{"n": 1, "what": "layered template: 11 layers cut from archive f006 by bone-Voronoi over the SAM mask", "looked_at": "proof/template_layers.jpg", "saw": "head, torso, tail and four two-segment legs separate cleanly; f006 chosen over f008 because its tail is visible and all four legs are apart", "decision": "keep"}
{"n": 2, "what": "fit v1: soft-silhouette + keypoint render loss at 192x144, finite-difference least_squares, sequential over 50 frames", "looked_at": "build/cache/fit_attempt1.log", "saw": "IoU 0.82-0.91 on profile frames, but ~50-280 s per frame on the shared machine (load 12-21); frames 19-25 fall to IoU 0.2-0.6", "decision": "killed at frame 25 (no params saved); rewrote as resumable chunks at 96x72, one BLAS thread"}
{"n": 3, "what": "provisional render with the f006 texture only (no learned model) to test stage, person, IK and timing", "looked_at": "scratchpad sheets of frames 0-660", "saw": "door leaf hinged on the far jamb covered the lit room; the template head held high reads as howling; lying reads; front view of the turn is a leg tangle", "decision": "hinge on the near jamb, brighter room, head neutral lowered, far legs fade toward front view"}
{"n": 4, "what": "fit v2: two chunks (0->24 forward, 49->25 backward)", "looked_at": "proof/fits_sheet.jpg (first version)", "saw": "front and right-facing frames (26-49) stuck in local minima: tail outside the mask, head on the wrong end, IoU 0.40-0.55", "decision": "multi-start (tail hung down, head lowered, both mirror states) + a sigma-4 blur level for a wider basin"}
{"n": 5, "what": "fit v3 for frames 17-49 with multi-start", "looked_at": "proof/fits_sheet.jpg", "saw": "IoU 0.55-0.85 (mean all 50: 0.73 at 96x72); front frames 26-32 fit 0.75-0.85; head on 43-47 (dog lowering its head to the bowl) still poorly placed; 19-25 masks swallow the shadow", "decision": "exclude 19-25 by eye from learning (23, 49 also flagged by mask area); keep"}
{"n": 6, "what": "learn: view latent, view regression of shape, view-binned appearance atlases (w^2 kernel), held-out evaluation", "looked_at": "proof/learned_atlas.jpg, proof/eval.json", "saw": "held-out MAE 0.093 learned vs 0.107 single-drawing rig vs 0.098 drawing swap; atlases blurry grey averages, tail low-confidence everywhere", "decision": "try a sharper kernel"}
{"n": 7, "what": "atlas kernel w^4 (nearest views dominate) + larger joint caps on knees", "looked_at": "proof/eval.json, scratchpad learned2.jpg", "saw": "held-out MAE 0.098 (slightly worse than w^2, still below the rig 0.107, equal to swap); knee gaps closed; texture marginally crisper", "decision": "keep w^4 for the film (visual), report both numbers"}
{"n": 8, "what": "model raw failures at generated poses", "looked_at": "scratchpad learned3-7.jpg", "saw": "(a) front view: foreshortening only x stands the head upright (giraffe); (b) look back over the shoulder: the reflected head kept its unreflected tilt and pointed at the sky; (c) look in: x-foreshortening of an up-left head makes it point up, and the model cannot tell a head turned away from one turned toward us", "decision": "corrections: head tipped toward camera with psi_h, reflected tilt for reflected head, look-in muzzle levelled + face shadowed; learned head tilt from the view regression now applied"}
{"n": 9, "what": "full render 768 frames, trace, check_handmade", "looked_at": "invariants.json, contact sheet film1.jpg", "saw": "I1 failed at retreat->watch (my bbox-height criterion caught gait bob at 1.83 s); after measuring the top of the dog instead: all onsets within 0.17 s; ALL OK", "decision": "keep the top-of-dog criteria"}
{"n": 10, "what": "person at the end faced away from the dog at the sill; offered hand rested on the knee", "looked_at": "frame 672 (28 s)", "saw": "the hand did not read as offered; the person sat with its back to the dog after it passed", "decision": "hand reaches lower and further; the silhouette turns toward the door 21.9-22.5 s (scaled through edge-on); re-render all"}
