{"n": 1, "what": "cut all 50 archive-dog frames by their SAM masks; contact sheet", "saw": "a real turn in the footage: f0-19 left profile, f22-35 facing the camera, f40-47 right profile head low. Masks of f20-25 include lawn; f23 unusable", "changed": "garbage mattes for f6,8,12,20-25; f23 skipped; largest-blob filter"}
{"n": 2, "what": "lying dog from frame-10 sublayers (head, torso, two forelegs, hind leg) re-assembled by pivots", "saw": "first try: forelegs fanned like prongs and the hind leg went below the floor", "changed": "one near foreleg + one far foreleg rotated flat, lower hind leg tucked forward, torso bottom laid on the floor line; reads as a sphinx-lying dog"}
{"n": 3, "what": "first stage frame (9 s)", "saw": "NaN from the porch-post shading blacked out the right half; the hallway plate showed only its floor because the oblique projection lifts deep things", "changed": "clip before power; back wall moved to D=-1.5 m"}
{"n": 4, "what": "occlusion DAG on the first frames", "saw": "spurious depth 'conflicts' between the porch floor and standing actors (antialiased edge pixels) and between the sill and the casing", "changed": "typed layers: supports (floor, sill) sit under what stands on them; the room sits behind anyone in the doorway; conflict only if the minority side is >2% and >20 px"}
{"n": 5, "what": "lying across the sill, first geometry", "saw": "the whole dog came out 'in' (behind the frame plane) or lay on the porch in front of the door: the body-centre offset put the loin left of the jamb", "changed": "one fixed mapping body centre = x - 0.14 m; lying body angled 31 deg into the doorway (depth slope 0.6), so its own depth pass crosses the casing plane under the jamb: dog_in < jamb_R < dog_out, and merging the two sublayers makes a 2-cycle (logged per frame)"}
{"n": 6, "what": "draft render, every 6th frame", "saw": "the dog is small in the wide two-shot; the turn back at 8.0-8.3 passed through camera-facing frames so the head left the door; the lie-down started from a head-up rig pose (pop from the head-low walk); the look-in head was a black ball floating off the neck against dark boards", "changed": "camera pushes onto the dog for the refusal (3.5) and the turn back (7.7) and pulls back for the causes; TURN_BACK skips camera-facing cels; lie starts head-low (-78 deg); look-in = foreshortened profile head turning away (T) then the frontal-head outline seen from behind (B), attached to the neck, dusk on it and the room's light around its edge; room floor brightened so the silhouette reads"}
{"n": 7, "what": "the push", "saw": "the seated person's arm never appeared: its depth pass put the whole arm behind the wall plane, so the facade correctly held it out", "changed": "arm in front of the wall, only the hand in the opening on the leaf's face; the hand follows the leaf's free edge; push 6.0-6.6, door swings 6.15-7.0"}
{"n": 8, "what": "sound", "saw": "reverb by direct convolution took minutes on the shared machine", "changed": "FFT convolution; measured: soft lines 11.5 dB under the calls, 15.6 dB over the room tone (not listened to)"}
{"n": 9, "what": "first full render (stopped at frame ~60) and review render, every 3rd frame", "saw": "at 2.33 the body slid ~0.1 m when the head-low walk cel gave way to the head-up cel (cels were registered by their alpha centroid, which follows the head); cel 16 read headless; cel 22 a narrow blob in the turn; the push-in at 3.5 lost the person, so the reach was a hand at the edge of frame; the dog small in the wide; the person walked through the bench (an order flip with 5833 px of overlap at 6.0 s: a pop the assertion caught); the look-in head a blue balloon", "changed": "profile cels registered by the rump (barely moves when the head does), camera-facing cels by their centre; dropped 16 and 22; one framing 0-4.7 s holding the caller and the dog; a bigger dog (shoulder 0.70 m); the person steps out past the bench front then back onto the seat"}
