Machine backlog until the 4 PM reset
The hourly check (a session cron) reads this, sees which machine is idle, and queues the top unblocked item for it. Each job names its machine (ROSTER.md). Mark items done with the queue seq and the time from date. Keep Claude spend small: queue, screen with contact sheets, log, post to the board.
Running or done
- [done 00:45] tower 1: overnight batch 1 (music, voice, props, textures): 27/27 ok
- [done 01:05] tower 1: stunt motion batches 1 and 2 (69 and 299 takes), ranked; previews of the top 3 per move in
~/mj-tmp/prev2on tower 1 - [done 01:25] laptop batch 2 (seq 564); tower 1 music batch 2 (6/8, seq 566) and props redo (seq 567); screened 03:05
- [queued 03:05] laptop: look frames 2 and chair redo (seq 568); tower 1: music retry (seq 569), Wan previz of 6 stunts (seq 570)
- [done 03:57] look frames 2 (posted), previz 1 (5B: crude); music retry failed twice (cut the fanfare from the main title)
- [queued 04:00] tower 1: kimodo fails batch 3 (seq 571), voice batch 2 full script (seq 572); laptop: props batch 6 (seq 573); Sonnet worker: laptop code-project packaging test
- Next for tower 1 after those: Wan 2.2 14B first-last-frame previz using look frames 2 as keyframes (klein edit makes the end frame)
- Rule from check 1: leave at least an hour of queued work per machine at every check.
Next, in order
- Laptop, when batch 2 ends: turntables of its 12 props, plus a contact sheet of the 8 look frames. Screen both and post the good ones to the board as dailies. Redo misses as batch 4 on the laptop.
- Tower 1, after the props redo: turntables, screen, post (batch 1's keepers and the redo together).
- Tower 1: voice batch 2. Director walk-through lines for the three scenes (STORY.md), in Chatterbox and Dia, 3 takes each, run through the Whisper takes check. Wait for Don's casting pick if it's on the board; otherwise use the top 2 reels by the checks.
- [queued seq 570] Tower 1: Wan 2.2 previz, text only (never MetaHuman frames): one 5 s 480p clip per hero stunt (cart vault, gate slide, beer tent haymaker, chair smash, cannon launch, truck landing) as choreography reference. About 2.2 min each.
- Laptop: look frames batch 2 from Don's notes on the first set.
- Tower 2: stunt previews as a queued CPU job (
game/motion/preview_takes.py), so future motion batches preview themselves. - Music: after batch 2, screen with CLAP and audiobox-aesthetics scores (free critics on tower 1), and post the best take per cue as a daily with audio. The board shows images and video only, so audio needs a small board edit first.
Waiting for 4 PM (needs Claude effort)
- [DONE 11:26] Take 1 root alignment (b1c8e55): all boundaries under 4.6 cm a frame. Next: re-cut the slide clip's standing start; ease the vault flub drift.
- A/B retarget trial (Don's X find, 11:40): toyxyz-blender-retargeter (GPL-3.0 Blender add-on, run as a tool, no code shipped; github.com/toyxyz/toyxyz-blender-retargeter) against our UE IK Retargeter on the same 4 takes. Measure foot slide (contact-frame foot travel), the joint-limit gate and the root drift. Adopt only if it wins.
- Unique faces (Don's X find, 12:35: TREE's MetaHuman Maker makes custom faces, 5.8 only): stock presets are recognisable to judges. Test MetaHuman Creator's own blend and sculpt in 5.9 by script: Rex as a blend of Bruce with 1 or 2 other presets, kept close to Don's pick C; the same for Walt. Compare on one sheet against the stock faces.
- [DONE 13:50] Take 1 flubs swapped to vault_tumble_roll_on-s33 (s35 does not tumble) and mirror_bounce_on-s11; take 2: chair_eat_stagger-s6, table_short_roll-s17 (s13 stays standing).
- Take 2 fixes (92eb188): the heavy's haymaker throw (batch 11), cast a real heavy MetaHuman (tower 2, after the costume), [DONE 13:50: b2 beat, trestle lift for the slide, eat-the-chair clip, Rex_final into KidKatapult] still open: a b2 shrug-off branch.
- Ranking: add gag checks per move (a tumble needs the head below the hips; a fall needs the pelvis near the floor) so a take that skips the gag can't win.
- Costume into KidKatapult (C8, commit 35dacc2), on Rex_final in both takes; package size check.
- Costume C9c (f9a212d): flares good. Boots read as blobs, so fit the best TRELLIS boot from batch 12 rigidly to the foot bone, with the toe split at the ball. Close the dark notch on the right outer shin.
