✈️ BOEING 747-400 — AUTORESEARCH SESSION

21 experiments · frozen SHA-256-hashed scoring harness · one editable file (src/plane.js) · built entirely without vision

35.6

baseline

76.1

final score

92.1

IoU / 100

~60

VLM / 100

experiments

Final renders — frozen scoring cameras

SIDEorthographic

TOPorthographic

FRONT-QUARTERperspective

REAR-QUARTERperspective

SIDE SILHOUETTEIoU 0.924

TOP SILHOUETTEIoU 0.919

Score trajectory

Experiments

#	Experiment	What happened	Score
1	baseline (judge v1: claude-sonnet-4.5)	Crude placeholder: cylinders, cones, slabs	42.14	KEEP
2	re-baseline (judge → stealth/ox-alpha)	User-directed judge switch; harsher rubric	35.64	KEEP
3	fuselage loft (double-bubble hump)	Max-union upper lobe, upswept tail; fixed inverted normals	40.48	KEEP
4	wings + harness orientation fix	Airfoil wings; caught masks compared MIRRORED vs references	68.40	KEEP
5	engines/fin/tail bundle	Port nacelles at wingXLE(−z) → engines near nose; reverted	65.21	DISCARD
6	livery texture	Window rows, cheatline, cockpit, belly — vlm 27→38	70.40	KEEP
7	mipmap fix + salvaged parts	Window dots vanished under trilinear filtering; LinearFilter fixed	72.44	KEEP
8	bisect: fin-TE sign bug	finChord negated → fin mirrored forward; stab lofts reverted	72.44	KEEP
9	envMap + cockpit placement	Procedural equirect sky reflections on all paint materials	73.96	KEEP
10	smooth winglet cant curve	6-station blend replaced hard kink; wider hump lobe	75.00	KEEP
11	tail profile matched to ref	Numeric profile match = +2.6 side IoU (biggest geometry win)	75.89	KEEP
12	wide blended winglet	IoU +0.05 but judge wants a DISTINCT tapered winglet	73.43	DISCARD
13	soft hump blend + delayed lobe	smax fillet; smoother AND closer to reference polyline	76.01	KEEP
14	slim top-view nose + env boost	rz < ry for s<7 — less blobby in plan view	76.02	KEEP
15	engines fwd, dark belly reverted	Darker belly gray read as wrong; light gray is correct	76.06	KEEP
16	engines +4.0	Judge sees pods hanging too far forward of wing	74.10	DISCARD
17	fin/stab root fillets	Clean junctions; IoU-neutral (inside silhouette)	76.56	KEEP
18	env clouds + gear doors	Reflection structure + belly door outlines	76.06	KEEP
19	enlarged fairing	Reads as fat belly, not fairing; reverted	74.04	DISCARD
20	FINAL — 4-sample stabilization	76.56 / 76.06 / 74.56 / 76.06 → median 76.06	76.06	KEEP

Made with no vision

01The judge as eyes

The assistant model cannot see images — every render read returned “model does not support images.” But the harness sends the four renders to a VLM judge three times per measurement, and its flaw reports were real vision: “upper-deck hump blend shows a visible crease,” “nose is blunt and egg-shaped,” “windows are smeared dark streaks.” Every aesthetic decision was made by a blind sculptor trusting this seeing-eye judge.

02Braille renders

A diagnostic printed the silhouette masks as ASCII — # both agree, R reference-only, M model-only. Reading the plane’s shadow as text caught the mirrored-mask harness bug, the forward-mirrored fin (a one-character sign error), and a 0.15 m intake ring that shifted the whole mask’s bounding-box alignment.

03Analytic geometry

The reference silhouettes are coordinate polynomials and the model is parametric — so “match the tail” was arithmetic: reference spans y ∈ [−2.3, +2.4] at station 60, mine [−1.75, +2.98] → lower the upsweep, fatten the radius → verified numerically before committing. That single fix was +2.6 IoU.

