BEATS.md
Source game/takes/BEATS.md. Last git commit 2026-11-13. About 1,481 words.
Between-scene beats (CONTENT-PLAN section 4): the laptop ticket
Three takes on Stage 7, one after each scene's edit bay: beat_1.take.json (after the chase), beat_2 (after the brawl), beat_3 (after the cannon, before the premiere). Written by make_beat_takes.py (rerun it if a line is re-recorded: every time comes from the wav lengths). The shoot, not the film: no 1978 grade, neutral work light, no cues, no score, no music.
Walt stands by his chair, Rex (BP_Rex_final in the 1978 costume) by the craft table. The set's squawk box (the studio phone, on apple box apple_3 beside the chair) buzzes and squawks; Walt answers the front office about the budget and switches it off; Rex fusses about his close-up and his double; Walt answers him. Then Walt reacts to the dailies, good or rough.
| beat | main run | then | lines (main) | good | rough |
|---|---|---|---|---|---|
| beat_1 | 0 to 18.00 s | jump to good 23.00 to 29.45 or rough 31.00 to 35.95 | walt_b01, rex_b01, walt_b02 | walt_b03 | walt_b04 |
| beat_2 | 0 to 17.83 s | good 23.00 to 28.31 or rough 30.00 to 34.71 | walt_b05, rex_b02, walt_b06 | walt_b07 | walt_b08 |
| beat_3 | 0 to 18.09 s | good 24.00 to 31.10 or rough 33.00 to 39.95 | walt_b09, rex_b03, walt_b10 | walt_b11 | walt_b12 |
A play is 22.5 to 25 s, not the plan's 15: the lines run at Walt's real pace. The skip after 2 s covers repeat plays. To get near 15 s, drop Rex's exchange (his line and Walt's answer, about 9 s) from BEATS in the script and rerun.
How the game should play them
- After a scene's edit bay closes (the player has cut the reel), play
beat_<n>for that scene (sequence taggedDIR_beat_<n>), then go on to the next scene's call (beat 3: on to the premiere). Once per scene, never on a retake. - Skippable with any input (keyboard, pad, touch) once 2.0 s have played (
beat.skippable_after_s). Input before 2 s is ignored, so the button that closed the edit bay can't skip it. A skip stops the beat's audio and moves on with the usual fade. A skip during a branch does the same. - The pick. Read the stars of the reel the player just cut for that scene (
FKKReel::Stars(Tuning)in KKPremiere.h, the number the critic will use).beat.pick.good_if_stars_at_leastis 3: 3 stars or more playsgood, 1 or 2 playsrough. If no reel exists (a debug jump), use the best attempt'sFKKAttempt::Stars. - The jump. Play the sequence from 0 to
beat.jump_at, then seek to the chosen branch'st0and play to itst1, then the beat ends. This is the resets' mechanism (sequence-only sections after the main run). Each branch opens on its own shot, Walt's single (s2_good/s2_rough), so the jump is a camera cut and the small pose change at it doesn't show. Never play through pastjump_atwithout the seek: the stretch fromjump_atto the good branch is empty (the two-shot holds over idle loops), and the rough branch would follow the good one. - No cue gate: director_cues.py writes an empty table for these (warning W03 "the take has no cues" is expected). Nothing goes into the edit bay, the premiere or the blooper reel.
- Visibility: show
DIR_beat_<n>plus the Stage 7 set (KK_set_stage7); hide every scene'sDIR_actors andKK_settake_dressing. The beats carry their own work lights (work_sun,work_sky), like the walk-throughs: work lights off while a beat is up, or deletelightsin the script to keep Stage 7'sKK_worklightrig on. Not both. - Grade: no
gradekey on purpose. Skip any game-side grade or letterbox forbeat_takes.
Builder fields
beat(top level, read by the game only):after,set,skippable_after_s,grade: "none",worklights,jump_at,branches.good/branches.rough(t0,t1,line),pick. director_build.py ignores it.- Lines in a branch carry
branch: "good"or"rough"(information only; the builder places them at theirtlike any line). rexis the travel actor andwalta follower, as in the walk-throughs, so clips join with blends. Every clip runs back to back from 0 to the end, the last one held. Nobody walks; the travel is only the clips' own small drift.- Rex speaks as himself here (speaker
rex), so his face anims are/Game/KKCast/LipSync/rex/rex_b0non the actorrex. - Moves (all imported): Walt
KimodoWalt/idle_director_s7loop,point_action_s1at the box (beats 1 and 3) orshout_cut_s4to wave it off (beat 2),arms_folded_watch_s9while Rex talks,point_action_s1at Rex on beat 2's answer,clap_once_s6in every good branch,facepalm_s7in every rough one. RexKimodo/mark_check_nervous_s6looped (he checks his costume) andbow_to_crew_s13in beat 3's good branch, when Walt calls the cannon the shot of the picture. - Cameras:
s1a slow push on the two-shot from the south-east (24 mm; Rex, Walt, the box and the craft table in frame, lamp_3 and flag_1 behind),s2_*a static 40 mm single on Walt from the same side. The director's chair sits low in the foreground of s2 (its footprint is 1 cm off the line of sight, but its 95 cm top is far under the 158 cm lens).
Missing pieces (not this ticket)
- Sounds: done (set pass 2, MOSS-SoundEffect v2.0, Apache-2.0).
phone_ring,phone_squawk(gain 0.8) andphone_clickare ingame/audio/sound_map.jsonand staged byScripts/stage_audio.py, so the three takes'audiocues on the squawk box play them and the build no longer warns "sound name ... is not in sound_map.json". SM_squawk_box: no mesh yet; it builds as its placeholder (a small dark box, 24 x 16 x 12 cm, on apple_3).
Build and check (laptop)
run_director_build.ps1 -Take beat_1(and beat_2, beat_3), then rerunset_build.py allas after any director build.- Read the log:
PATH rexandPATH waltshould stay within a few cm of their start; if the idle loops drift visibly, droptravel_actorandfollowersfrom the script (the clips then play in place, without blends). frame_checks(s1 at 1.0 s and one second before the jump): Rex, Walt and the box in frame.
Lip sync for tower 2
All 15 lines have lipsync: true; until the anims exist the build warns per line and the faces play eye life only. Make them the way the cast's were (game/casting/ls_solve.py, ls_run.py, ls_check.py, LIPSYNC_IMPORT.txt):
/Game/KKCast/LipSync/walt/walt_b01towalt_b12(12 lines)/Game/KKCast/LipSync/rex/rex_b01torex_b03(3 lines)
Voice
Made on tower 1 exactly as the walk-throughs (~/mj-tmp/voice-beats, and ~/mj-tmp/voice-beats2 for seven lines trimmed after the first pass, new seeds): Maya1 with gen_final.py's settings (gravel for Walt, matinee for Rex, temperature 0.4, top_p 0.9, seeds 4321 + 1000 x idx + take, idx 200 to 214 and 300 to 306), check_takes.py (Whisper) with WER at most 0.05, Audiobox aesthetics, ECAPA similarity against walt_l01_t1 / rex_l02_t2 with the cast's threshold (median minus 0.10: Walt 0.354, Rex 0.378), finalize.py (0.15 s pads, -18 LUFS). Rows are in game/audio/voice/manifest.json with a beat field.
All 15 kept lines passed WER and similarity, none flagged. walt_b08 needed the retry round (takes 0 to 3 under the similarity threshold, takes 4 and 5 passed, 5 kept). rex_b01 is the first-pass text: the trimmed one ("...and the kid smaller") failed WER in all 11 takes because Whisper hears "the kids smaller", which sounds the same.
Story fixes 2026-11-13
All 15 beat lines were read against the new cannon climax (Johnny's evening show turned on the robbers' truck, the back-first landing on the boxes, the effects man firing a spring cannon) and the chase changes (a robber shakes Kid off, the second leap is a pickup on the parked truck). None contradicts them, so no beat line was re-recorded and the beat takes are unchanged.
In the game (wired Fri 2026-11-13, laptop)
Phase Beat, the same StartSide code as the walk-throughs (see WALKTHROUGH.md, "In the game"). NextScene (the Printed screen's way on, tap or button, or the bot) plays beat_<scene number> first, once per scene (bBeatDone), then AdvanceScene does what NextScene always did (next scene's call, or the premiere after beat 3). PrintCut builds the reel it just cut (FKKPremiere::BuildReel) and keeps its stars; the beat picks good at 3 stars or more, else rough (no reel: the best attempt's stars). TickSide plays from 0 to jump_at, seeks to the branch's t0 once, and ends at its t1; a skip during either part ends the beat and moves on. Visibility: DIR_beat_<n> stays, every scene's DIR_ actors and KK_settake_ dressing are hidden, Stage 7 stays (it has no scene tag). The beat takes' tables come from director_cues.py --write-anyway (E07, the gap between the main run and the branches, is the sequence-only section by design); KKBEAT reel stars|beat ... branch=/jump log lines show the pick. Test: KidKatapult.Sides.WalkAndBeatTablesCarryTheirMetadata (Tests/KKSideTests.cpp).
BRAWL-CONTACT-PLAN.md
Source game/takes/BRAWL-CONTACT-PLAN.md. Last git commit 2026-11-13. About 1,237 words.
Brawl contact plan (planning step before the engine ticket)
Fri Nov 13, 06:15. Numbers from brawl2_plan.py on the last layout dump and the clip tracks in D:\megajam\brawl2\work\tracks.
0. What the numbers add
Placement is half the miss. Three clips also aren't the moves they're named for.
- brawl_beat_a_s17 (B03 duck and uppercut): fastest hand 1.8 m/s, head dips 12 cm.
- brawl_beat_b_s1 (B06 "full spin" sell): hips turn 18° at most.
- heavy_wild_swing_miss_s12: fist at 115 cm at full speed. It can't pass over a head.
- These sell: heavy_haymaker_throw_s10 (10.4 m/s), brawl_uppercut_s3 (8.9 m/s), bandstand_fall_s1.
The joint log's noise is 19-25 cm a frame (stale ticks), so no 5 cm check means anything in the game until it's fixed. Every branch is Kid's, so the heavy can be re-solved freely; Kid stays in Kimodo (SOMA) poses where branches attach.