- Takes (a3c7fcb): the builder must not delete shared lights (tag take lights, leave untagged ones); re-check take 1 renders. Cannon flight: pitch Rex horizontal along the arc, frame closer, a get-up clip for the flop rejoin, model the truck. Swap all takes to the Maya1 voices in game/audio/voice (c27c994).
- Look pass (after 0851fb7): take 1 dim and blue (8 lux sun), so warm the dusk key and fill. Cannon camera too steep, with a black floor and the truck lost, so lower it and keep the truck in frame. Then the 1978 grade (style-film-1) on the film view.
- Costume v2 (take 1 failed 10:25, crude script garments): director designs it. Rex's posed body from a runtime frame, Blender cloth drape of pattern garments (leather jacket, open collar shirt, flared jeans), skin-weight transfer, import on the MetaHuman skeleton, Chaos Cloth on the jacket hem.
- [DONE 06:05] Jam project game/KidKatapult created (2df8b0e); size trim running (Sonnet worker)
- Don: allow the Windows Firewall prompt for ".NET Host" when a packaged game first runs (system setting, his click)
- [DONE 05:13] Hair in packaged builds: fixed by packaging the laptop code project (D:\megajam\codetest, frames in frames/). Next: convert the jam project the same way when it exists; a cloth garment test; size.
- MetaHuman Maker test on a UE 5.8.4 copy.
- [DONE 08:30] Retarget pipeline (7f7014a): SOMA to MetaHuman, gate clean except wrists. Next: hand pose offset for wrist twist; a ranking term for action energy (peak root speed, limb travel) so tame takes don't win.
- [DONE 09:35] Take 1 end to end (d0639f5), greybox. Next, in order: flub sections in the sequence; real props and the money man on the set; clip blends; light and grade (style-film-1 look); cue window tuning from Don's +216 ms data and a pass-rate target; costume for Rex; then Don plays it.
- A video critic for motion (qwen3-vl 32b on clips, then a fresh Opus reviewer on the finalists).
TOP PRIORITY (Don, Thu 23:03): flawless branches
Rule now in lab/STANDARDS.md (Game director). A hit plays the take exactly as authored; a miss plays a pre-authored variation that blends in and out with no pop, slide or snap; each cue has a commit cutoff before its branch motion starts. Ticket for the laptop right after the sets: (1) every flub branch ends in the main path's pose and position by its own motion (root-aligned branch clips that rejoin at a shared pose, or a recovery clip), no root easing; (2) crossfade or inertial blend 0.2-0.4 s at every branch entry and exit, measured: max joint jump per frame at each seam under 5 cm and no root speed spike; (3) a commit time per cue (window close plus blend lead), logged; (4) the same for wreck resets: they cut away (Walt's "Cut!" and a camera cut) instead of snapping the actor; (5) the rope flub's lying-to-standing in 0.25 s gets a real get-up clip. Checked by numbers per seam, then Don's eyes on the packaged build. Run it as the Fable test (Don, Thu 23:10): worker model fable, effort low, on the laptop right after the sets; inputs game/takes/SEAMS.md (seam audit), batch 16 recovery clips, game/ACTIONS.md (action bible and interaction methods). Record tokens, time, compile rounds and seam numbers in devlog/trials.md against the Opus tickets.
Jam week queue (Thu 17:00)
- [DONE 17:55 9cc8c4b, 855 MB] Laptop: import Heavy_v1 (game/casting/HEAVY_IMPORT.txt), retarget the brawl's heavy clips to him, swap the take's heavy actor; package and check the zip against 950 MB (trim stills assets, shared textures).
- [DONE 18:00 0839158, on the board] Heavy costume. NEXT on the laptop after the premiere: import the costume (HEAVY_IMPORT.txt), swap actors.heavy.bp, re-check size; open: boot toes, coat underarm in big swings. Was: 1970s work jacket or denim vest over a tight tee, work pants, boots (reuse the Rex costume pipeline in game/costume); shoulders broader. Nothing to Don before it.
- Tower 1: prompt redos for moves with no passing gag take (gate_slide starts in the air, kungfu_flurry too slow, flub_late_jump, dust_off, brawl_beat_a); write new ranking.md back to tower 1.
- Walt build: bald, flat cap, Walter face.
- Premiere and credits; between-scene beats.
- Edit bay and call sheet dressed in the real look (fonts, paper, film-can textures).
Jam week pace (Don, Thu 16:10: monitor usage, push a little more than overnight)
- Weekly quota reset to 0% at Thu 16:00; next reset Thu Nov 19 16:00. The jam ends Sun Nov 15 21:00 (77 h).
