Four days.
One way home.

Preparation began with Unreal Engine 5.9 MegaJam Preview, Codex integration and editor connections. The repository and initial project established the workspace that both of us would build on.

The history spans two calendars: the initial Unreal project and design-document commit is dated November 12 in China, which was still November 11 in Eastern Canada. We retain that distinction here; these are recorded milestones, not a reconstruction of when every individual asset was made.

Tian Sang’s browser level editor and five authored prototype levels gave the idea a practical shape. Props, paths and timing could be laid out in a scene and exported into the native game.

The starting puzzle was deliberately small: a falling rock, a fixed performer, and a bad take. Removing a threat could make the scene safe; reversing it could preserve an optional reward. A second puzzle asked a flying charge to solve the rock problem through timing.

Chai Yuntian’s six-scene C++ prototype brought removal, reversal and timing changes into Unreal. The team kept the early scope focused on a single-player keyboard-and-mouse puzzle.

Success had to mean more than watching a promising preview. The player would explicitly call Action, complete the performance, and then see the same scene as a cinematic replay.

Layered miniature backdrops dressed the six-scene prototype. At the same time, contact corrections and recorded outcome comparisons checked that presentation changes did not quietly alter the player’s take.

The visible surface, the physical support and the recorded motion had to agree. That problem would return throughout the jam in feet, rails, rope anchors and moving platforms.

Tian Sang’s tutorial and audio work landed at 18:33 Beijing time. Chai’s integration followed at 08:50 Eastern / 21:50 Beijing. The handoff combined editing interactions, saved cuts and sound with the miniature scenery and lighting.

Preservation was explicit: authored physical items and catalog entries kept their gameplay fields. Even a squashed bomb silhouette was repaired without changing its physical box. Shared chapter work was a conversation between mechanics and presentation, not a clean division down the middle.

Square wooden controls, a hinged Action clapper and larger timeline tracks turned ordinary interface elements into miniature production tools. The scene and the edit needed to feel like parts of the same object.

Frozen story and spatial chapters were integrated with the workshop, then checked in Windows at three interface resolutions. Readability, framing and how much stage remained visible became practical design constraints.

Stale previews and scene framing received attention alongside a lit mine set, local mist and handmade brambles. The visual world grew richer, but the player still needed to read the route and recognize what could go wrong.

Layered scenery and local effects became ways to explain depth, distance and danger. A scene could not depend on a dramatic camera angle that hid the important collision.

Tian Sang refined tutorials and wooden controls, introduced Barlow typography, and integrated contact-driven stage Foley. Interaction audio helped edits feel deliberate and physical.

The soundscape combined original procedural cues with separately credited recordings. The licensed stage music, sampled audience reactions and later generated scene themes each retained their own provenance.

The route selected ten scenes while preserving the opening tutorials. The mill and suspended cargo became a continuous puzzle, and later rescue chapters required the companion to arrive alive.

That changed the meaning of success. Reaching the hero’s exit could no longer stand in for a successful rescue. The authored solutions and failure cases needed to match the story the game was telling.

Two short Kling videos generated through fal helped inspect walking and idle motion. One did not produce the requested full turn, so it was not accepted as a turning solution.

Those videos stayed visual references. They were not converted into skeletal animation or presented as gameplay footage. The team continued with bounded edits to existing clips, planted idle poses and a dedicated death reaction.

The Grand Finale joined the route as an eleventh scene. Successful cinema could continue the journey, and puppet presentation moved to 24 frames per second while the underlying puzzle simulation stayed at 120 Hz.

With many pieces moving at once, the team tightened integration: freeze a handoff, bring code and content into main, then start the next test or recording batch. Every result needed to identify the actual build it described.

Full-route review exposed premature freezes, rescue timing problems and a restored barrel that invalidated an old finale solution. A successful build was clearly not enough to establish that the route worked.

The team investigated timing neighborhoods in the native simulation while retaining the authored geometry and permissions. Reference cuts were corrected instead of moving the exit or weakening the arrival rule. Contact failures that remained were kept visible in the records.

The final stretch addressed render lifecycle stalls, puppet skinning, warm follow light and green rim light, first-click menu handling, camera layout and native UHD recording. Six outdoor themes were generated with Stable Audio 2.5 through fal.ai and edited into fixed offline loops.

English submission copy, credits and provenance were prepared alongside the game. These changes are part of the development history; they do not turn an earlier package’s passing checks into acceptance of a newer build. Final full-route recording and human review remain distinct steps.