1. Beat sheet (world m from the layout, x = Kid's start facing)
| s | beat | Kid | heavy | contact |
|---|---|---|---|---|
| 0-4 | marks | (0,0) facing 0° | (0.75,0.05) facing 180° | |
| 4-6.4 | B03 | ducks, uppercuts (NEW) | hook at head height, spins round | hook 15 cm over the ducked head at 5.14; fist on jaw 6.21, his reaction starts that frame |
| 6.4-9.43 | B06 | guard, about (-0.4,-0.2) | haymaker_take trimmed to under 1 m, steps back in (bridge), haymaker | fist on Kid's jaw 9.43 |
| 9.43-11.9 | sell | snap and full spin, about 1.2 m back (NEW), travel 1.0 | follows through | |
| 11.9-13.2 | B08 | shoved | walks in, shoves | palms on chest at 13.10 |
| 13.95-17.1 | B10 | rolls clear | chair onto the table | as built |
| 17.1-23.2 | B13 (NEW) | dusts off near (-1.8,-0.3), runs at 21.3 | turns his back, struts 4.5 m to the bar crowd, turns at 22.2, steps at Kid | replaces the reversed walk |
| 23.4-24.5 | B14/15 | kick, lands near (2.9,1.55) | pull-back to near (4.9,2.3) | as built |
| 24.5-26.4 | B17 | ducks, uppercuts (second seed of B03's clip) | charges in (NEW), hooks, spins | fist on jaw; s1_upper moves 25.3 to 26.37 so he's square again |
| 26.4-29.4 | re-square | pivots to keep him in front | shrug_off_s11, circles a quarter turn left (NEW) | axis turns 60-90° off the rope line |
| 29.4-31.2 | B19 | bandstand_fall_s1 as authored | haymaker | fist on jaw on the fall's first reaction frame; shoulders break the riser's balsa front edge, 1.2 m behind his heels |
| 31.2-45 | out | gets up, bows 35.0 | dusts off | no joint inside unbroken riser |
Bandstand: move the fall. Turning the rope line can't fix it. The kick lands where the heavy faces the swing, and a haymaker drives Kid straight back, so any fall after the kick goes down the swing line. Fighters re-square by circling, which costs one heavy stalk; Kid's pivot rides on the new B17 clip, and the fall clip stays. The riser is hollow, breakaway front boards over a pad (1978 practice), first guess near (3.8,-0.8). The solve places it 30 cm or more clear of Kid and the rope over 21.3-24.6 s, not covering Kid in the camera.
2. Method: contact by placement, continuity by generation, residue by IK
- Staging solve (numpy). brawl2_plan.py grows into stage.json. It places whole clips rigidly (yaw plus xy) so each fist meets its target bone at impact, and starts the reaction on that frame (2 frames of shift at most). Kid anchors wherever a branch touches him.
- Joins are generated by Kimodo. New moves and every join gap over 2 cm become clips pinned 3 frames at each end (kimodo_constrained.py; batch 20 hit 0.0-0.3 cm, 0% velocity mismatch). No root ramps, no crossfades over gaps.
- Both men are world-baked in one Blender scene on the real Rex and Heavy skeletons, exported from UE, so no retarget follows the solve (that hop cost 16-17 cm of wrist drift). plib two_bone_ik closes the residue at impact (target 2 cm, 4 frames in, 6 out), feet lock on plants, and IK lifts the hook's path for both ducks. Each actor gets one continuous bake, cut per beat, with hard switches. The riser is the P2 fracture baked against Kid's bones.
- All six branches are regenerated against the new Kid path by kimodo_branches20.py with stage.json placements. IK windows avoid the constraint frames. The keyed rope and pull-back clips exist only on MH, so a pose converter to SOMA supplies their frames. rope_miss's cue moves before takeoff.
Rejected: Kimodo generating reactions to a strike, since it has never held a contact (rope 0/70, cart 1/40). Keying every beat from scratch is too slow, so it stays the per-clip fallback.
3. Tickets
| # | where | ticket | check | est. |
|---|---|---|---|---|
| 1 | laptop CPU, now | Staging solver, riser spot, batch 21 constraints, gag rules (duck 25 cm or more, fist 6 m/s or more, spin 270° or more) | fists within 10 cm on placement alone; riser clear 30 cm | 1.5 h |
| 2 | tower 2 | Round trip UE-FBX-Blender-FBX against the tracks; MH to SOMA converter | 0.5 cm; 1 cm | 1 h |
| 3 | tower 1 | Batch 21: new moves plus bridges, 40 seeds | gag; seams 2 cm or less, velocity within 20% | 1.5 h |
| 4 | laptop UE | Retarget picks, export real-body FBX | tracks match | 45 min |
| 5 | tower 2 | Two-actor bake, riser, mannequin sheets | fist under 5 cm at impact; joins under 5 cm a frame; plant slide 2 cm or less; reaction within 1 frame | 2 h |
| 6 | tower 1, after 3 | Batch 22: six branches | batch 20 numbers, no joint inside props | 1 h |
| 7 | laptop UE | Log after the anim update; "world" sections in director_build; brawl3_take.py; cue move; build, package, bots, seam_check | log p99 under 3 cm when still; all checks in the package; under 940 MB; cue tests pass | 2.5 h |
| 8 | fresh Opus | STANDARDS review from sheets and numbers, no MetaHuman frames | sign-off | 30 min |
Critical path 1-3-4-5-7, about 8 h 15 min (laptop estimates have run 2x low, batches 2-3x high). Opus on 1, 5, 7, 8; Sonnet on the rest.
4. Risks and the fallback
- If Kimodo fails a move, key it in Blender. If the round trip fails, do the IK in UE with armfix's solver.
- If the cue move breaks tests, keep 25.3 and use a harder bridge out of the pull-back. Under a 60° turn the riser goes.
- A contact that still misses gets sold like a film fight: shoot along the punch line from behind the striker's shoulder (within 15° of the lens, 50-85 mm), cut to the reaction close-up on the impact frame, and put the hit sound on that frame. Check: in the camera's projection the fist overlaps the face and is in front of it, with at most 40 cm of hidden depth.
B05, B18, B08 and B11 wait.
SEAMS.md
Source game/takes/SEAMS.md. Last git commit 2026-11-13. About 8,913 words.
Seam audit: every branch entry and exit (flawless branches)
Written Thu Nov 12 2026, night, for Don's top-priority ticket (BACKLOG.md "flawless branches", rule in lab/STANDARDS.md, Game director). Offline: no editor, no compile, take JSONs untouched. Numbers come from game/motion/seam_audit.py (re-run it after any take change): the three take JSONs, the director build logs in D:\megajam\builds\logs\director-build-*.log (ALIGNED, FLUB and PATH lines), and the source Kimodo npz from tower 1. The full5 bot runs have no PATH or BOUNDARY lines (those are written by the build, not the game), and full6 has no runs yet, so the build logs are the evidence.
Verdict
15 seams: 9 flub branches and 6 wreck resets (the 4 applause sections are not branches). Against the new rule:
- Root easing: 8 of the 9 flubs bring the fail clip onto the main path with a root ramp over 10 cm (up to 392 cm), which is exactly the "root sliding to rejoin a path" the rule forbids. Only early_sell (3 cm) and stage_fright (11 cm) are near zero.
- Pose gap at the crossfade: every flub entry and exit is crossfaded over blend_s = 0.25 s (7.5 frames). A pose gap of G cm max becomes about G / 7.5 cm per frame, so the max gap must stay under about 37 cm to meet the 5 cm target. 15 of the 18 flub seams are over it (worst: bad_landing entry 187 cm, belly_flop entry 164 and exit 147, rope_miss exit 123). Counting the real crossfade (which also carries the clips' own motion) the largest step per frame is over 5 cm at 17 of 18.
- Branch length: three fail clips are cut short because the main path rejoins too soon (rope_miss plays 38% of its clip, mirror_dizzy and stage_fright 52%). A clip cut at half way cannot end in the main pose.
- Wrecks: the reset shot is already a camera cut, but the action does not match across it: pull-backs of 2 to 108 cm at the cut in, teleports of 3 to 343 cm at the rejoin, pose gaps up to 182 cm, and the entry heading is off by up to 56 degrees (the reset starts at its own heading).
- Commit cutoff: every flub judges at cue + 0.432 s and starts leaving the main clip at cue + 0.533 s (16 frames), so the commit leads the branch motion by 0.10 s and the whole 0.25 s crossfade is after the commit. That holds, with little to spare; log
commit = cue_t + after_msper cue as the ticket asks. - Cues swallowed by a branch: early_sell swallows b2_uppercut, short_slide swallows b5_chair, belly_flop swallows c3_tuck and c4_land, and mirror_spin leaves 0.15 s before c4_leap. They need an explicit "scored as missed" rule.
- Main-path joins (not branches, flagged because they break the 5 cm rule too): the build's raw frame step in the blend is 28.3 cm at cannon 29.3 s (the launch), 24.6 cm at cannon 32.22 s (the landing), 13.9 cm at cannon 16.0 s (platform steps), 12.2 cm at chase 4.6 s and 10.0 cm at chase 30.8 s (tailgate leap). The baked hard-switch gap is 0.00 cm at every main join, so these are motion speed, not pops. Brawl is clean (2.1 cm max).
How it is measured
- A flub branch: the gate jumps into a parallel timeline at J = cue + 16 frames, the main clip A crossfades into the fail clip F over blend_s (0.25 s), F plays at the rate the build picked (log "fall rate", 0.64 to 1.25) and crossfades into the next main clip B, the gate returns at rejoin_t. Entry gap = pose of A at J against F frame 0; exit gap = F at the start of the B crossfade against B's first frame (with its trim).
- Pose gap: 24 body joints (no fingers), pelvis at the origin, hip line turned to a common heading, centimetres of the source rig. "mean / max" is over joints. "(x/fr)" is the max joint step per frame inside a linear crossfade of the two clips, the quantity the ticket wants under 5 cm. The rig scale on the MetaHuman is close to 1, so the numbers hold in the build.
- Root: from the build's FLUB line: "ramp onto the next clip R cm XY, Z cm height over T s of the fail clip". The ramp is an S curve over T - delay clip seconds, so the peak speed is 1.5 x R / (that window at the playback rate).
- Exit heading: how far the fail clip's end heading is from where clip B's baked heading is (the root ramp does not turn the clip, the crossfade has to).
- A wreck is a hard cut: at the judge time (cue + after_ms 432) the gate jumps to the reset section and the reset camera. Entry gap = main clip at the judge time against reset frame 0; pull-back = distance the main path moved between the cue time (where the reset is anchored,
at_t) and the judge time. Exit gap = reset's last frame against the main clip at rejoin_t; teleport = main-path advance from cue to rejoin minus the reset's own travel (collinear estimate from the 0.5 s PATH samples, so rough). - Severity = root cm (eased, or pull-back plus teleport) + worst pose gap cm + exit heading degrees + half the height ramp. It only orders the table.