- Target: about 60% used by the deadline, leaving 40% for the four days after. Pace line: used% = 60 x (hours since Thu 16:00 / 77), about 0.8 point an hour, roughly 2.5x the overnight burn.
- Keep 3 to 4 workers busy across the three machines while under the line; at more than 3 points over, drop to two workers and Haiku helpers; at more than 6 over, director only.
- Fable only for the approved trial unless Don says more. Hourly checks log to game/devlog/usage-pace.md.
Ration plan, CLOSED Thu 16:07 (Don, 01:42: use the remaining quota up to the 4 PM reset, no prompting from him)
- Reading at 01:42 ET: weekly all-models 92% used, resets 16:00 ET (14 h 18 m). Fable 3%, not used.
- Budget: spend 7 points by 16:00 and keep 1 point in reserve, so the pace line is used% = 92 + 7 x (hours since 01:42 / 14.3). That's about 0.5 point an hour.
- Every hourly check reads
get_usagefirst and logs the reading ingame/devlog/usage-pace.md.- More than 0.5 point below the line: spend the slack on the next "needs Claude" item. Use a Haiku helper first, and one Sonnet worker for a scoped ticket with a check.
- On the line: queue, screen, post.
- Above the line: queue only, skip screening reads, and write a two-line log.
- The biggest cost is this chat's context (about 320k tokens re-read every turn). Keep turns few and short, and send file-heavy reading to Haiku helpers.
Next on the laptop after the sound wiring (Thu 19:30)
- DONE Thu 19:55: costume FX1 (Rex, heavy, Walt) imported and worn in all three takes (outfit field), Walt's WC1 costume folder and flat cap (walt_cap actor on the head socket, Movable) in. Rex's 8 Rex_v2 textures restored; M_CostumeV4 had no default textures and failed to compile in the cook (default material on all of Rex's FX1 costume), fixed in Scripts/fix_costume_defaults.py. Packaged D:\megajam\builds\full3, 911 MB zip, premiere bot loop clean (0 errors in the build's log).
- Accepted for now: upper-arm skin through Rex's jacket and the heavy's coat in the haymaker and stride (up to 26 points, about 2 cm). Next try is a pose-space corrective; only if Don sees it.
- DONE Thu 19:55: Walt's batch 14 moves on Walt_v3 (/Game/Motion/KimodoWalt): idle_director-s7 loop, clap_once at "action", shout_cut at "Cut!", facepalm in each wreck reset, applaud_slow in a new applause section (sequence 100 s) that the game jumps to when a take ends at 5 stars. arms_folded_watch-s9 and point_action-s1 are retargeted but not placed yet (walk-through pointing has no hook).
- [DONE 2026-11-12, zip 783.2 MB decimal in D:\megajam\builds\full4] Lip sync in: 39 face anims under /Game/KKCast/LipSync, builder fields line.face_anim / lipsync / face_actor and actor.face_component (SCRIPT-FORMAT.md), 40 ms lead, speakers fixed (walt, kid = Rex, ad), blinks and saccades baked into the anims plus EL_walt / EL_rex eye life under the gaps (ls_eyes.py). Size: body and clothing textures capped (D:\megajam\builds\full4_rules\caps.json), pkg_smoke.ps1 gate at 940 MB decimal. Still open: judge sync by eye in the editor (no MetaHuman frame goes to an AI model).
- Sets (7ee1055, layouts checked, 215 items, 0 conflicts): after the take extension lands and batch 15 props arrive, a laptop ticket: rerun layout_tool.py paths from the new bot runs, apply proposed/take_props.diff (real cart, gate, platform, crash pad), apply proposed/KKGame_settags.patch (so a director rebuild no longer deletes a set), run set_build.py after every director build, package and check the gate. Stand-ins until batch 15: trucks, armored car, mirror maze, string lights, bleachers, pennants, bar, benches, bunting, tent roof, dolly, cables.