{"n": 11, "what": "independent critic, round 1: DOES NOT PASS narrowly; five conditions (depth vs person 19.5-24.6 s, ghost limbs in the turn back 8.20-8.40, knot in the refusal turn 3.50-3.70, look-in barely changes the silhouette, head not below withers in the approach)", "looked_at": "critic report", "saw": "the critic credits the learned per-view texture and the end-card raw-vs-corrected strip", "decision": "repair all five, keep timing/sound/rights"}
{"n": 12, "what": "round 2 repairs: dog path at Z=-0.25 and person on a drawn crate at Z=1.35 (feet always lower than the shoes, no overlap); turns as whip-arounds that never show 55-125 deg; hard far-leg visibility switch, mirror crossfade removed; neck joint drops 30 template px when the head is low; look-in raises the head into the lit doorway, held 27.0-28.2 s; door-light tint on the dog cut to 0.32", "looked_at": "scratchpad r2a (turn frames 82-92, 179-204), r2b/r2c (depth, look-in, head low), film frames at 22.75 and 24.25 s zoomed", "saw": "four legs and head distinguishable in each turn frame, no translucent limbs; person's shoes ~40 px above the dog's back; head top 13.4 px min below withers (measured); look-in silhouette clearly different from rest", "decision": "make.sh end to end: ALL OK, 912 frames, 11.9 MB at CRF 23 -> re-encoded CRF 25, 5.0 MB"}
{"n": 13, "what": "critic round 2: DOES NOT PASS narrowly (depth, ghosts, turns, head-low all fixed). Round 3 repairs: back-of-head layer (skull curve, two ears, rim) swapped in for the look-in, head level, held 27.0-28.2; lying forelegs flat with drawn paw pads, far legs and hind shins hidden, hind thigh behind the torso; standing lower-leg bend clamped (no hooks); breathing while lying; person body no longer turns (only the head profile); flite leading silence trimmed, calls placed at 0.88/3.30; end card rewritten so lying/looking back/looking in appear only as corrected", "looked_at": "scratchpad r3a-r3c, fl.jpg (two foreleg angles), frames extracted from the encoded film.mp4 at 3.7, 6, 7, 22.0-22.4, 25, 26.6, 27.5, 28.1, 30 s and the end card; voice-band envelope of film.mp4", "saw": "ears and skull read from behind in the doorway; lying paws read as paws; person keeps full width; call onsets 0.938 and 3.358 s", "decision": "check_handmade ALL OK; 912 frames; 3.9 MB"}
{"n": 14, "what": "critic round 3: DOES NOT PASS narrowly (cross-pose identity of the look-in head). Round 4: back-of-head fill = learned head fur texture + rim, folded dog ears, neck taken from the learned head layer (no gap); turn back with 3 solid intermediates and far legs hidden above 40 deg; lying hides both hind segments, tail rests on the boards; low hand below the knee by 7.15 s; standing hip height 64->55 template px so feet reach the floor (reach error mean 8.6->2.1 px; the 'sill overdraws legs' was legs ending in the air at the sill line); correction layers listed in assets.json", "looked_at": "every frame 8.0-8.6 s cropped from film.mp4, crops at 23.4, 23.8, 25.0, 27.2, 28.2 s and the hand at 7.2 s, all from the encoded film", "saw": "turn back: solid single bodies, <= 4 legs, no dissolve; look-in head is fur with ears, attached; hand below the knee, forward", "decision": "check_handmade ALL OK; 912 frames; 3.9 MB"}
{"n": 15, "what": "critic round 4: DOES NOT PASS, close. I first extracted 5.0, 6.5, 24.5, 27.0, 31.0 s from film.mp4 at full resolution: the critic was right (foreleg reaching forward, tail trailing back on the floor = frozen leap); my round-4 report of 'tucked' was wrong. Round 5: forelegs bent under the chest, shortened and drawn behind the torso; hind segments hidden; tail curled forward; back-of-head fill 92% fur, body-layer feather, no rim; turn back without stepping, far legs hidden only above 80 deg; end card rewritten to claim only what assets.json lists", "looked_at": "the same five frames and every frame 8.10-8.30 s, from the re-encoded film.mp4", "saw": "lying legs under the body, paws close to the chest (paw pad still a little hooked); back of head reads as fur with ears; turn back shows four legs in most frames, slightly bunched at 8.25-8.30", "decision": "check_handmade ALL OK; 912 frames; 3.9 MB"}