{"n": 10, "what": "per-frame assertions on the new geometry", "saw": "the 'who hides the dog' record skipped the post (bug: it started from a dog_in layer that was empty); the leaf near its hinge cut through the dog's chest and the standing person (hinge on the wall's outer face); with the casing plane at 0.005 m the lying dog came out entirely in front of the jamb; the verifier disagreed on the lamp glow (unclipped) and on the sill/floor tie", "changed": "hiders per sublayer; hinge on the inner face of the wall (D=-0.16); lying depth shifted so the depth pass crosses the casing plane at x=0.54 under the jamb; facade at the casing plane; verifier compares clipped images and stands the sill on the boards. After: no conflicts, verifier 0 px on the test frames"}
{"n": 11, "what": "looking in (26.4)", "saw": "a dark frontal head read as a ball; a profile head lifted toward the bulb is what the brief rejects (a lift toward the light)", "changed": "the back of the head as a matte and a fill: the frontal head outline of f28 filled with the dog's own neck fur from f10 (no face shows), lit like the body, the room's light around its edge, reached through a foreshortened profile turning away; measured onset = the top of the lying dog rising"}
{"n": 12, "what": "final render (768 frames), measure, assemble, check", "saw": "onsets first measured on the screen bbox were wrong (the boil cels moved the centroid 2 cm and triggered 'retreat' at 0.63 s; the sheared lying body's bbox stayed tall, so 'lie down' read at 26.3 s); a float floor held one turn frame twice, which broke the 'sustained change' test for the refusal", "changed": "onsets measured on the dog's own alpha pass in its layer space (head over the line of the back, frame-to-frame pixel change, height falling three frames running); int(round()) for the turn index; porch post moved to x=2.20 so it no longer covers the dog's nose at the start; frame 2's tail added. Result: invariants ALL OK, onsets dt 0/0/0/0/0/+0.166, look-in +0.10; DAG acyclic, 0 conflicts, 0 pops, verifier 0 of 192 frames over 50 px; film 912 frames, 38.0 s, 4.0 MB"}
{"n": 13, "what": "round 2, from critic round 1 (DOES NOT PASS)", "saw": "the critic's jamb defect was my bug: the dog layer's affine dropped the shear's constant term, so every sheared dog sat ~90 px low, on the porch in front of the sill while its depth said behind the casing; also the look-in disc, the turn cels (box, two heads, sniff tangle), the boulder walk cel, no contact shadow on dark boards, a straight matte edge on the hindquarters, head halos, the person cropped at 7.8", "changed": "b2 = Ya - k*va - shear*k*cx*ua; crossing at the jamb's outer edge (x=0.575); look-in = real frame-0 head seen from behind; TURN_AWAY 13,14,26,28,30,31,33,36,38,40 then a mirrored profile hold; head level over the shoulder, quicker rise, TURN_BACK 26,14,13,11,10; walk f43-45; wider contact pool; hind part behind torso; rim halved, halo removed; camera holds the two-shot at 7.6-9.4. Re-rendered 768 frames; invariants ALL OK; onsets unchanged within 0.17 s; 0 conflicts, 0 pops, verifier 0/192"}
{"n": 14, "what": "round 3, from critic round 2 (DOES NOT PASS, close)", "saw": "walk-in: three-quarter-front walk cels carry their hind feet high, so angled into the doorway the rear sat on the siding (feet 28 px above the porch line, measured); the depth crossing ran right of the opening so the wall held out a strip of the rear, which shrank to a sliver then popped at the cel-to-rig swap at 23.79; the over-the-shoulder head was a two-part lump at 7.67-7.79; flite's 0.27-0.35 s lead silence made every line late (first call audible 1.30, come in 26.80); the contact shadow was a halo round the outline", "changed": "frame-10 parts walk from x=2.05 (15 deg leg swing, level feet) and the same parts lie down; crossing held at x=0.575; hind part overlaps the torso, drawn behind; rising uses frontal head then real cels 28, 26, 14; voice lead silence trimmed and lines placed by audible onset (0.82, 3.24, 7.28, 10.42, 26.30); contact shadow cast below the contour only; colour bleed in cels; scene fade-in removed. Render stopped at 562 on a 1-px-overlap bug in build_dag (fixed; remaining 206 frames rendered with the same code). Measured: rear lowest pixel <=4.9 px above the porch line 22.1-24.4, adjacent rear-area change <=5.8%; ALL OK"}