QA tower 2 (2026-11-12, packaged build D:\megajam\builds\full2, bots on tower 2, logs only)
Runs: 36 single-scene matrix runs (mode x strictness x bot plan, chase, 53 to 57 s each), 6 full loops, 3 keyboard-driven loops, 1 smoke. Raw logs: tower 2, D:\megajam_qa\runs\. Scripts: game/qa_tower2/. No crash, no error line from LogKK, every run exited 0. Full loop (sheet, 3 scenes, bay, print, premiere, credits skipped, back on the sheet): 251 to 281 s with one attempt per scene, 361 s with two, 421 s with three. Frame rate (log frame counter, per take): chase and brawl about 85 fps, cannon about 127 fps; coarse 1% low (0.3 s windows) 55 to 85 fps, one 38 fps window in m_take_book_late near film 40.4 s (slap sound). The game logs no fps itself. Stars fit the plan: perfect 5 stars in all 9 mode x strictness cells; none and late 1 star in all 18 (3 flubs, 2 wrecks, still finishes and prints a cut); human (sd 45 ms) falls with difficulty: Rehearsal 3/3/2, The Take 3/2/1, Director's Cut 2/1/1 (easy/book/strict). Film stars equal the cut: 5 for perfect, 4 for the perfect,late,human mix, 1 for none.
- [ ] BUG (medium, intermittent, may be harness): the wrap screen ignored Enter after the 2nd attempt. 2 of 4 keyboard runs (k1 brawl, k3 chase): FINAL logged, "Player bShowMouseCursor Changed, False -> True", then Enter 3 s later (and two retries over 20 s) did nothing, no further LogKK line for 2 minutes, process Responding and rendering. The same step worked in k2 and k4 and in scene 1 and 2 of the other runs. Log (k3):
[2026.11.13-02.07.10:272][741]LogKK: KKRESULT_FINAL stars=1 score=0.0 print=0thenLogViewport: Player bShowMouseCursor Changed, False -> True, nothing after. Not proven to be the game (keys are SendInput to the foreground window; focus was not logged in k1 and k3, and k4 with focus logging did not repro). If it is the game, it is a soft lock. Needs a repro with a focus check and the mouse (click GO AGAIN) on a second machine. - [ ] BUG (medium, from code and log, not seen on screen): Director's Cut and The Take still draw the big yellow call text. HandleCueCalled sets CallText and bCallActive in every mode and the HUD draws them whenever bCallActive; only the marker ring uses Feedback.bShowMarker. The log prints
KK call c1_vault 'JUMP!'in all three modes (m_cut_*). The call sheet says Director's Cut is "No calls". Game/KKGame.cpp lines 596 to 606 and 1529 to 1535. - [ ] GAP (low): DESIGN.md says Director's Cut has "one extra stunt". The chase has the same 5 cues in all three modes (
KK take started, 5 cues, mode DIRECTOR'S CUT), brawl 5, cannon 4. Only the windows are tighter (perfect +/-26 ms, The Take 33, Rehearsal 40, all at By the book). - [ ] NOTE (design): the "late" bot adds 300 ms on top of the 216 ms the perfect bot already adds (KKSetBot, BotExtraMs 300), so it is 516 ms after the cue, and it scores 1 star even in Rehearsal on Easy (scrappy window +/-114 ms). Strictness cannot be seen with perfect or late, only with the jittered "human" bot. The perfect bot is the Don-like +216 ms case (it grades PERFECT with err 4 to 7 ms). If a slow but honest player should get 2 or 3 stars on Easy, there is no bot for that yet.
- [ ] NOTE (low): first launch on a fresh machine took 34 s from start to the call sheet and logged 55
Waited for PSO creationlines up to 1.6 s each (smoke run). Every later launch: 0. PSO precache is not shipped in the pak, so a player's first take after install can hitch. Worth a PSO cache or a warm-up on the sheet. - [ ] NOTE (low): every launch logs 16 engine warnings that are not ours: LogToolsetRegistry "Cannot register Toolset SequencerScripting..." (13), LogStreaming SkipPackage /MetaHumanCoreTech/RealtimeMono/DefaultSmoothing and /ToolsetRegistry/ToolsetManifestSeed (2), LogUObjectGlobals "Failed to find object" (2). A clean packaged log should not carry editor-only toolset or MetaHuman realtime references; check the cook excludes them. LogRenderer "The film grain texture is not streamed in" appeared in 4 of 36 matrix logs.
- [ ] NOTE (harness): the premiere bot quits 3 s after the call sheet reopens, so "start again" was tested only with no bot (k4_nobot: Enter on the sheet after the credits rolled scene=chase and
KK take startedfollowed). Back and forth in the bay (Esc to the wrap, Enter back into the bay, Left, Right, Down, Up, C) worked in k2 and k4 and kept the picks. - Not covered: touch and controller input (SendInput keyboard only), a second display, long idle on the sheet, Alt-tab during a take, real fps counters (nothing in the game logs them), the soft lock rule "no input moves on for 30 s" was checked by log gaps over 20 s (none in 45 runs) and by the keyboard scripts' wait timeouts (the two stalls above).