Seams, worst first
Gap columns are mean / max joint distance in cm, with the per-frame step in brackets.
| # | seam | clips (main A, branch, main B) | entry gap | exit gap | root | heading | severity | fix |
|---|---|---|---|---|---|---|---|---|
| 1 | chase gate_bump (wreck, cue c2_slide at 9.60 s) | take_chase_split_b_s53 to gate_too_slow_s6 to take_chase_split_b_s53 | 61 / 182 (cut) | 17 / 56 (cut) | pull-back 108 cm at the cut in; 343 cm teleport at the rejoin | 1 at entry | 634 | camera cut + new recovery clip |
| 2 | chase mirror_spin (flub, cue c3_mirror at 13.40 s) | mirror_stumble_s2 to mirror_bounce_on_s11 to take_chase_split_c_s36 | 16 / 46 (6.1/fr) | 20 / 67 (8.9/fr) | 392 cm in 1.7 s, peak 353 cm/s; height -9 cm | 132 | 596 | root-aligned branch (new clip) |
| 3 | chase mirror_dizzy (flub, cue m1_mirror at 28.80 s) | mirror_stumble_s11 to mirror_dizzy_s1 to tailgate_leap_s3 | 15 / 68 (9.1/fr) | 16 / 46 (6.1/fr) | 372 cm in 1.1 s, peak 524 cm/s; height +5 cm | 85 | 527 | root-aligned branch (new clip) |
| 4 | cannon belly_flop (flub, cue c2_fire at 29.45 s) | cannon_launch_flight_s11 to cannon_belly_flop_s25 to bow_to_crew_s13 | 60 / 164 (21.8/fr) | 70 / 147 (19.6/fr) | 169 cm in 2.2 s, peak 115 cm/s; height +86 cm | 39 | 414 | shared transition pose + recovery clip |
| 5 | brawl short_slide (flub, cue b4_table at 13.18 s) | brawl_beat_c_s7 to table_short_roll_s17 to dust_off_s1 | 25 / 69 (9.1/fr) | 20 / 67 (8.9/fr) | 251 cm in 3.5 s, peak 108 cm/s; height +5 cm | 7 | 329 | root-aligned branch (new clip) |
| 6 | chase vault_clip (flub, cue c1_vault at 5.20 s) | take_chase_split_a_s79 to vault_tumble_roll_on_s33 to take_chase_split_b_s53 | 16 / 58 (7.7/fr) | 20 / 58 (7.8/fr) | 183 cm in 2.9 s, peak 96 cm/s; height +21 cm | 13 | 265 | root-aligned branch (new clip) |
| 7 | brawl rope_miss (flub, cue r1_swing at 23.15 s) | rope_swing_kick_s1 to flub_late_jump_s25 to brawl_uppercut_s3 | 7 / 33 (4.4/fr) | 59 / 123 (16.4/fr) | 87 cm in 0.5 s, peak 251 cm/s; height +50 cm | 12 | 247 | longer branch + recovery clip |
| 8 | brawl bad_landing (flub, cue s2_fall at 30.25 s) | bandstand_fall_s1 to limp_off_s2 to bow_to_crew_s13 | 87 / 187 (25.0/fr) | 11 / 41 (5.4/fr) | 26 cm in 3.8 s, peak 10 cm/s; height +1 cm | 4 | 218 | shared transition pose + recovery clip |
| 9 | chase tailgate_miss (wreck, cue c4_leap at 16.40 s) | take_chase_split_c_s36 to tailgate_miss_s5 to take_chase_split_c_s36 | 23 / 58 (cut) | 43 / 139 (cut) | pull-back 66 cm at the cut in; 11 cm teleport at the rejoin | 56 at entry | 215 | camera cut + recovery clip |
| 10 | chase tailgate_hang (wreck, cue t1_leap at 32.55 s) | tailgate_leap_s3 to truck_hang_on_s37 to dust_off_s1 | 42 / 91 (cut) | 22 / 83 (cut) | pull-back 50 cm at the cut in; 62 cm teleport at the rejoin | 10 at entry | 204 | camera cut + recovery clip |
| 11 | cannon land_slip (wreck, cue c4_land at 33.02 s) | cannon_tuck_land_s20 to cannon_land_slip_hang_s36 to bow_to_crew_s13 | 20 / 52 (cut) | 14 / 45 (cut) | pull-back 49 cm at the cut in; 71 cm teleport at the rejoin | 0 at entry | 172 | camera cut |
| 12 | brawl eat_the_chair (wreck, cue b3_sell at 9.43 s) | brawl_beat_b_s1 to chair_eat_stagger_s6 to brawl_beat_b_s1 | 15 / 72 (cut) | 18 / 70 (cut) | pull-back 21 cm at the cut in; 4 cm teleport at the rejoin | 17 at entry | 96 | camera cut |
| 13 | cannon stage_fright (flub, cue k1_wave at 9.20 s) | cannon_wave_crowd_s31 to mark_check_nervous_s6 to high_five_crew_s2 | 20 / 80 (10.7/fr) | 21 / 56 (7.4/fr) | 11 cm in 1.3 s, peak 13 cm/s; height -1 cm | 3 | 95 | blend only (retime) |
| 14 | brawl early_sell (flub, cue b1_duck at 5.14 s) | brawl_beat_a_s17 to haymaker_early_sell_s2 to brawl_beat_b_s1 | 10 / 28 (3.8/fr) | 7 / 25 (3.4/fr) | 3 cm in 2.6 s, peak 2 cm/s; height -2 cm | 4 | 37 | blend only |
| 15 | brawl chair_whiff (wreck, cue b5_chair at 15.86 s) | brawl_beat_c_s7 to chair_whiff_s7 to brawl_beat_c_s7 | 7 / 18 (cut) | 4 / 9 (cut) | pull-back 2 cm at the cut in; 3 cm teleport at the rejoin | 18 at entry | 23 | none |
The fix per seam
- chase gate_bump (camera cut + new recovery clip). Wreck. Keep the reset shot's camera cut and add Walt's "Cut!" at it, but make the action match across it. Re-anchor the reset at the judge time (cue + 0.43 s) instead of the cue time: that removes the 108 cm pull-back. Two new clips, chained: slide_stop_lying (starts in the slide, stops) then getup_run_back (gets up and runs on); set rejoin_t so the main path is where that clip ends (about 3.5 m on). Today gate_too_slow ends 6 cm from where it began and the main path is 349 cm ahead.
- chase mirror_spin (root-aligned branch (new clip)). Replace mirror_bounce_on-s11, which stays near the mirror and is then dragged 392 cm in 1.7 s (3.5 m/s at the peak) while its hips turn 132 degrees against the next clip. stumble_recover_run bounces off, recovers and keeps running about 4.3 m, ends heading with the run, so the ramp goes to 0 and the exit turn is small. The 2.3 s branch fits a 3 s clip at rate 1.25. Blend end 16.25 s leaves 0.15 s before cue c4_leap (16.4 s): fits.
- chase mirror_dizzy (root-aligned branch (new clip)). Replace mirror_dizzy-s1, which spins in place, plays only 52% of its length (1.7 s branch, 3.5 s clip) and is then dragged 372 cm in 1.1 s (5.2 m/s peak) and turned 85 degrees. dizzy_stagger_run spins while staggering forward 3.7 m and sprints into the run-up of tailgate_leap. The branch only holds about 2.15 s of clip, so the spin has to come early; seam_rank reads the exit pose at 2.15 s. Cue t1_leap is 1.5 s past the blend: fits.
- cannon belly_flop (shared transition pose + recovery clip). Entry: the main flight is high in the air (4.3 m), the flop clip starts 164 cm away in joint space (40 cm height bridge): flight_belly_flop_lie starts in the flight and lands flat. Exit: the flop ends flat, the bow starts standing (147 cm); getup_dazed_prone, chained after it, ends on his feet. The build's arc carries the root, so no travel target. The branch swallows cues c3_tuck and c4_land; log that and score them as missed.
- brawl short_slide (root-aligned branch (new clip)). table_short_roll drags 251 cm of root over 3.5 s because the main path barely moves here (0.3 m) while the clip slides 2.5 m. table_roll_off_getup slides short, rolls off and gets up standing for dust_off (exit gap 67 cm today). Cue b5_chair (15.86 s) falls inside the branch, which ends at 17.97 s; it is swallowed, so commit its outcome as a miss or move the rejoin.
- chase vault_clip (root-aligned branch (new clip)). 183 cm of ramp over 2.85 s (96 cm/s) and 58 cm entry and exit gaps. vault_trip_roll_run trips, rolls on the ground and runs on about 7.6 m into take_chase_split_b. Cheaper first try: start the existing vault_tumble_roll_on 1.2 s in (trim), which lowers the entry gap from 16 to 4 cm mean and 58 to 14 cm max. Next cue c2_slide is 0.35 s past the blend: tight but fits.
- brawl rope_miss (longer branch + recovery clip). The branch is only 1.17 s long and the 3 s fall plays 38% of itself, so the fall is cut off in mid-air and the hips are 50 cm too low against the next clip (123 cm exit gap, 87 cm ramp in 0.5 s, 2.5 m/s peak). Rejoin one clip later (skip_clips 1, bandstand_fall at 28.85 s, a 5.2 s branch) and append getup_fight_ready (the real get-up from flat on the chest that the ticket asks for). The branch swallows cue s1_upper. Needs plan_flubs to take a list of branch clips (A, fall, get-up, B) or the two clips baked into one.
- brawl bad_landing (shared transition pose + recovery clip). Entry is the worst of any flub (187 cm max, 25 cm per frame of crossfade, an 88 cm height bridge): the main clip is flat on his back at the cue, limp_off starts standing. The existing dust_off-s3 (gag 40 of 40) starts flat on his back and ends standing: entry 33 / 118 and exit 6 / 20 against 87 / 187 and 11 / 41 today. The limp is lost, the pose is right; the new back_getup_limp_off and getup_clutch_back did not work (0 of 30 each). Exit is fine (41 cm) and the root ramp is only 26 cm.
- chase tailgate_miss (camera cut + recovery clip). Wreck. The reset ends flat on the ground (hips 14 cm) and the cut lands on a running main clip (139 cm gap). getup_run_back after tailgate_miss; the heading also jumps 56 degrees at the entry cut (the reset starts at its own heading, the main clip is turned 56 degrees), so turn the reset clip onto the main heading.
- chase tailgate_hang (camera cut + recovery clip). Wreck. The reset ends hanging 122 cm up, the main clip is standing on the bed (83 cm exit gap). hang_pull_up_stand ends standing on the platform. Pull-back at the entry cut is 50 cm.
- cannon land_slip (camera cut). Wreck. 52 cm entry and 45 cm exit gaps, 49 cm pull-back, 71 cm teleport. Keep the camera cut and Walt's "Cut!"; re-anchor at the judge time. No new motion needed if the cut is clean.
- brawl eat_the_chair (camera cut). Wreck. The stagger travels 74 cm and the main path 71 cm, so the exit is a 4 cm step, but the pose gap is 70 cm max and the heading is turned 17 degrees at the entry cut. Cut, and turn the reset onto the main heading.
- cannon stage_fright (blend only (retime)). Small root ramp (11 cm over 1.3 s) but 80 cm max entry and 56 cm exit pose gaps with 11 cm per frame of crossfade; 52% of the 4 s clip plays. Re-trimming does not help (mean 20 to 18 cm). A 0.4 s crossfade takes the step from 10.7 to about 6.7 cm per frame; check it by eye.
- brawl early_sell (blend only). Root ramp 3 cm, pose gaps 28 and 25 cm, crossfade 3.8 cm per frame at the entry and 3.4 at the exit: already inside the 5 cm per frame target. The branch swallows cue b2_uppercut.
- brawl chair_whiff (none). Wreck. 18 cm entry and 9 cm exit gaps, 2 and 3 cm root steps. Already clean.
What the laptop ticket needs from the engine side
plan_flubstakes one fail clip (chain A, F, B). The get-up and recovery fixes need a list of branch clips (A, F, R, B), or F and R baked into one asset.- Reset anchoring: use the judge time (
cue + after_ms) forat_t, and pickrejoin_tfrom where the recovery clip ends on the main path. - Turn the reset clip onto the main heading (today its yaw is 0 against a main clip turned up to 56 degrees).
- Drop the root ramp (set
rampto 0) once the replacement clips are in; keep the check that logs it. - Crossfade 0.25 to 0.4 s at each branch seam; measure the max joint step per frame in the packaged build, not only the source clips.
