Where the tree points go

In short

The three AI animals: better, and the gate to go on was met. After the fixes the red-dot trees have far fewer points than today's rule (flamingo 31 to 19, agama 88 to 37, scorpion 142 to 76) and sit on the marked joints: the flamingo's neck follows its curve and the scorpion gets a point at each leg bend. Marks with no limb near are now dropped instead of piling into the body. The gate was judged on the run before the last fix, where the agama had 29 points, one end per foot and a clean Skin body; the last fix (pruning toes only from marks on side branches, so a tail is never cut) brought a few agama toes back (37 points) and a fold in its Skin body.

The artist-skeleton check fails its pass line. On ten rigged models, the red-dot trees reach recall and precision of 0.80 on none (best: chimpanzee 0.70 / 1.00, gecko 0.75 / 0.81), beat both shape rules on 3 of 10, and the two tries agree on 66% of joints (needed 80%). Two reasons, from the pictures and the 2D diagnostic below: (1) the marks themselves are fair where joints are visible (in 2D: gecko 0.94 / 1.00, spider 0.84 / 0.85, chimpanzee 0.80 / 0.96) and poor where they hide (eagle wings under feathers 0.44 / 0.75, penguin flippers and legs inside the body 0.38 / 0.28); (2) the bigger loss is turning marks into 3D on the carved shape: the joints are snapped onto the carved shape's centre line, which runs elsewhere where the four views merge limbs (the dog's legs: 0.59 / 0.62 in 2D but 0.14 / 0.11 in 3D). Three of the ten answer keys are weak from the files themselves (giraffe: leg bones stored flat, no limb joints; turtle: flipper bones carry no weights; gecko: walking legs have no mirror partner), so those numbers say little.

Part 1: three AI animals

These drawings have no artist skeleton. The tables are diagnostics; the owner's eye supplies the final check.

Flamingo

No limb near: 5 named marks dropped (neck-body attachment, left wing shoulder, right wing shoulder, left foot attachment, tail base). Hidden in every view: 0 named marks skipped (). Unmarked tip branches pruned: dots: 1, named: 1.

Agama

No limb near: 12 named marks dropped (neck-body attachment, left front shoulder, left front elbow, left front wrist, right front shoulder, right front elbow, right front wrist, left hind hip, left hind knee, left hind ankle, right hind knee, right hind ankle). Hidden in every view: 0 named marks skipped (). Unmarked tip branches pruned: dots: 6, named: 4.

R2 budget could not be met: 37 requested; the required structure alone needs 38 points.

Scorpion

No limb near: 30 named marks dropped (left leg 1 body attachment, left leg 1 proximal bend, left leg 1 distal bend, left leg 2 body attachment, left leg 2 proximal bend, left leg 2 distal bend, left leg 3 body attachment, left leg 3 proximal bend, left leg 3 distal bend, left leg 4 body attachment, left leg 4 proximal bend, left leg 4 distal bend, right leg 2 proximal bend, right leg 2 distal bend, right leg 3 proximal bend, right leg 3 distal bend, right leg 4 body attachment, right leg 4 proximal bend, right leg 4 distal bend, left pincer body attachment, left pincer elbow, left pincer wrist, left pincer movable finger hinge, right pincer body attachment, right pincer elbow, right pincer wrist, right pincer movable finger hinge, tail base, tail terminal bend, stinger base). Hidden in every view: 0 named marks skipped (). Unmarked tip branches pruned: dots: 0, named: 2.

Part 2: artist-skeleton check

The answer key is hidden from marking. Only weighted deform joints are eligible. Nearby joints merge within 3% of the longest side; leaves of the original skeleton and zero-length duplicates go. Helpers outside the four-silhouette hull grown by one cell go. Limb joints are key joints farther than tau/4 from the mid-plane x = 0 with a left/right partner: another key joint within tau of their mirror image. All other key joints are main chain. A key with no limb joints is unavailable for recall and fails (a).

Score all tree points except degree-one ends. Recall is the share of key limb joints found; precision is the share of tree points near any key joint. Tau is 5% of the posed model's longest side. Repeat agreement is the one-to-one matched intersection divided by the union. One kept dots try or two empty tries are unavailable and contribute zero to the median over models. An R2 tree whose point budget cannot be met is unavailable and that model fails (b).

Dots: fixed pass lines

• (a) recall and precision both ≥ 0.80 on ≥ 70% of models: 0/10; FAIL.
• (b) F1 strictly beats both R0 and R2 on ≥ 70% of models: 3/10; FAIL.
• (c) median repeat agreement ≥ 0.80: median 0.657 (10/10 models; 1 unavailable, counted as zero); FAIL.

