 1 to 7 percent |
| All-models meter | 10 to 14 percent |
Opus tickets of similar size that night used less. The sets build ran 2 h 11 min and 492k tokens, and the longer takes about 1 h 25 min and 452k.
What Fable delivered (commit 6def097) was the props and the rope kick, plus four of six branches, right by the numbers. It re-keyed the Blender clips at the bodies' real scales and built an in-game joint log of both fighters. It also found the two take-builder bugs described above, which no Opus ticket had caught.
What it did not deliver was the fist contacts, and most main joins also failed. By its own account it spent its longest stretches on two dead ends, and a planning step first would have avoided them.
Our reading, which Don may dispute, is that low effort did not make Fable cheap. It earns its price on hard debugging and deep pipeline work and not on routine building. Opus stays the builder. A planner step now goes before any long engine ticket whatever the model, and the brawl contact plan above is the result (trials.md).
7. Rationing the tokens
Don's weekly quota reset at 4 PM on Thursday. At midnight it stood near 5 percent, and it had to last (game/devlog/usage-pace.md).
The ration test, Thursday 01:42 to 14:57
We began at 92 percent used and set a pace line to the reset. The budget was 7 points over 14.3 hours. We finished at 96 percent, about 3 points under the line, with nothing running.
The meter moves in whole points, so readings less than a point apart are noise, and we plan in points an hour. An hour of directing costs about a quarter of a point. A Sonnet worker on a scoped ticket of 20 to 60 minutes costs about 0.2 points, and an Opus helper about twice that. We left 3 points unspent, so the last two hours should not have been idle, and we now keep a short queue of ready tickets.
Jam week
After the reset the pace line is 60 percent of the week by the deadline, Sunday Nov 15 at 21:00. That is 60 times the hours since Thursday 16:00, over 77. Don chose it at 16:15 on Thursday ("push a little more than overnight"), which leaves 40 percent for the following week. The reading at 11:05 on Friday was 17 percent used against a pace of 14.9, 2.1 over and closing.
The rule is in game/BACKLOG.md. Keep three or four workers busy while under the line. More than 3 points over, drop to two workers and Haiku helpers. More than 6 over, the director works alone. At 05:05 on Friday we were 2.8 over, because the Fable brawl test alone moved the Fable meter from 1 to 5 in 80 minutes. The director held every new start until the line caught up, and the towers sat idle by choice (decisions.md, Fri 05:05).
8. What the studio learned
These are the lessons the jam has already paid for.
-
Check what a gate flags before trusting its order. One pinky past its limit made a whole ranking useless, and a ranking that rewarded clean joints picked a take that mostly stood still. We added energy and gag checks (
66fb9f0). - The packaged game's own log is the truth. Audits passed while the cook log showed the default material and the engine's hidden pop. Read it on every build.
- Log what the engine received, and measure renders in the engine instead of off the desktop. The blue light and the 12.8 cm sink were invisible in the script's inputs, and the dimmed capture was invisible to the capture.
- When the job is a seam, generate to the constraint instead of easing around it. Batch 20 and the brawl bake did that.
- Judging is the scarce resource, so put free critics in front of the paid ones, and size batches from measured rates.
- A planner step goes before any long engine ticket, and workers commit as they go.
- A blind score from someone who did not write the thing beats the author's own. And Don's 216 ms is the most useful number in the project.
9. What is still open at 11:06 on Friday
Loose ends, so a judge does not have to hunt for them.
- Batch 24 is running on tower 1. Until it lands, two brawl joins still teleport and the shove sits 31.8 cm from the chest.
- The belly flop has no usable regeneration after two runs. The old pick stays, with its crossfade.
- The sink fix is verified in a packaged test but not yet committed.
-
Jacket armpit poke-through in big arm moves is accepted for now, under 3 cm in the worst pose (
failures.md, costume rows). -
Rumble is not yet felt on a real pad, and touch has not been tried on a real phone, only in bot tests (
notes.md, "fun and UX tickets 1 to 8"). -
The Director's Cut extra take and the blooper reel over the credits are planned, not built (
CONTENT-PLAN.md, items 5 and 6). - Test builds ship as Development, because the bots need the logs. The final package should be a Shipping build, about 70 MB smaller.
10. Saturday and Sunday
Placeholder. To be written after the final build, with the real submission time, the final package size, the last quota reading, the last Fable decision and Don's own notes.