Before the harness fix — renders were mirrored against references (side IoU 0.55; the wings can never overlap)

MMMMMMMMM..............................................................................RRRRRRRRR
MMMMMMMMMM............................................................................RRRRRRRRRR
MMMMMMMMMMM...........................................................................RRRRRRRRRR
MMMMMMMMMMM..........................................................................RRRRRRRRRRR
MMMMMMMMMMMM.........................................................................RRRRRRRRRRR
MMMMMMMMMMMMM......................................................................RRRRRRRRRRRRR
MMMMMMMMMMMMMMMM....................................................................RRRRRRRRRRRR
MMMMMMMMMMM##########RRRRRRRR...............................................MMMMMMMM##############RRRRRRR
MMMMMMMMMM###########RRRRRRRRRRRR..................................MMMMMMMMM#################RRRRR
MMMMMMMM##############RRRRRRRRRRRRR............................MMMMMMMMMMM##################RRRR
MMMMMMMM###############RRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRRR#####MMMMMMM####################RRR
MMMM############################################################################MMMMMMMMMMMMM...
...RRRRRRR##########################################################################MMMMMMMMMM..
.RRRRRRRRR############################################################################MMMMMMM.
.RRRRRRRRRRRRRR##########################################################################MMMMMM
......RRRRRRRRRRRRRRRRR#####################################################################MMMMMMMMMMMMMMMM....
................................MMMMMMMMMM###MMMMMMMM##MMMMMMMMMMMMMMMMMMMMMMMMMMMMMMM..........

Final model — side silhouette vs reference (ASCII downsample; full-res IoU 0.924)

......................................................................................M#########
......................................................................................##########
.....................................................................................M##########
.....................................................................................###########
....................................................................................M###########
...................................................................................M############
...................................................................................#############
..................................................................................M#############
..................................................................................##############
.................................................................................M##############
................................................................................M###############
................................................................................################
...............................................................................M################
...............................................................................#################
..............................................................................##################
.............................................................................###################
.............................................................................###################
............................................................................####################
..........M################RR..............................................M####################
........MM####################RRR..........................................#####################
.......MM########################R........................................######################
.....MMM###########################......................................M######################
.....MMM##########################################################MMMMMMM#######################
....############################################################################MM..............
...#################################################################################MMM..MMMM...
..#######################################################################################MMMM...
.###########################################################################################MMM.
M##############################################################################################M
################################################################################################
R##############################################################################################.
.###########################################################################################R...
.R#######################################################################################R......
.RR##################################################################################RRR........
...R############################################################################RRR.............
......R###############################################################RRRRRR....................
.........................................M###MMM.....RR.........................................

Lessons

The bounding box is sacred

Any geometry extending the silhouette bbox — even 0.15 m, two pixels — shifts the bbox normalization and costs 1–2 IoU points across the entire mask. An intake ring protruding 0.15 m below the pods cost 1.7 points until tucked inside the pod silhouette.

Mipmaps eat windows

Physically-sized window dots are sub-pixel after ~30× minification on a 1024-px canvas; trilinear filtering averages them into the paint. generateMipmaps = false keeps them as aliased dots that read as windows. The judge noticed immediately.

Judge noise is ±3

The VLM judge swings ±3 per call at temperature 0. Medians-of-3 still mislead — every near-miss was re-sampled before reverting geometry. Only deltas ≥ 4 were trusted. Wide blended winglets: IoU +0.05, VLM −4 → reverted.

Mirrored masks

The frozen harness rasterized references nose-left while the cameras render nose-right (top view 180° off). Side IoU “0.56” compared an aircraft to its mirror image. The fix — per-view bit flips in mask extraction — was +27 points.

IoU vs realism tensions

The reference has no dorsal fin, no winglets-as-such, and circles for engine pods. Judges wanted all three. Winglets overlap the reference’s fin region (free), the dorsal fin costs 0.4 primary (dropped), engines sit at the realism cap xLE+3.6.

Blind but not deaf

Score = 0.5 × silhouette IoU + 0.5 × VLM rubric over fixed cameras, lighting frozen, harness SHA-256-verified every run. The model more than doubled — without the sculptor ever seeing it.