- QA fixes (after the take extension; from QA tower 2, 6737ed6): (1) the wrap screen sometimes ignored Enter after attempt 2 (2 of 4 keyboard runs, right after the mouse cursor turned on): make menu input independent of cursor and focus state, and log every accepted input on menus; (2) Director's Cut and The Take must not draw the big call text (Director's Cut: no calls at all); (3) difficulty: a human-like bot (sd 45 ms, centred) gets only 3 stars on Rehearsal/Easy; target 4 there, 3 on The Take/By the book, 2 on Director's Cut/Out for blood, tuned by bot runs; (4) warm the shader cache so a first launch doesn't stall 34 s (a PSO cache or a warm-up level).
- QA fixes status, Fri Nov 13 (fun tickets 1 to 8): (1) menu input now reads a key queue filled by the controller's InputKey plus a key state edge, in game and UI input mode with the mouse never locked, and logs every input (KKMENU); the stall itself was not reproduced and the fix has not been tried with real keys (the QA harness steals focus). (2) done: Rehearsal draws the word, ring and buttons, The Take a subtitle, Director's Cut nothing. (3) done by the star points and the bot table in devlog/notes.md. (4) still open: the PSO warm-up.
- Flub branch sliding (3dfe3f1): flub branches ease Kid back onto the main path by up to 3.7 m (maze flub), and the rope flub goes from lying to standing in 0.25 s. Fix with root-aligned flub clips (align the flub's end to the main path, as KimodoAligned does for main clips) or a recovery clip (dust_off / get-up) instead of a slide.
- SFX model bake-off (scout, Thu 21:55): MOSS-SoundEffect v2.0 (Apache-2.0 per its card; verify the repo LICENSE and weights terms first). On tower 1, remake the 10 weakest code-synth effects (crowds, footsteps, body roll, table slide, reel start, projector, ambience beds) and rank against ours with CLAP and audiobox; Don's ears decide. DONE (job sfx-moss-1): licence clean (Apache-2.0 repo LICENSE and card), MOSS beat ours on 16 of 17 by CLAP and Audiobox, winners in game/audio/sfx/*_m.wav plus phone_ring, phone_squawk, phone_click; sound_map proposal in prep/moss-sfx/SOUND_MAP_PROPOSAL.md; reel at D:\megajam
eview\sfx_ab\sfx_ab.mp4. Waits on Don's ears.
- Walk-throughs in (after the sets ticket): game/takes/WALKTHROUGH.md has the full spec (1314614): play walk_<scene> before a scene's first attempt, skippable after 2 s, show the walk take's actors plus the scene's set, hide the cast, props and film lights, no grade; import lip sync stage2 (21 anims, 6729ca3) and add lipsync:true on the walk lines; wire Walt's six calls walt_c01..c06 into the takes per the table in WALKTHROUGH.md; consider retargeting take_walkthrough-s11 for Kid's walk.
- Beats in (after the walk-throughs): game/takes/BEATS.md (5ab495b): three Stage 7 beats after each edit bay, good/rough Walt line by the scene's stars, skippable after 2 s; lip sync stage3 (15 anims, tower 2 now); phone sounds from the SFX worker; a squawk box prop (1970s studio intercom, small; next prop batch, or the batch 2 megaphone as a stand-in). Beats run 22-25 s: keep them, they're content.
Quality strategy tickets (Don, Thu 23:20)
- [DONE 2026-11-13 08:40] Package profiles:
pkg_smoke.ps1 -Profile jam|ultra(KidKatapult/Config/Profiles/README.md). All 253 baked texture caps moved out of the assets (Max Size back to 0; capped textures now carry a Project texture group, and the jam device profile gives each group its MaxLODSize; cooked sizes of the capped textures match the old baked ones, 89.4 MB both). Jam also drops SM5 shaders, the Vulkan debug layers and opencv_world455.dll, and GC_stunt_rig is reimported with constant topology (120.6 to 76.0 MB). Jam zip 857.6 MB (was 927.0; D:\megajam\builds\jam1), ultra 1617.3 MB (D:\megajam\builds\ultra1, no gate); the packaged premiere loop ran with 0 errors on both. Still baked in the set meshes: Nanite kept-triangle percentages and distance field scales from sets/set_slim.py (need a re-import to undo). - Trailer and extras at ultra: an MRQ render path for the itch trailer, key art and the dev-log video at the highest settings (path tracing or high Lumen, motion blur, 24 fps), with the 1978 film finish done in-engine or with non-AI ffmpeg filters (gate weave, grain, scratches, halation) because MetaHuman frames never go to an AI model.