- Walt's "Cut!" at each wreck cut (his shout_cut clip exists).
Batch 16 and 16b (recovery motion, tower 1)
Jobs mj-kimodo-seams-16 (seq 589, 300 takes, 1246 s) and mj-kimodo-seams-16b (seq 591, 150 takes, one phase per prompt, run after the SFX bake-off released the GPU lock), prompts in game/motion/kimodo_seams16.py, Kimodo-SOMA-RP-v1.1, 30 seeds each, 0 failed generations. Each prompt starts in the pose the seam leaves the actor in and ends on his feet in the main path's direction. Gag rules are in gag_checks.py; game/motion/seam_rank.py ranks each move against the real main-path clips named in game/takes/seam_specs.json. Score = entry mean cm + exit mean cm + 25 x travel error in m (lower is better), with the exit read at the last frame the branch can play (play_s). Previews of the top three per move: ~/studio-ai/renders/megajam/mj-kimodo-seams-16/previews/ and mj-kimodo-seams-16b/previews/ on tower 1.
| seam | best tag (chain) | gag | entry mean / max cm | exit mean / max cm | travel m (target) | score |
|---|---|---|---|---|---|---|
| chase mirror_dizzy | dizzy_stagger_run-s1 | pass (14/30) | 15 / 68 | 12 / 44 | 4.1 (3.7) | 36 (next s9: 47, but turns 119 deg) |
| cannon belly_flop | flight_belly_flop_lie-s9 then getup_dazed_prone-s18 | pass, pass | 16 / 52 | join 5 / 13, exit 5 / 20 | 0.6 (0.5) | 26 (today 60 / 164 in, 70 / 147 out) |
| brawl rope_miss | getup_fight_ready-s16 (after flub_late_jump) | pass (30/30) | 12 / 36 | 2 / 7 | 0.5 (0.5) | 14 (next s25, s3) |
| brawl short_slide | table_roll_off_getup-s23 | pass (6/30) | 29 / 54 | 5 / 19 | 0.4 (0.5) | 36 (next s6: 38) |
| chase tailgate_hang | hang_pull_up_stand-s21 | pass (1/30) | 27 / 69 | 5 / 17 | 0.8 (0.5) | 40 (s26: 25, fails the hang test by a hair) |
| brawl bad_landing | dust_off-s3 (existing, redo-13) | pass (40/40) | 33 / 118 | 6 / 20 | 0.7 (0.6) | 39 (today 87 / 187 in) |
| chase gate_bump | slide_stop_lying-s30 then getup_run_back-s16 | pass, pass | 37 / 123 | join 10 / 26, exit 10 / 33 | 1.8 + run-out | 56 (today 61 / 182 in, 17 / 56 out) |
| chase tailgate_miss | getup_run_back-s5 (after tailgate_miss) | pass (8/30) | 46 / 112 | 17 / 51 | 3.6 (3.0) | 78 (s10: 89) |
| chase mirror_spin | stumble_recover_run-s26 | pass (9/30) | 9 / 32 | 18 / 82 | 6.6 (4.3) | 85 (s2: 95) |
| chase vault_clip | vault_trip_roll_run-s23 | pass (2/30) | 18 / 61 | 23 / 58 | 9.4 (7.6) | 85 (round 2 vault_dive_roll_run-s13: 95) |
The two chained rows are the best pair by total score (entry + join + exit mean) among gag-passing takes, 30 x 26 pairs for the belly flop and 3 x 8 for the gate. The join is where one new clip hands to the next; 5 / 13 cm and 10 / 26 cm are inside the 37 cm limit for a 5 cm per frame crossfade.
What the numbers say:
- Kimodo follows one phase per prompt. Round 1 prompts that chained phases (slide, bump, get up, run; fly, flop, get up; lie, get up, limp) mostly came out in the wrong order: slide_bump_getup_run never went low (0 of 30), flight_belly_flop_getup and back_getup_limp_off started standing and ended lying. Round 2 split them into one phase per clip, and the pairs join to within 3 to 10 cm mean. The single-phase get-ups worked from the start: getup_fight_ready 30/30, getup_run_back 23/30 start low.
- getup_clutch_back (0 of 30) and slide_bump_getup_run (0 of 30) found nothing usable; they stay in the file as the record.
- Travel overshoots the main path in the running clips (6.6 m against 4.3, 9.4 against 7.6): set the clip's
travelscale in the build (0.65 and 0.8), as mirror_spin already does with 0.3. That keeps the ramp at 0 and the pose untouched. - Exit gap max stays 44 to 82 cm in the run-outs because the last played frame is mid-stride against the start of the next run clip. Start the next main clip on the best-matching stride frame (the audit's best-trim search finds it) instead of at 0.
- gate_bump entry is still 123 cm max: the cut lands mid-slide against a clip that starts lying. It is a camera cut, so match the action by choosing the cut frame (best-trim search) rather than by blending.
- Not covered by a new clip: land_slip, eat_the_chair, chair_whiff, early_sell, stage_fright. They need the crossfade and cut changes in the list above, no new motion.
Batch 20: constrained branches
Jobs mj-kimodo-branches-20 (flubs, 400 takes) and mj-kimodo-branches-20b (wrecks, 240 takes), tower 1, 0 failed generations. Code: game/motion/kimodo_branches20.py (plan from the take JSONs, constraints, seam numbers, gag) and rank_branches20.py; results and ranking.md in ~/studio-ai/renders/megajam/mj-kimodo-branches-20 and ...-20b. Each clip starts on the main path pose at the commit (flub: film_t; wreck: cue + 0.432 s) and ends on the main path pose at rejoin_t, full-body constraints on the first and last 3 frames, text guidance 4.0, 40 seeds. Rank: gag, both seam targets, joint-limit violations, whole-clip peak joint acceleration. Seam gaps are 0.00 to 0.60 cm at every pick, pelvis offset 0.0, velocity mismatch 0 percent. Clip length is whole frames, so the corrected rejoin_t is film_t + (frames - 1)/30; the clip then plays at rate 1.0 (drop rate_min, the root ramp and the fail-clip retime).
| seam | pick | backups | gag pass | seam gaps start / end (max joint) | joint-limit violations | accel clip new / source (cm/frame^2) | rejoin_t -> corrected (frames) |
|---|---|---|---|---|---|---|---|
| early_sell (flub) | early_sell-s28 | s24, s3, s25 | 22/40 | 0.01 / 0.02 cm, vel 0.0% / 0.0% | 1 | 10.03 / 3.34 | 9.0967 -> 9.1067 (104 fr) |
| short_slide (flub) | short_slide-s14 | s20, s11, s1 | 20/40 | 0.09 / 0.02 cm, vel 0.0% / 0.0% | 2 | 5.09 / 3.84 | 17.9667 -> 17.98 (129 fr) |
| rope_miss (flub) | none (closest rope_miss-s17, fails fall) | 0/40 | 24.9667 -> 24.9833 (40 fr) | ||||
| rope_miss_long (flub) | rope_miss_long-s33 | s22, s11, s26 | 40/40 | 0.26 / 0.0 cm, vel 0.0% / 0.0% | 44 | 21.21 / 13.85 | 29.2167 -> 29.2167 (167 fr) |
| bad_landing (flub) | bad_landing-s22 | s16, s28, s32 | 40/40 | 0.0 / 0.0 cm, vel 0.0% / 0.0% | 31 | 13.29 / 4.13 | 35.3667 -> 35.3833 (139 fr) |
| belly_flop (flub) | belly_flop-s17 | s33, s30, s7 | 40/40 | 0.02 / 0.01 cm, vel 0.0% / 0.0% | 428 | 14.06 / 0.51 | 34.7 -> 34.7167 (143 fr) |
| stage_fright (flub) | stage_fright-s1 | 1/40 | 0.0 / 0.0 cm, vel 0.0% / 0.0% | 17 | 10.54 / 8.78 | 11.7667 -> 11.7667 (62 fr) | |
| vault_clip (flub) | vault_clip-s29 (ranker put s23 first, but its whole-clip peak is 130 against 17) | s23, s36, s4 | 12/40 | 0.0 / 0.12 cm, vel 0.0% / 0.0% | 62 | 27.94 / 17.07 | 9.3667 -> 9.3667 (110 fr) |
| mirror_spin (flub) | mirror_spin-s24 | s38, s3, s4 | 40/40 | 0.0 / 0.0 cm, vel 0.0% / 0.0% | 0 | 7.05 / 8.21 | 16.3667 -> 16.3667 (74 fr) |
| mirror_dizzy (flub) | mirror_dizzy-s23 | s26, s5, s40 | 15/40 | 0.0 / 0.0 cm, vel 0.0% / 0.0% | 0 | 15.2 / 9.93 | 31.1667 -> 31.1667 (56 fr) |
| gate_bump (wreck) | gate_bump-s19 | s9, s15, s6 | 40/40 | 0.6 / 0.06 cm, vel 0.0% / 0.0% | 0 | 12.42 / 6.85 | 12.0 -> 12.0 (75 fr) |
| tailgate_miss (wreck) | tailgate_miss-s1 | s15, s38, s18 | 37/40 | 0.06 / 0.51 cm, vel 0.0% / 0.0% | 1 | 13.67 / 8.46 | 16.95 -> 16.95 (90 fr) |
| tailgate_hang (wreck) | tailgate_hang-s4 | s24, s19, s22 | 37/40 | 0.09 / 0.03 cm, vel 0.0% / 0.0% | 0 | 5.77 / 4.09 | 34.0 -> 34.0 (105 fr) |
| land_slip (wreck) | land_slip-s5 | s38, s25, s19 | 40/40 | 0.0 / 0.0 cm, vel 0.0% / 0.0% | 24 | 13.88 / 7.94 | 34.7 -> 34.7 (120 fr) |
| eat_the_chair (wreck) | eat_the_chair-s16 | s35, s26, s39 | 40/40 | 0.01 / 0.0 cm, vel 0.0% / 0.0% | 0 | 5.27 / 3.85 | 11.5 -> 11.5 (75 fr) |
| chair_whiff (wreck) | chair_whiff-s25 | s8, s1, s30 | 27/40 | 0.0 / 0.0 cm, vel 0.0% / 0.0% | 0 | 2.69 / 0.33 | 17.0 -> 17.0 (90 fr) |
Notes for the laptop ticket:
- Put the corrected rejoin_t from the last column in the take JSONs, set the branch clip rate to 1.0, set
rampto 0, and import the picks from the tower folders above (npz and bvh share the tag). Wrecks: the clip is as long as the reset (reset_ms), and the reset starts at the main path pose at the judge time (cue + after_ms), so re-anchor the reset there and cut at that frame. - rope_miss: the existing rejoin (24.97, 1.3 s) is too short to fall; 0 of 40 fell. Use rope_miss_long-s33, which rejoins one main clip later at 29.2167 (skip_clips 1, the start of bandstand_fall plus the take's 0.37 s lead). It swallows cue s1_upper (score it as missed).
- stage_fright is weak: 1 of 40 passed the nervous-hands gag (the 2 s branch is short). vault_clip passes 12 of 40 and the picks rise 2.4 m at the rejoin (the main clip is mid-leap there); look at it by eye. mirror_dizzy 15 of 40.