Named: fixed pass lines

• (a) recall and precision both ≥ 0.80 on ≥ 70% of models: 0/10; FAIL.
• (b) F1 strictly beats both R0 and R2 on ≥ 70% of models: 1/10; FAIL.
• (c) median repeat agreement ≥ 0.80: not measured: the named variant read each sheet once.

elephant

Tau 26.77 cm; mesh/split box offset 6.521 cm (1.22% of longest side).

dots repeat: 0.261.

Named repeat: not measured (one read per sheet).

giraffe

Tau 25.00 cm; mesh/split box offset 0.269 cm (0.05% of longest side).

r2: unavailable (no limb joints in key; requested 38, actual 38, structure 21).

dots: unavailable (no limb joints in key).

named: unavailable (no limb joints in key).

dots repeat: 0.464.

Named repeat: not measured (one read per sheet).

chimpanzee

Tau 7.50 cm; mesh/split box offset 0.125 cm (0.08% of longest side).

dots repeat: 0.688.

Named repeat: not measured (one read per sheet).

eagle

Tau 10.00 cm; mesh/split box offset 0.240 cm (0.12% of longest side).

r2: unavailable (R2 point budget unavailable; requested 43, actual 64, structure 64).

dots repeat: 0.647.

Named repeat: not measured (one read per sheet).

penguin-standing

Tau 3.98 cm; mesh/split box offset 0.071 cm (0.09% of longest side).

dots repeat: 0.636.

Named repeat: not measured (one read per sheet).

turtle-loggerhead

Tau 5.00 cm; mesh/split box offset 0.047 cm (0.05% of longest side).

dots repeat: 0.667.

Named repeat: not measured (one read per sheet).

gecko-house

Tau 0.60 cm; mesh/split box offset 0.008 cm (0.06% of longest side).

dots repeat: 0.714.

Named repeat: not measured (one read per sheet).

spider-rigged

Tau 0.42 cm; mesh/split box offset 0.005 cm (0.06% of longest side).

r2: unavailable (R2 point budget unavailable; requested 118, actual 161, structure 161).

no second try passed the outline check

dots repeat: unavailable (one try).

Named repeat: not measured (one read per sheet).

dog-shepherd

Tau 3.36 cm; mesh/split box offset 0.040 cm (0.06% of longest side).

dots repeat: 1.000.

Named repeat: not measured (one read per sheet).

chicken-standing

Tau 2.25 cm; mesh/split box offset 0.052 cm (0.12% of longest side).

r2: unavailable (R2 point budget unavailable; requested 35, actual 43, structure 43).

dots repeat: 0.857.

Named repeat: not measured (one read per sheet).

The marks alone, in 2D (diagnostic)

This diagnostic was added after the verdict. It is never a pass line and does not change the verdict above. Hidden joints are unknown, so 2D recall is optimistic: all artist key joints are projected, including occluded joints. Recall uses limb joints; precision and nearest-joint distances use every key joint. Tau is 5% of the longest side.

Values are unweighted means over the four views. Median cm is the mean of the four view medians. Red dots pool all kept tries without deduplication; per-try and per-view results are in marks2d.json. Undefined values are unavailable.

How this was made

1. Render posed meshes and skeletons at the same frame, turn head toward −Y and up +Z, scale to approximate real dimensions, centre the box in X/Y and set ground to zero. Reuse v54's clay cameras and green sheet layout. Split each sheet with its measured real longest side; report and reject box offsets above 2%.
2. Carve with two pixels of outline growth and the shared 256-cell centre line. R0 stays unchanged. R2 keeps all centre-line roots, forks and tips and repeatedly adds the largest centre-line or radius departure from a straight interpolated run until the dots count is reached. An infeasible structural minimum is reported.
3. Dots keep the two-view support and unambiguous one-view rules from step A. Named marks hidden in every view are skipped and reported separately. A visible named mark must lie within max(3 pixels, half the projected cell radius) in every supplied view or it is dropped as “no limb near”.
4. The centre-line main run is the path through each root carrying the most centre-line length. Only marks on side branches off that run start distal pruning. Beyond such a marked joint, a fork drops unmarked tip branches except the child with the most remaining bone length (volume breaks ties), which keeps the limb's end. Elsewhere an unmarked tip stays only if its reach past its parent surface is at least 2.5 parent radii. Branches with joints stay. Former forks with one child go. Only substantial bulges add interior shape points. Linked points closer than the larger radius merge: forks inherit joint identity; crowded joints keep the better-supported position. Tip extension moves unmarked ends and adds no interior points.
5. Render balls and Skin from the same four trees, without a ball fit, membranes or shells. All numerical pass verdicts come from summary.json; visual readings await the director.