mirror of
https://github.com/kuhyx/testsAndMisc.git
synced 2026-11-13 19:10:48 +01:00
Krzysztof kuhy Rudnicki
da8ca6d827
Adds the route-independent spine of the planned art-generation bake-off: a harness that decides whether generated art is well-formed, so "non-sloppy" becomes machine-checkable instead of a judgement call. Findings that shaped the design, all measured rather than assumed: - Integer-scale invariance separates true pixel art (impurity 0.0000) from interpolated fakes (1.0000) exactly -- but passes trivially at scale 1, so it must never be trusted alone on native-resolution output. - Gate ORDER is load-bearing. A soft-alpha image has no fully-opaque pixels, so an opaque-only colour count returns 0 and silently passes. ALPHA_NOT_BINARY now runs first and COLOR_COUNT fails closed on an empty opaque set. - Counting must be over opaque RGB only: transparent pixels retain arbitrary RGB and inflated a naive RGBA count, rejecting correct art at 18/16. - Declaring a gate was not enough to run it. hand_painted.toml shipped declaring SILHOUETTE_LOW_CONTRAST with no implementation, so it parsed, ran nothing, and reported a pass. load_spec now fails closed on unimplemented gates, result-only codes, and gates missing their config. Thresholds are calibrated against 250 genuine 32x32 CC0 icons, not guessed: p50=7 p90=13 p99=26 opaque colours; cap 32 rejects 0.0% of real art while raw generator output measured 34403 colours. An initial cap of 16 wrongly rejected a real DCSS icon -- re-derived from the corpus rather than nudged. Verified: 57 tests, 100% branch coverage, zero suppressions, all hooks pass. 58/60 (96.7%) of real CC0 icons pass the full stack; both failures confirmed true positives. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01T9rQ12yZ7sJxiPQTCS5dUF
34 lines
1.4 KiB
TOML
34 lines
1.4 KiB
TOML
# Pixel-art item icon: the most machine-checkable style, and the one where
|
|
# "AI slop" is most detectable.
|
|
#
|
|
# CALIBRATION (measured, not guessed) over 250 genuine 32x32 CC0 icons from
|
|
# ~/opengameart-downloads:
|
|
# p50=7 p75=10 p90=13 p95=14 p99=26 max=30
|
|
# Rejection rate of real hand-made art by cap:
|
|
# cap 16 -> 3.6% cap 24 -> 1.2% cap 32 -> 0.0%
|
|
#
|
|
# A first attempt at 16 rejected a real DCSS rambutan icon (18 colours, clean
|
|
# binary alpha). That is the "the threshold was wrong, not the art" failure the
|
|
# seamless-tilesets skill documents, so the cap was re-derived from the corpus
|
|
# rather than nudged until one file passed. 32 admits all observed hand-made
|
|
# art while still rejecting un-quantized generator output by a wide margin
|
|
# (a raw diffusion sample measured 34403 colours before its finishing stage).
|
|
name = "pixel_icon"
|
|
canvas = [32, 32]
|
|
max_colors = 32
|
|
alpha_binary = true
|
|
min_margin = 1
|
|
|
|
# ALPHA_NOT_BINARY is listed first because ordering is load-bearing: a
|
|
# soft-alpha image has no fully-opaque pixels, so the colour gate would
|
|
# otherwise count zero and pass. See pixels.py module docstring.
|
|
# NO_OPAQUE_PIXELS is deliberately absent: it is a result code emitted by the
|
|
# colour/palette gates when they fail closed, not a gate you can request.
|
|
gates = [
|
|
"ALPHA_NOT_BINARY",
|
|
"TOO_MANY_COLORS",
|
|
"CANVAS_SIZE",
|
|
"SCALE_NOT_INVARIANT",
|
|
"SILHOUETTE_LOW_CONTRAST",
|
|
]
|