- belly_flop: every take has 400+ joint-limit violations, so the numbers say the flight start pose itself is out of range; the pick is by gag and seams only and needs a look. The seam is in the clip frame: the build's arc drives the root, so the end pose was placed 0.5 m ahead of the start pose and the arc must be turned off or reduced to the clip's own travel after commit.
- The main path is rebuilt offline from continuity of the take's clips and ignores the engine's slide, lift, arc, tilt and z_w keys, so check the seams in a build. land_slip, tailgate_hang and tailgate_miss start from a main clip that is already low or in the air: tailgate_miss rejoins 0.12 s of film after the judge time, so the reset clip ends about 0.3 m from where it began.
- None of the picks has been viewed by eye. No previews were rendered.
Brawl rebuild: seams measured in the packaged game (Fri Nov 13, build D:\megajam\builds\brawl2)
The rebuilt brawl (game/takes/brawl2_take.py) was packaged and played by the perfect, late and none bots with -KKPelvisLog; the game now logs 18 joints of Kid and the heavy and the prop bones every tick (KKJ lines, about 83 a second). game/takes/brawl2/seam_check.py reads them. Nobody looked at a frame.
How to read it. "jump" is the largest unexpected move of any joint over one 30 fps frame (the move over the next three ticks minus the move over the last three). "usual" is the 99th percentile of the same number away from every seam. It is high (19 to 25 cm) because a body's logged pose is sometimes one tick stale, which a fast hand turns into a false step; the median is 1 cm. So a seam at or under "usual" is not shown to be a pop, and the 5 cm target cannot be confirmed below about 10 cm with this log. The build's own pose-fit numbers (CONSTRAINED and RESET lines) are exact and are given for the branches.
| run | seam | jump cm | joint | usual | root step cm/s | usual | under 5 cm |
|---|---|---|---|---|---|---|---|
| perfect | kid main: into brawl_beat_a_s17 | 34.83 | hand_l | 23.46 | 73.3 | 128.3 | NO |
| perfect | kid main: into brawl_beat_b_s1 | 13.03 | hand_l | 23.46 | 10.6 | 128.3 | NO |
| perfect | kid main: into brawl_beat_c_s7 | 32.98 | hand_l | 23.46 | 3.1 | 128.3 | NO |
| perfect | kid main: into dust_off_s1 | 8.83 | foot_l | 23.46 | 2.5 | 128.3 | NO |
| perfect | kid main: into vr_kid_rope_swing_kick | 17.41 | calf_r | 23.46 | 44.8 | 128.3 | NO |
| perfect | kid main: into brawl_uppercut_s3 | 4.42 | calf_r | 23.46 | 25.3 | 128.3 | yes |
| perfect | kid main: into bandstand_fall_s1 | 33.29 | hand_r | 23.46 | 9.2 | 128.3 | NO |
| perfect | kid main: into dust_off_s18 | 73.86 | hand_l | 23.46 | 20.1 | 128.3 | NO |
| perfect | kid main: into bow_to_crew_s13 | 11.55 | hand_l | 23.46 | 3.1 | 128.3 | NO |
| perfect | kid main: into dust_off_s1 | 14.93 | hand_r | 23.46 | 4.6 | 128.3 | NO |
| perfect | heavy main: into mark_check_nervous_s6 | 9.12 | hand_l | 19.25 | 47.7 | 82.2 | NO |
| perfect | heavy main: into heavy_wild_swing_miss_s12 | 17.11 | hand_r | 19.25 | 25.8 | 82.2 | NO |
| perfect | heavy main: into haymaker_take_s13 | 13.96 | lowerarm_r | 19.25 | 43.5 | 82.2 | NO |
| perfect | heavy main: into heavy_haymaker_throw_s10 | 13.61 | ball_r | 19.25 | 5.6 | 82.2 | NO |
| perfect | heavy main: into walk_to_mark_s2 | 12.48 | ball_l | 19.25 | 33.6 | 82.2 | NO |
| perfect | heavy main: into brawl_shove_chest_s3 | 33.12 | hand_r | 19.25 | 14.1 | 82.2 | NO |
| perfect | heavy main: into cs_heavy_chair_swing_2h | 63.2 | hand_r | 19.25 | 14.0 | 82.2 | NO |
| perfect | heavy main: into mark_check_nervous_s6 | 26.19 | ball_l | 19.25 | 3.4 | 82.2 | NO |
| perfect | heavy main: into walk_to_mark_s2_rev | 34.79 | lowerarm_r | 19.25 | 52.4 | 82.2 | NO |
| perfect | heavy main: into walk_to_mark_s2_rev | 18.28 | ball_l | 19.25 | 136.9 | 82.2 | NO |
| perfect | heavy main: into walk_to_mark_s2_rev | 18.67 | ball_r | 19.25 | 133.8 | 82.2 | NO |
| perfect | heavy main: into walk_to_mark_s2_rev | 19.66 | ball_r | 19.25 | 138.5 | 82.2 | NO |
| perfect | heavy main: into walk_to_mark_s2_rev | 25.07 | ball_l | 19.25 | 147.0 | 82.2 | NO |
| perfect | heavy main: into mark_check_nervous_s6 | 25.53 | hand_r | 19.25 | 137.5 | 82.2 | NO |
| perfect | heavy main: into vr_heavy_pullback | 25.53 | hand_r | 19.25 | 173.9 | 82.2 | NO |
| perfect | heavy main: into heavy_wild_swing_miss_s12 | 36.71 | hand_l | 19.25 | 52.6 | 82.2 | NO |
| perfect | heavy main: into walk_to_mark_s2 | 14.95 | ball_l | 19.25 | 36.6 | 82.2 | NO |
| perfect | heavy main: into heavy_haymaker_throw_s10 | 5.95 | hand_l | 19.25 | 8.0 | 82.2 | NO |
| perfect | heavy main: into dust_off_s1 | 58.07 | hand_l | 19.25 | 1.7 | 82.2 | NO |
| perfect | heavy main: into mark_check_nervous_s6 | 15.18 | hand_l | 19.25 | 7.6 | 82.2 | NO |
| perfect | heavy main: into dust_off_s1 | 21.21 | lowerarm_r | 19.25 | 5.5 | 82.2 | NO |
| late | flub early_sell: in at film 5.673 | 8.77 | lowerarm_r | 24.26 | 43.3 | 131.0 | NO |
| late | flub early_sell: back at film 9.357 | 9.8 | hand_l | 24.26 | 2000.8 | 131.0 | NO |
| late | flub short_slide: in at film 13.713 | 4.6 | hand_l | 24.26 | 41.8 | 131.0 | yes |
| late | flub short_slide: back at film 17.980 | 2.39 | hand_l | 24.26 | 708.2 | 131.0 | yes |
| late | flub rope_miss: in at film 23.683 | 53.58 | hand_r | 24.26 | 207.5 | 131.0 | NO |
| late | flub rope_miss: back at film 29.217 | 9.83 | hand_l | 24.26 | 586.7 | 131.0 | NO |
| late | flub bad_landing: in at film 30.783 | 6.61 | ball_l | 24.26 | 51.3 | 131.0 | NO |
| late | flub bad_landing: back at film 35.633 | 6.44 | hand_r | 24.26 | 490.8 | 131.0 | NO |
| late | wreck eat_the_chair: cut in (main film 9.956) | 7.35 | ball_l | 24.26 | 51.2 | 131.0 | NO |
| late | wreck eat_the_chair: cut back (to film 11.523) | 1.33 | lowerarm_r | 24.26 | 127.7 | 131.0 | yes |
| none | wreck eat_the_chair: cut in (main film 10.035) | 15.75 | lowerarm_r | 25.18 | 49.9 | 126.0 | NO |
chair_whiff has no row: a missed b4 swallows b5, so no bot reaches it. Its build numbers: pose gap 0.01 cm at the cut in, 0.70 cm at the cut back.
Branches, build numbers (pose gap after the fit, max joint):
| branch | in | out | left at the rejoin, spread over the branch |
|---|---|---|---|
| early_sell-s28 | 17.4 cm (head), 0.25 s crossfade | 1.1 cm after a 0.25 s crossfade, heading off 9 deg | pelvis 13.8 cm over 3.4 s (0.20 cm per frame at most) |
| short_slide-s14 | 1.5 cm, hard switch | 0.8 cm, hard switch | pelvis 4.1 cm over 4.2 s |
| rope_miss_long-s33 | 124.6 cm (hand), pelvis 26 cm off and 22 cm low, 0.4 s crossfade | 2.4 cm, hard switch | none: locked on its exit |
| bad_landing-s22 | 0.0 cm, hard switch | 0.5 cm after a 0.25 s crossfade, heading off 10 deg | pelvis 7.1 cm over 4.6 s |
| eat_the_chair-s16 (wreck) | 0.13 cm at the judge time | 0.56 cm | 0.09 cm |
| chair_whiff-s25 (wreck) | 0.01 cm | 0.70 cm | 0.09 cm |
What the numbers say.
- The batch 20 clips whose both ends were made on clips still in the take are seamless: short_slide, bad_landing and both wrecks, and the out side of rope_miss. In the game they measure 1.3 to 9.8 cm, inside the log's noise.
- early_sell-s28 does not start on the main pose (17 cm at the head; the closest frame of brawl_beat_a is 8 cm off), so batch 20's 0.01 cm was measured on a different main pose than the build plays. It needs a rerun against the build's pose at 5.6733 s.
- rope_miss_long-s33 was made against rope_swing_kick_s1. The keyed swing replaced that clip, so the entry cannot match. It needs a rerun against the keyed clip, which first needs that clip in the SOMA skeleton, and the cue should move before the takeoff (the miss is judged after the kick has landed).
- A wreck is called 85 to 170 ms after the judge time the reset clip was made for (the late guard). The game now enters the reset that far in (
JumpToSectionAt), which took the cut back from a 33 ms flash of the rest pose to 1.3 cm; the cut in still measures 7 to 16 cm. - The main path is not seamless and was never going to be with these clips: joins between unrelated Kimodo clips are 0.25 s crossfades over pose gaps of 20 to 80 cm, which is 9 to 74 cm per frame here. Two things were hiding how bad it was. Ease durations were written in display frames where the engine wants ticks, so every crossfade in the packaged game was a 0.3 ms hard switch (fixed in director_build.py; chase and cannon get it at their next build). And a held section looped its clip instead of holding (fixed: a hold is its own section).
- Root speed: on the main path the pelvis speed steps more than the usual (82 cm/s for the heavy, 128 for Kid) only at the joins between the repeats of the heavy's backward walk and the idle after them (134 to 147 cm/s) and at the pull-back (174 cm/s, which is the yank itself). The branch rows over 400 cm/s are the timeline jump, not motion.
Contacts, perfect run.
- Rope kick: the ball of Kid's foot 24.7 cm from the heavy's chest joint at 23.40 s (about 20 cm is sole on chest).
- Rope grip, 38 ticks held: left wrist 10.44 to 14.06 cm from the grip joint, right 19.29 to 20.18 cm. The spread while held is 3.62 cm, over the 2 cm target, and the right wrist sits about 9 cm lower on the rope than authored. brawl2_armfix.py puts both wrists 0.0 cm from the authored points in the clip, so the difference is between the rope's bake and the body's 1.147 scale.