- Action bible (f0a1089): 55 of 79 action rows have gaps (16 motion wrong, 12 no clip, 11 prop static, 9 prop absent, 7 contact missing). Plan and ticket order in game/ACTIONS-PLAN.md; Blender halves of proof tests P1-P6 running on tower 2. Note: studio-pipeline docs commits from the jam land on branch virtual-staff (that checkout's branch); merge to main after the jam.
Story and stunt fixes (review Thu 23:40; decisions.md)
Cheap, now (offline worker): STORY.md climax rewrite, new and changed lines voiced on tower 1 (Walt's landing line, the effects man, "Keep rolling", the pickup line), speaker retags, walk and beat scripts updated, a patch for longer_takes.py/take JSONs applied once the laptop's sets ticket is done. Into the laptop tickets: T7 cannon gets the helmet and goggles (head socket), the truck at 35-40 m with a box stack and mattresses under a cargo tarp, the barrel pitched to the arc, flight at real speed (about 1.8 s), a back-first landing clip, the effects man firing; T3/T4 brawl gets the heavy's real swings, the chair from a visible spot with a two-hand hold onto the back or the table, the table collapse, candy-glass mugs, the rope to the grid with the heavy's chest pad and a pull-back; T6/T10 chase gets the truck at about 4.5 m/s, a hidden grip, a robber (the heavy) shaking Kid off into loose straw over boxes, the second leap as a pickup on the parked truck; T9 the bandstand under the fall with a raised deck and a balsa rail.
Set pass 2 (director's look at the set-only views, Fri 00:15)
- Beer tent: good (striped tent, tables, bar, bandstand with bunting). Needs patrons.
- Midway: reads like a storage yard (crates, pallets, barrels) and the popcorn cart is a crumpled-foil blob. Needs: a regenerated popcorn cart (5 candidates, classic 1970s wheeled cart with glass case and striped canopy), string lights with warm bulbs over the run, booth signs, fewer crates.
- Cannon field: the cannon and platform read well; the far tents are flat striped boxes, the truck is tiny and far, the crash mats are red blobs (replaced anyway by the box stack under a tarp, T7), lots of bare dirt and sky.
- All sets: no people. A carnival needs a crowd. Crowd extras: one shared body mesh and skeleton (an existing cast body, no new MetaHuman), a few material and hat variants, offset Kimodo loops (cheer, clap, point, drink, eat), 10-20 per set, cheap on size because mesh and textures are shared.
- Content/Sets (248 MB source) is not in git; tarred to tower 1 backups.
- MOSS-SoundEffect won 16 of 17 cues by the critics (497787e; licence Apache-2.0, verified). Next laptop slot: apply prep/moss-sfx/SOUND_MAP_PROPOSAL.md to game/audio/sound_map.json (one line per cue, keep ours for the crowd gasp), restage audio, and add "MOSS-SoundEffect v2.0 by OpenMOSS (Apache-2.0)" to the credits. Don can overrule any cue from the A/B reel on the board.
- Set pass 2 assets ready (Fri 01:15): batch 19 keepers in D:\megajam\props\batch19 (popcorn cart c1, sign board c1, tent c1 scaled to about 6 x 6 x 5 m), string lights and bunting meshgen scripts, crowd loops (idle_crowd_watch-s20, cheer_arms_up-s12, clap_steady-s13, eat_popcorn-s20, gasp_hands_to_face-s12, drink_from_cup-s4; point_and_laugh needs a re-prompt). Laptop ticket after the fun fixes: place them via the layouts (string lights over the midway run, bunting in the tent and cannon field, real tents replacing the flat boxes, fewer crates), crowd extras on one shared body with material and hat variants, 10-20 per set, numbers-only checks plus set-only views.
- Constrained branches proven (batch 18, 69d9f8b): Kimodo full-body constraints on the first and last 3 frames pin pose and pelvis velocity at both seams (0.01-0.02 cm, 0%), and with text guidance 4.0 the middle sells the action (22/40). Batch 20 regenerates all flub branches this way. Cart vault and rope swing still fail in Kimodo (0-1 of 70): key them in Blender on tower 2 after the chair swing, or try end-effector (hand) constraints.
- Story fixes in (01bf1d6, Fri 01:37): STORY.md climax rewritten, Walt's landing line re-recorded ("Land flat on your back and let the boxes do the rest"), the effects man cast (Maya1 brooklyn, distinct from Walt and Rex), "Keep rolling" and the pickup line made, Rex's lines retagged. Lip sync still owed on tower 2: walt_w12 (redo), walt_c07, walt_c08. Don to listen to walt_c07. "Keep rolling" plays once T7 makes the cannon wreck a run-on; the pickup line waits for T6/T10.