- Chair: both wrists 0.0 cm from the authored rail points from frame 13 to the hit, up to 3.2 cm short over the first 13 frames (he cannot quite reach the rail on the floor). Read in the clip (brawl2_armfix.py), not in the game: the chair is one bone, so the log cannot give the rail.
- Feet: Kid slides while planted in 11 of 12 planted runs, worst real ones 25 to 32 cm at 8.8 to 15 s, which is
travel0.5 on brawl_beat_b and brawl_beat_c (kept because the batch 20 branches were made on that path); the 64 cm at 31.2 s is him getting up from his back, not a plant. The heavy: 12.6 cm at the crossfade out of the chair swing, 7 cm at 31.5 s.
Why the bandstand is not under the fall. bandstand_fall_s1 throws Kid 1.2 m straight back from where he stands. He stands 0.4 m ahead of where the rope swing put him down, and the swing came in along that same line: its last 2.4 m of flight are over the ground the bandstand's 2.4 m deck would need, at foot height 0.3 to 1.3 m, against a deck at 0.59 m with rails to 1.41 m on all four sides and a roof at 2.58 m. He would swing through it one beat before falling onto it. The fall is therefore a flat back fall on the boards, which is what the clip is (pelvis 7 cm above the floor at rest), and the smashing sounds are gone. To land him on a deck the fight has to move 2.5 m on after the swing, or the fall has to turn him: new motion either way.
Cannon rebuild (T7, Fri Nov 13): every seam of the three takes in the packaged game
Package D:\megajam\builds\cannon2 (927 MB), bots perfect, late, none, human sd 45 and 75 on chase, brawl and cannon with -KKPelvisLog; python brawl2/seam_check.py D:\megajam\builds\cannon2\runs --take <chase|brawl|cannon> (now takes --take). Full tables: D:\megajam\builds\cannon2\seams_all.md, json per take in runs\seams-<take>.json. "jump" is the largest unexpected move of any joint per 30 fps frame (cm); "usual" is the 99th percentile of the same number away from seams in that run. Root steps of about 2000 cm/s at a branch entry or exit are the checker dividing across the timeline jump, not a pop (the pelvis trace through the belly flop entry is continuous).
Cannon, the new parts (perfect, late):
| seam | jump cm | usual | note |
|---|---|---|---|
| climb-in into the keyed clip (28.95 s, lying back in the bore) | 11.3 | 20.8 | the first build pitched him forward (137 cm): the climb-in's yaw is baked outside the tilt, so leaning back is -43.7 deg |
| wave s39 (turns him 180 deg while he waves) | 20.3 | 20.8 | |
| into the climb-in | 5.7 | 20.8 | |
| stage_fright-s1 in / back | 1.9 / 16.7 | 24.5 | back: the branch's own walk ends 92.6 cm ahead of the high five, spread over the 2 s branch (2.3 cm a frame) |
| belly_flop-s17 in (31.01 s, top of the arc) / back (35.85 s, the stand) | 35.0 / 15.6 | 24.5 | in: a 0.6 s crossfade from the tuck to s17's upright first frame (it was made on the old flight); 160 cm with 0.3 s |
| land_slip-s5 in (32.95 s, in the boxes) / back (36.91 s, in the wave) | 21.2 / 9.9 | 25.8 | from t7_landslip.ps1 (human sd 160, seed 19: no other bot misses c4 with c2 hit); 0.4 s crossfades both ways, the 68 cm pelvis gap spread over the branch |
The keyed clip read back from the game: gravity 1126.9 cm/s2 and 1491.6 cm/s across (expected 1125 and 1491 at Rex's 1.147), apex pelvis 1514 cm; at rest in the boxes the pelvis is at 229.5 cm (200 x 1.147), standing in the crater the balls of the feet are at 235 cm over the 231 cm crater floor. The packaged game plays the retargeted pelvis 12.8 cm under the editor's mesh-retargeted evaluation; the first build had his feet 9 cm into the mattress, so the keyed clip and its branches are lifted 12.8 cm. Planted foot slide in the cannon take is worst in the old main path (123 cm in the sprint to the platform, 12.5 to 14.3 s; 62 cm in the first wave, 4.6 to 11.5 s), not in the rebuilt part. Bots on the final package: chase 5/1/1/5/3 stars (perfect, late, none, human 45, human 75), brawl 5/1/1/4/3, cannon 5/1/1/4/3, 0 error lines in all 15 runs and in the slip runs; premiere loop 15 of 15 sections planned = played, 0 errors, slow motion in at film 29.636 s and out at 32.420 s. Chase and brawl main joins and wreck cuts are unchanged from the brawl rebuild's numbers (cuts are camera cuts, 125 to 370 cm by design); the chase now carries a joint log ('joint_log' in chase.take.json). Not in git: Content/Props/Cannon2/GC_stunt_rig.uasset (120.6 MB, the repo keeps large media out); rebuild it from D:\megajam\proofs\cannon\k_stunt_rig.abc with KidKatapult/Scripts/t7_import.py (T7_STAGES=rig,mats).
T7 follow-up (Fri Nov 13): the ladder climb and two branches, ready for the laptop (not yet in the take or the package)
Ladder climb, k_kid_ladder_climb (keyed on tower 2: game/proofs/scripts/lad_dump.py, lad_geom.py on the laptop, lad_climb.py; numbers in game/proofs/cannon/lad_result.json and lad_geom.json; mannequin sheets game/proofs/cannon/lad_contact_*.jpg). Files in D:\megajam\proofs\cannon\ladder\ (hashes in SHA256SUMS.txt, the FBX checked against tower 2): k_kid_ladder_climb.fbx (175 frames, 5.80 s, proof skeleton, root bone pelvis, retarget to MH-Rex as the other keyed clips), k_kid_ladder_climb.json, SM_cannon_platform_ladder_gap.glb. It replaces walk_to_mark_s2 at 16.0 s (its lift_keys and slide go): out of the sprint, two stopping steps round the near rail to the ladder foot, hands on the rails, both feet onto each rung in turn (the rungs stand 71 cm apart in the game, so a foot cannot skip one), hands onto the two guard posts at the top, right foot from rung 3 onto the deck, the left hangs while he stands up on the deck, then comes over the edge, then two side steps and a quarter turn onto the mark, ending exactly on the wave's first frame at 21.8 s.
- Frame 1 is chase_sprint-s7 at clip 3.2 s (sprint frame 97); frame 175 is cannon_wave_crowd-s39 frame 1 turned -5.753 deg (the layout's psi of the sprint minus the wave's) and lifted onto the 298.2 cm deck. The clip is keyed in the sprint's own retarget space, so the take places it with
"rigid": true, "blend": 0.0and"psi_abs": -5.753(the sprint's psi; the pose fit gives the same), no lift. The wave at 21.8 s keepsface_rawand gets"blend": 0.0: its first frame is the clip's last. - The ladder parts are keyed 12.8 cm (0.1116 m proof) higher than the deck and the ground: the package plays every retargeted clip 12.8 cm under proof x 1.147 (Kid's balls of the feet are 10 to 14 cm under the floor in the joint log of the sprint and on the deck in the wave), so the hands and feet meet the real rungs in the game while the joins with the neighbours stay exact. If that sink is fixed for all clips, rebuild with ZOFF 0 in lad_geom.py.
- The guard rail closes the ladder top in the set (mid rail 43 to 58 cm and top rail 100 to 116 cm over the deck, right across the ladder).
SM_cannon_platform_ladder_gap.glbis the platform with those two rails cut between the posts (329 of 21,940 faces, nothing else moved, same materials): import it over SM_cannon_platform for the cannon set, or he climbs through the rails. The opening is 46 cm clear (proof), so his shoulders pass it with 3 to 4 cm to spare. - Checks (proof body, mannequin): hands within 1.62 cm (left) and 1.37 cm (right) of the rail or post surface on every gripping frame (frames 34 to 128 and 38 to 134), fingertips 0.02 cm into the timber at worst; feet on the rungs within 1.57 cm, the deck stance 1.64 cm (the wave's own pose); no clipping (the closest body part, the left hand, passes 0.12 cm from the rail at frame 31); joins 0.031 cm into the sprint and 0.0003 cm into the wave; two builds identical (0.0 cm); FBX read back 0.033 cm; planted feet slide 1.0 cm at most (right foot on rung 3, frames 97 to 112).
- Known limits: the climb is fast (match feet, a step every 10 frames, knee up to 27 cm a frame on the first rung); the placement in the game is modelled from the layout and the joint log (the wave's pelvis predicted to 3.3 cm of the log), so measure the two joins in the next package; contact is exact on the proof body, expect a few cm on Rex (as the crater).
Branch picks (tower 1, game/motion/kimodo_cannon_t7.py, rank_cannon_t7.py; jobs mj-kimodo-cannon-t7 (4c3e58f) and -t7b (1a03c86); picks with SHA256SUMS in D:\studio-ai\renders\megajam\<job>\picks\, checked on the laptop):
| branch | pick | seed | entry gap cm | exit gap cm | velocity mismatch | gag | note |
|---|---|---|---|---|---|---|---|
| stage_fright | stage_fright-s1 (t7) | 1 | 0.0 | 0.0 | 0.0 % / 0.0 % | nervous: pass (1 of 40) | ends on the high five where the build puts it: the end pin moves 1.6 cm from the commit (director log), the joint log says the main pelvis moves 8.2 cm there, so expect a rejoin gap of about 6 cm spread over the 2 s branch (0.1 cm a frame, was 2.3); 19 joint-limit frames; rejoin_t 11.7667, blends 0 |
| belly_flop | none usable | 0.0 | 0.0 | 0.0 % / 0.2 % | t7: 1 of 40, t7b: 0 of 40 | t7 (arc height kept, 11 m to fall in 1.4 s): only s34 passes the gag, but it shakes (peak joint acceleration 752 against 12.8 in the source), 1031 joint-limit frames and lands 110 cm off and 42 cm high of the boxes. t7b (whole arc removed, he starts at his contact height): seams all 0, steadier (accel 65 to 76, 62 to 131 joint-limit frames), but every seed lands on its back (chest_front_y +0.98 on s33). Stopped after two tries; the old belly_flop-s17 with its 0.6 s entry crossfade stays in the take. Next lever: a mid-clip pin at contact (face-down pose) or a belly-first source pose |
The 12.8 cm sink: cause and fix (Fri Nov 13, wiring run)
Cause. AnimationLibrary.get_bone_pose_for_time(anim, "pelvis") does not return the stored track: it retargets to the skeleton's reference pose and returns the stored value minus (0, 0.62, 12.8) cm for Rex_final's body mesh (stored 94.0, read 81.2 on walk_to_mark_s2, every frame, every clip tried, Walt's and the heavy's included in kind). The mesh-retargeted pose evaluation (AnimPoseExtensions with the mesh, what _joints uses) returns the stored value. The game plays the stored track as it is. director_align.sample_source read with the first call and realize wrote with set_bone_track_keys, which stores exactly what it is given, so every baked ("aligned") copy lost the constant (written 88.06, read back 75.26 in a scratch copy). Unbaked first clips were right; every baked clip had Rex's full-length legs under a pelvis 12.8 cm too low, and each first-to-second clip handover stepped the pelvis 13 cm. Not Rex's mesh (Body at 0, scale 1, bounds z 0 to 147), not a capsule (the BP root is a scene component), not the retarget, not the builder's z. Probes: D:\megajam\builds\logs\sink_probe*.log. Fix. director_build._store_offset measures the constant per clip and body mesh, director_align.realize(..., off) adds it back to every baked key. Removed: ZFIX in cannon2_take.py (take file patched in place by sink_unfix_cannon.py: keyed clip lift, climb-in lift keys, belly_flop and land_slip lifts, applause lift), ZOFF in lad_geom.py (0), the ladder clip rebuilt on tower 2. Joint log, packaged D:\megajam\builds\wire1 (perfect bot), planted-foot sole height (ball z minus 0.9, feet_check.py): was a median 10.7 (chase), 10.0 (brawl), 11.0 (cannon) cm under the support in cannon2; now the unbaked first clip and the baked clips agree: floors +2.1 to +2.8 cm (the walk pose's ball sits 2.1 cm over the flat-foot rest; the unbaked first clip reads the same), chase truck bed 116.1 against 114 (+2.1), cannon deck 301.2 against 298.2 (+3.0), crater 233.8 against 231 (+2.8). Medians of the absolute error 2.1 (chase), 2.5 (brawl), 2.4 (cannon). Outliers: lying poses (hay throw landing +10 cm, a prone body) and the leap's landing frames (-7 for 0.45 s).