- Rope swing kick keyed (98d879d): joins in 0.02 cm, hands on the rope, both feet on the heavy's chest pad, rope from a 7 m grid anchor (needs a canvas slot), heavy pull-back clip. For the brawl ticket. Check: cs_chair_intact_hold.fbx (fec761f) is probably rotated 90 degrees (one-bone prop keyed through the pose; the vault worker hit the same bug).
- Cart vault: low flat-top peanut wagon on tower 1 next, then keyed in Blender (decision Fri 02:47).
- Set pass 2 DONE (Fri 02:56, laptop): batch 19 cart (vault cart 0.43 scale), sign boards on booth roofs, carnival tents (about 6 x 6 x 5 m) with four more on the cannon field, string lights zig-zagging over the midway run (warm emissive bulbs), bunting in the beer tent and between the flag poles, crates 7 to 1, and 43 crowd extras (19 midway, 12 tent, 12 field) on the shared Rex body with the six batch 19 loops, 4 outfit colours and 5 hats. Details in game/sets/README.md "Set pass 2". Package D:\megajam\builds\sets2 (851.5 MB zip).
- MOSS-SoundEffect applied: sound_map.json swapped per prep/moss-sfx/SOUND_MAP_PROPOSAL.md (gasp kept ours; walk_dirt_run and reel_start added as new entries nothing plays yet), phone_ring, phone_squawk, phone_click mapped for the beat takes, credit line added to premiere.json.
- crowd faces: build 2-3 cheap extra faces (MetaHuman blend of presets, low LOD, shared body) on tower 2. Until then every extra has Rex's face, so the layout keeps extras 8 m or more from every camera position and puts brimmed hats on those within 16 m. Not started.
- point_and_laugh crowd loop still needs a re-prompt (not in the crowd mix).
- Cart vault keyed (14b6d71, Fri 03:05): over the batch 21 wagon (c2, four wheels, 0.9 m top), joins 0.01 cm, palms on the top, nothing clips, the wagon tips 1.8 degrees and settles. Reads as a tucked squat vault, a little floaty at the top. For the chase ticket.
- Set pass 2 coverage (packaged sets2, bare share of the shot, worst per take): chase 1.9%, brawl 0.0%, cannon 41.2% at 22 s (31.4% and 29.3% at the other two; before the pass 37.0%). The cannon is over the 40% target by sky and foreground dirt because the batch 19 tents are smaller than the flat boxes they replaced; accepted by the director, and the cannon set changes again in the cannon ticket. Set-only views: D:\megajam
eview\sets2\.
- Brawl rebuild (6def097, Fable test): in: the batch 20 branches, the heavy's swing, haymaker and shove, the keyed chair smash with the table and candy glass, the rope kick from the grid. Fixed for every take: crossfades were hard switches (frames vs ticks), and "hold" looped; chase and cannon need a rebuild to get it. Still failing: fist contacts 98-144 cm short (B03, B06, B17, B19: clips made one man at a time), main joins jump 9-74 cm, the bandstand fall (the rope swing comes in on the same line), the heavy's reversed walk-back. Next: a two-person contact pass (world-bake the brawl and pin fists and impacts in Blender, method 2.1), with a planning step first.
- Brawl staging solved (699896e): predicted contacts at impact 0.2-6.8 cm (uppercuts on a stand-in until batch 23), riser placed, fight turned 85 degrees off the rope line. Set fix needed: move bench t_e_bench_a and table t_e east (the heavy's pull-back foot lands in them at 23.93 s).
- Cannon rebuild (T7, Fri 07:40, laptop): in: the keyed tower 2 clip on Rex (bore hold, fire on Walt's word, true arc at real speed, back first into the box stack, sit up, stand, wave) riding the rolling flatbed; the stunt rig (boxes, mattresses, tarp) as a geometry cache on the truck; barrel at 46 deg, muzzle 378 cm, deck 298 cm, front wheels on a crate; helmet and goggles on the head socket (hair groom hidden); the wave turns him to the barrel, the climb-in lies him back into the bore on the clip's first frame (seam 11 cm in the joint log, under the take's usual 21); belly_flop-s17 now flops into the same boxes from the top of the arc, land_slip-s5 is a run-on with walt_c07, stage_fright-s1 constrained; premiere plays release to contact at 0.4. Chase and cannon rebuilt with the fixed builder (crossfades in ticks). Package D:\megajam\builds\cannon2, 927 MB (gate 940). Seams and what is open: game/takes/SEAMS.md "Cannon rebuild". Still open: the run-up onto the raised deck does not climb the ladder (it lifts over the deck edge), the belly flop entry is a 0.6 s crossfade between a tuck and an upright pose (s17 was made on the old flight), stage_fright-s1 slides 92 cm to the high five, land_slip-s5 rejoins with a 68 cm spread and a 0.4 s crossfade (only a wide human bot reaches it: c2 missed swallows c4), the slow motion leaves the score 1.7 s ahead until it resyncs at the landing. The zip is 13 MB under the gate.
- Cannon rebuilt (f054b77): real-speed 37 m flight, back-first landing in the crushing box stack on a rolling truck, helmet and goggles, "Keep rolling!" run-on, chase and cannon rebuilt with the crossfade fix. Package 927 MB (13 under the gate): profiles and size ticket running. Open: belly flop entry needs a new branch clip (made on the old flight), stage fright slides 2.3 cm a frame, the ladder climb lifts him over the deck edge, music runs ahead in the slow-motion flight.
- Package profiles (acba558): pkg_smoke.ps1 -Profile jam (857.6 MB zip, 940 gate) and -Profile ultra (1617 MB, full quality, for the post-jam Director's Cut edition). Every texture cap now lives in the jam device profile; source assets are back at full size. Jam vs ultra frame stats agree within 0.03 at 720p. For the submission: build Shipping instead of Development (about 70 MB smaller) once the bots are no longer needed.
- Ladder climb keyed (69bab36): hands on rails within 1.6 cm, feet on rungs within 1.6 cm, joins 0.03 cm; needs the opened platform SM_cannon_platform_ladder_gap.glb (its rails crossed the ladder top). stage_fright-s1 constrained pick. Belly flop: no usable regeneration after two runs; the old s17 with its crossfade stays; next try pins a face-down pose at the box impact.
- ROOT CAUSE FIXED (Fri Nov 13 wiring run, SEAMS.md "The 12.8 cm sink"): the director's baker wrote the library's pelvis read, 12.8 cm under the stored track; compensations removed. Was: in the packaged game every clip plays about 12.8 cm low (joint log: feet 10-14 cm under the floor, the deck too). The cannon clips and the ladder were lifted 12.8 cm to compensate. Find the cause once (Rex's mesh or root offset, the retarget, the actor's capsule or the builder's z), fix it for all takes, then remove the per-clip lifts (ZOFF in lad_geom.py, the cannon lift in cannon2_take.py).
- Brawl two-actor bake (9455382): all five gated hits land within 1.41 cm (were 12-23 cm), reactions start within 1 frame, planted feet slide under 0.7 cm, deterministic. The riser's balsa break is keyed (Bullet threw pieces). Uppercuts read a little like short straights. Two pelvis teleports left (Kid 17.60 s, heavy 10.97 s): bridges in batch 24 with the six branches. Then ticket 7 (laptop build, import the baked FBX, check against bake_run1.npz) and ticket 8 (fresh review).
- Wiring run (Fri Nov 13): sink fixed, ladder climb (four rungs, real heights) and the opened platform and the stage_fright t7 pick in the cannon, walk-throughs and beats play from the call sheet (WALKTHROUGH.md, BEATS.md "In the game"), Walt's calls placed (walt_c02 left out, it overlaps walt_l06). Still open: Rex's off-camera lines have no face flag by design; ladder foot contact 3.7 cm (was 1.6); a squawk box mesh; the belly flop pick.
- Batch 24 (39955e4): six brawl branches and two bridges regenerated against the baked fight, pin gaps 0.0 cm. Ticket 7 next on the laptop once the wiring ticket frees it: import the baked FBX and batch 24 picks, build brawl3, measure seams and contacts in the joint log, package, bots. Watch rope_miss_long-s22 (166 joint-limit hits; backup s3) and bad_landing's 6 cm floor sink in the lying pose (fix in the bake if the sink root-cause fix doesn't cover it).
Post-jam ideas from Don (Sat Nov 14, 02:5x)
- Phone as controller: gyro aim plus fire from a local web page over Wi-Fi, a UE WebSocket receiver, no app install.
- Retro devices: confirm absolute-mouse light guns and Wiimote with a sensor bar in mouse mode on real hardware; Steam Input gyro profiles.
- Two players: two new hires on the job with Lenny (story change: who gets snatched, split routes).
- A free-roam FPS mode alongside the ride.
- DLSS with frame generation for the commercial version (follow the vendor attribution rules then, which the jam bars).
- Commercial release: clear every item in spoiled/ship/LICENSES-COMMERCIAL.md first.