Cannon wiring (ladder, platform, stage_fright)
Ladder rebuilt with real heights (ZOFF 0): the step from rung 3 to the deck is 93 cm with the rungs where the set has them, past the reach, so the climb now uses all four rungs (lad_climb.py: climb plan, pelvis keys, grips; ZOFF 12.8 keeps the old plan). Result (lad_result.json): hands within 1.63 and 1.39 cm of the grips, feet within 3.7 (left) and 2.2 cm of the rungs (was 1.6), joins 0.03 cm into the sprint and 0.0003 cm into the wave, planted slide 0.8 cm at most, foot IK error 2.7, hand 3.0 (was 3.9 and 1.9); a toe passes 3.4 cm into the platform edge at frame 126. The old ZOFF 12.8 files are in D:\megajam\proofs\cannon\ladder\zoff12.8\. ladder_import.py brought the FBX in; the opened platform GLB replaced D:\megajam\props\sets_made\SM_cannon_platform.glb (the closed one is SM_cannon_platform.closed_ladder.glb) and was reimported over SM_cannon_platform. cannon.take.json: k_kid_ladder_climb at 16.0 s, rigid, blend 0, psi_abs -5.753, no lift or slide (wire_cannon_ladder.py); the wave at 21.8 s has blend 0. stage_fright_s1 is the t7 pick (its BVH copied from mj-kimodo-cannon-t7/picks, the batch 20 one kept in D:\megajam\retarget\takes_backup_before_t7pick).
Wiring run result (D:\megajam\builds\wire1, 868.4 MB zip, gate 940)
Bots on the package, 0 error lines in all runs: chase 5/1/1/4 stars (perfect, late, none, human 45), brawl 5/1/1/4, cannon 5/1/1/4. Premiere bot full loop with the walk-throughs and beats playing out (walk_chase 30.0 s, beat_1 24.5 s of 35.95 (good branch, 4 stars), walk_brawl 26.8, beat_2 23.1, walk_cannon 30.3, beat_3 25.2), premiere three reels, credits, back to the sheet: 703 s, 0 errors. D:\megajam\builds\cannon2 was overwritten by a first run of wire1_run.ps1 with the wrong default folder (stopped after the chase runs; the cannon2 chase logs are gone, its tables above are the record).
WALKTHROUGH.md
Source game/takes/WALKTHROUGH.md. Last git commit 2026-11-13. About 2,166 words.
Walk-throughs (CONTENT-PLAN section 3): the laptop ticket
Three takes, one before each scene: walk_chase.take.json, walk_brawl.take.json, walk_cannon.take.json, written by make_walk_takes.py (rerun it if a line is re-recorded: every line time comes from the wav lengths). They are the shoot on Stage 7, not the film: no 1978 grade, neutral work light, no cues, no score. Walt walks Kid a few steps to the first mark, points it out, and the far marks get slow inserts while his lines run over them. Kid has one line in each (Rex's voice, the takes' convention: rex_*.wav spoken by the actor kid).
| take | length | shots | lines |
|---|---|---|---|
| walk_chase | 30.0 s | s1 lead dolly on Walt and Kid at the cart, s2 swing gate, s3 mirror maze front, s4 the two of them, s5 the truck | walt_w01 to w04, rex_w01 |
| walk_brawl | 26.8 s | s1 lead dolly along the long table, s2 the heavy on his mark with the bar behind, s3 the rope from below, s4 the two of them, s5 the bandstand | walt_w05 to w08, rex_w02 |
| walk_cannon | 30.3 s | s1 the two of them walking to Kid's mark (Stage 7's lamps behind), s2 wide of the cannon and grandstands (the effects man off camera), s3 the two of them, s4 down the barrel to the truck | walt_w09 to w12, fx_w01, rex_w03 |
How the game should play them
- Before a scene's first attempt, play
walk_<scene>(sequence taggedDIR_walk_<scene>), then go straight into<scene>as now. Not before a retake or a reset. - Skippable with any input (keyboard, pad, touch) once 2.0 s have played (
walkthrough.skippable_after_s). Input before 2 s is ignored, so the button that confirmed the call sheet cannot skip it by accident. A skip cuts straight to the take with the usual fade; stop the walk's audio. - No cue gate: the cue table director_cues.py writes for a walk take is empty (it warns W03 "the take has no cues", which is expected). No HUD calls, no score, no takes for the edit bay, nothing in the premiere or the blooper reel.
- Visibility. A walk take needs its scene's set but not the scene's cast, props or film lights: show
DIR_walk_<scene>(its own Kid, Walt, props, work lights, cameras) plus the set dressing taggedKK_settake_<scene>(walkthrough.set_take), and hide the rest ofDIR_<scene>. Today the sets also carryDIR_<scene>, so the rule "drop every other take'sDIR_actors" would hide the set; withgame/sets/proposed/KKGame_settags.patchthe game already drops sets byKK_settake_, which makes this one line: keepKK_settake_<walkthrough.set_take>. - Light: the walk takes carry their own neutral lights (
work_sun,work_sky), so the game treats them like any take (work lights off while it is up). If you would rather keep Stage 7'sKK_worklightrig on for walk-throughs, deletelightsfrom the three files inmake_walk_takes.pyinstead; do not run both. - Grade: the files have no
gradekey on purpose, so their cameras get no film look. If the game adds any grade or letterbox of its own when a take starts, skip it forwalk_takes.
Builder fields
walkthrough(top level, read by the game only):for_take,set,set_take,skippable_after_s,grade: "none",worklights. director_build.py ignores it (load_take checks only the required keys).- Walt is a
follower(his clips carry their own travel, like the heavy's), because he walks:KimodoWalt/walk_to_mark_s2beside Kid,point_action_s1at the marks,arms_folded_watch_s9while Kid talks,idle_director_s7between. The brawl walk also has the heavy as a follower holdingmark_check_nervous_s6on his mark. - Nothing else is new. Props use the real mesh paths from
game/sets/proposed/take_props.diffwith their placeholders, so they build as boxes until imported.
Build and check (laptop, after the current take worker is done)
run_director_build.ps1 -Take walk_chase(and walk_brawl, walk_cannon), then rerunset_build.py allas after any director build (sets README).- Read the build log:
PATH kidandPATH walt. walk_to_mark covers about 1.5 m in its first 1.2 s; two of them should leave Kid about 3 m on from his start. Chase: Kid must pass the cart (610, -125) with at least 80 cm clear and Walt must stay off the booths. If the walk curves, movelocationinmake_walk_takes.py, not the clips. frame_checks(s1 at 1.0 and 4.5 s): Kid, Walt, and the cart or table in frame.- Two known set issues, for the sets worker, not this ticket: the brawl's bandstand stands at (1537, -35) in the layout but the take's fall lands near x 0; and the mirror maze front (batch 15) is still a stand-in, so chase shot s3 shows only the booths behind its spot until it imports.
Lip sync (later)
The 21 new lines (15 walk lines, 6 calls) need face anims before the faces move: /Game/KKCast/LipSync/walt/walt_w01 to walt_w12, walt_c01 to walt_c06, and /Game/KKCast/LipSync/rex/rex_w01 to rex_w03, made the way the cast's were (game/casting/ls_solve.py, ls_run.py, ls_check.py, LIPSYNC_IMPORT.txt). Every line in the walk takes already has lipsync: true; until the anims exist the build warns per line and the faces play eye life only.
Better Kid motion (optional)
take_walkthrough-s11 (mj-kimodo-takes-4 on D:\studio-ai\renders\megajam, 10 s, walks 2.66 m beside a director, points at a spot on the floor, crouches to touch it, stands and mimes a jump, no joint-limit frames) was made for exactly this and is not retargeted. Retargeted onto Rex_final it could replace Kid's two walk_to_mark clips in all three walks.
Voice
Made on tower 1 in ~/mj-tmp/voice-walk exactly as the final cast (~/mj-tmp/voice-final-work): Maya1 with gen_final.py's settings (gravel for Walt, matinee for Rex, temperature 0.4, top_p 0.9, seeds 4321 + 1000 x idx + take), check_takes.py (Whisper) with WER at most 0.05, Audiobox aesthetics, ECAPA similarity against the cast's reference takes (walt_l01_t1, rex_l02_t2) with the cast's threshold (median minus 0.10), finalize.py (0.15 s pads, -18 LUFS). Rows are in game/audio/voice/manifest.json.
All 15 walk-through lines passed WER and similarity in 3 takes. walt_w03 and walt_w04 use a passing take about 1 to 2 s tighter than the top aesthetics one (within 0.05), to keep the chase walk near 30 s.
The walk-throughs run longer than the plan's 20 s (chase 30 s, brawl 27 s, cannon 30 s); the skip after 2 s covers repeat plays. To shorten, cut walt_w03 (the mirror line) from the chase, about 5.6 s.
Walt's calls for the longer takes (walt_c01 to c06)
Six short calls the longer takes show as text only, recorded the same way: walt_c01 "Right!", walt_c02 "Swing!", walt_c03 "Hit!", walt_c04 "Fall!", walt_c05 "Wave!", walt_c06 "High five!" (0.9 to 1.5 s each, -18 LUFS). All six passed the Whisper check. c01 and c06 also passed the similarity check. c02 to c05 did not in 8 takes: ECAPA can't score a one-word, one-second clip, the same reason the cast's "Jump!" (walt_l08, 0.312) went in flagged. They follow that rule (the closest take that passes WER, similarity_flag in the manifest). Listen to those four before they ship.
For the laptop ticket, the takes' calls they voice (the take JSONs belong to the take worker, so wire them there):
| call text in the take | cue | wav | play at |
|---|---|---|---|
| RIGHT! | chase m1_mirror 28.8 | walt_c01 | t - call_lead_ms like the other spoken calls |
| SWING! | brawl r1_swing 23.15 | walt_c02 | same |
| HIT! | brawl b2_uppercut 6.21, s1_upper 25.3 | walt_c03 | same |
| FALL! | brawl s2_fall 30.25 | walt_c04 | same |
| WAVE! | cannon k1_wave 9.2 | walt_c05 | same |
| HIGH FIVE! | cannon k2_five 13.05 | walt_c06 | same |
Add each as a lines entry (speaker: walt, lipsync: true) and copy it into the flub branches as take_audio.py does for the existing calls. Lip sync for c01 to c06 goes in the same later batch as the walk lines.
Story fixes 2026-11-13 (decisions.md Thu 23:36)
The cannon climax is now Johnny's evening show turned on the robbers' truck (STORY.md), and Kid lands back first on a box stack under the truck's tarp. Lines made on tower 1 in ~/mj-tmp/voice-storyfix exactly as above (gen_final.py settings, seeds 4321 + 1000 x idx + take with idx 400 to 407, 3 takes each, Whisper WER at most 0.05, aesthetics, ECAPA), finalized at -18 LUFS:
| wav | who | text | result |
|---|---|---|---|
| walt_w12 (re-recorded) | Walt | "It'll look bigger on the way down. Land flat on your back and let the boxes do the rest." | take 0 of 3 passing, similarity 0.467 (threshold 0.354), 5.51 s |
| fx_w01 (new) | the effects man, off camera | "It's a spring, Kid. The bang's for the camera." | Maya1 brooklyn take 2, 2.95 s; see below |
| walt_c07 (new) | Walt | "Keep rolling, keep rolling!" | flagged: 8 takes, all pass WER, none reach 0.354; kept take 6 (0.337), 3.19 s |
| walt_c08 (new) | Walt | "Truck's parked, Kid. Give me the leap again for the insert." | take 1 of 2 passing, similarity 0.506, 3.37 s |
The effects man is a new cast voice from the Maya1 designs of the voice bake-off. A take is his only if it passes WER and sits under both Walt's and Rex's own same-speaker thresholds (ECAPA against walt_l01_t1 under 0.354, against rex_l02_t2 under 0.378); designs already cast are out (gravel for Walt, matinee for Rex, brisk for the AD). Of 15 takes over five designs, three passed (brisk t0, brooklyn t2, nervous t2); brisk is the AD's voice, so brooklyn t2 is cast: Walt 0.268, Rex 0.335, AD 0.229. Walt's reference scores 0.352 against Rex's, for scale. brooklyn's other two takes sounded too close to Walt (0.435, 0.438), so a re-record of this line must pass the same check again. Rows are in manifest.json with a voice_match field.
walt_c07 is shouted and short, the same ECAPA trouble as c02 to c05: listen before it ships.
- walk_cannon: fx_w01 plays after walt_w11 ("...you wait for fire") over the wide of the cannon (s2), speaker
fx, no actor,lipsync: false,off_camera: true. walt_w12 runs over s3 and the cut down the barrel. Duration 26.5 to 30.3 s. - walt_c07, the cannon wreck run-on. In
cannon.take.jsontheland_slipreset carries it asline(the flubs' schema). director_build.py readslineonly on flubs, so T7 has to turn the land_slip wreck into a run-on (the cannon doesn't do retakes) before it plays. - walt_c08, the pickup: for T6/T10, when the chase's second truck leap becomes a pickup on the parked truck. Not wired to a take.
- Rex's two lines (brawl 36.0 s rex_x02, cannon 35.7 s rex_x03) are speaker
rex, off camera, no face anim, in the takes and in longer_takes.py / make_cannon_take.py. The take-file half came asproposed/storyfix/(storyfix.patch and apply_storyfix.py).
Lip sync for tower 2 (later), made like the cast's: /Game/KKCast/LipSync/walt/walt_w12 (redo for the new text), walt_c07, walt_c08. None for fx_w01 or Rex's two retagged lines, since nobody on camera speaks them.
In the game (wired Fri 2026-11-13, laptop)
KKGame.cpp StartSide / TickSide / EndSide (phase Walk). The call sheet's run plays walk_<scene> before the scene's first attempt: StartTake asks SidesWanted() (the sheet is rolling, or -KKSides; -KKNoSides turns it off), keeps DIR_walk_<scene> in RemoveOtherTakes, hides every other DIR_ actor (the scene's cast, props and film lights, remembered and restored after) and every set but KK_settake_<set_take>, lets ApplyStageLights give the walk its own work lights, and plays the sequence. A retake (GoAgain) never goes through StartTake, so it never plays. Any key, pad button, mouse click or touch skips once walkthrough.skippable_after_s (2.0) has passed; earlier input is logged KKSIDE input ignored. The HUD draws no grain, no grade and no letterbox, only an "ANY BUTTON TO SKIP" card once input counts. After the walk its actors are destroyed, the scene's come back and the scene's call goes on (banner, 1.5 s, the take). The table the game reads is Content/Data/walk_<scene>_cues.json, which director_cues.py now writes with the take's walkthrough block (python prep/scripts/director/director_cues.py game/takes/walk_chase.take.json --out game/KidKatapult/Content/Data/walk_chase_cues.json). Bots: -KKSkipSidesAt=<s> skips every walk and beat that many seconds in; the default is to play them out. Logs: KKSIDE start|end|visibility, KKBOT skips. Walt's walk (KimodoWalt/walk_to_mark_s2) did not exist: retargeted from walk_to_mark-s2 onto Walt_v3 with RT_STAGES=retarget (cast3_retarget.ps1's recipe), then the three walks rebuilt. Walt's calls walt_c01, c03 (twice), c04, c05 and c06 are in the takes (wire_calls.py, at the cue minus call_lead_ms, lipsync: true, with the megaphone key-up); the director copies them into the flub branches. walt_c02 (SWING!) is not placed: at the cue minus 0.6 s it would sit on top of walt_l06 ("Now, Kid. Go!", 21.8 to 23.84 s), and walt_c05 starts at 8.9 s instead of 8.6 s so it follows walt_l17. Every spoken line with a face anim has lipsync: true except Rex's two off-camera lines (rex_x02, rex_x03), which have anims from the cast but are not spoken by anyone on camera.
WIRING-PLAN.md
Source game/takes/WIRING-PLAN.md. Last git commit 2026-11-13. About 766 words.
Wiring plan (Fri 2026-11-13, laptop, Sonnet): sink fix, cannon wiring, walk-throughs and beats
Started 10:05. Written 10:18 after the sink investigation, which is the only job that had to be proven before anything else.
1. The 12.8 cm sink: cause found (numbers, no engine change needed)
Probes: D:\megajam\builds\logs\sink_probe*.log (UnrealEditor-Cmd python, scratch scripts in the agent scratchpad).
- The packaged game plays the pelvis track as stored (no runtime pelvis retarget). Check: clip
walk_to_mark_s2, the first clip of the chase, is not baked. Game pelvis z 94.0 at film 1.0 s (cannon2 run, chase-perfect), the stored track is 94.0. AnimationLibrary.get_bone_pose_for_time(anim, "pelvis")does not return the stored track. It retargets to the skeleton's reference pose and returns it minus (0, 0.62, 12.8) cm on every clip, for Rex_final's body mesh (stored 94.0, read 81.2). The mesh-retargeted pose evaluation (AnimPoseExtensionswith the mesh, what_jointsuses) returns the stored value: 94.0. Constant over all frames and all clips tried (Kimodo, KimodoWalt, KimodoHeavy, the keyed cannon clip).- The director's baker (
director_align.sample_source,realize) reads with the first API and writes withset_bone_track_keys, which stores exactly what it is given. Every baked ("aligned") copy therefore loses (0, 0.62, 12.8): written 88.06, read back 75.26 in a scratch copy test. Unbaked first clips are right (feet 1.3 to 3.2 cm over the floor), every baked clip is 12.8 low with its legs at full Rex length: feet 10 to 14 cm under the floor, the deck too, and a 13 cm pelvis step at every first-to-second clip handover (95 to 82 in the chase at 2.6 s). - It is not Rex's mesh (Body at 0,0,0, scale 1, bounds z 0 to 147), not the capsule (none: the BP root is a plain scene), not the actor z (0), not the Blender or IK retarget (the stored tracks are right), not the builder's z placement. It is the bake.
Fix, once: _realize computes the offset per source clip and mesh (mesh-retargeted pelvis minus library read, frame 0) and realize adds it to every baked pelvis key. All baked clips then play at the stored height, consistent with the unbaked ones and with the editor's mesh evaluation. Then remove the compensations: cannon2_take.py ZFIX (keyed clip lift, climb-in lift target, flub and applause lifts) and lad_geom.py ZOFF. cannon.take.json is patched in place by a one-off script (the take file carries edits from other scripts; no regenerate).
Check: rebuild the three takes (director builds exit 0), package, read the joint log (KKJ): ball z at foot plants against the floor / deck / truck bed, within 2 cm, in all three takes. Per-take floor heights come from the take json and layout.
Feasible: yes, proven in the scratch copy test; the full proof is the package.
2. Cannon wiring (ladder climb, opened platform, stage_fright-s1)
- Ladder clip:
D:\megajam\proofs\cannon\ladder\k_kid_ladder_climb.fbxis keyed with ZOFF 12.8. With the sink fixed it must be rebuilt with ZOFF 0:lad_geom.py(laptop, numpy and trimesh) thenlad_climb.pyandlad_dump.py(tower 2, blender, via ssh). Feasible if tower 2 answers over ssh; if not, the clip plays 12.8 cm high on the rungs and I say so. - Import:
t7_import.pypattern (FBX onto the MH template skeleton under/Game/Motion/Keyed/Src, IK retarget to Rex_final into/Game/Motion/Kimodo/k_kid_ladder_climb). The opened platform GLB goes where SM_cannon_platform's GLB is imported (set import script). - Take: replace
walk_to_mark_s2at 16.0 s byk_kid_ladder_climb(rigid, blend 0, psi_abs -5.753, no lift, no slide), the wave at 21.8 s getsblend: 0, keepsface_raw. stage_fright flub anim becomesstage_fright_s1from the t7 pick (alreadyfail_anims.stage_fright; the picks live inD:\studio-ai\renders\megajam\mj-kimodo-cannon-t7*/picks).
3. Walk-throughs and beats in the game (C++, KKGame.cpp)
WALKTHROUGH.md and BEATS.md give the full spec. Take files and audio exist. Needs: StartTake hook for walk_<scene> before a scene's first attempt and beat_<n> after each edit bay, 2 s skip lockout, visibility rule (DIR tag plus KK_settake_<set_take> / KK_set_stage7), no grade, beat branch pick by stars (jump_at then seek), plus take edits: lipsync:true on every line with a face anim, walt_c01..c06 lines in the takes per the table, phone sounds already mapped. Check by run_cue_tests.ps1 and the packaged premiere bot. Risk: largest job; the game code for takes is one 2400 line file. If time runs out the walk/beat playback is the first thing to report undone.
Order and gates
1 sink fix and rebuild three takes (director builds), 2 ladder and cannon wiring, 3 walk/beat code and take edits, then build_editor.ps1, run_cue_tests.ps1, package to D:\megajam\builds\wire1 (pkg_smoke.ps1 -Profile jam, gate 940 MB), premiere bot full loop, joint log check, commit by explicit paths, push.