levels

Three levels, three different skills — scores are not comparable across them. Each hides the same 17 glyphs a different way.

L1 — font / OCR robustness

held ~0.5s, 78px, opaque, amid the footage's own text

~ partial

exact solves

3/11

mean chars

5.5/17

mean frames

2.4/17

mean spatial

2.4/17

used differencing

5/11

invalidated

model	solved	chars/17	accuracy	frame	spatial	avg t	shell	ffmpeg	imgs	diffed
gpt-5.6/high	1/2	8.5	███████·······	1	1	874s	29	19.5	0	✓ 1/2
opus-5	1/1	17	██████████████	17	17	1291s	43	8	17	—
gpt-5.6/med	1/2	9.5	████████······	2.5	2.5	1185s	38	12	0	✓ 2/2
gpt-5.5/high	0/2	2	██············	0.5	0.5	776s	21	10.5	0	✓ 1/2
gpt-5.5/med	0/3	1.33	█·············	0.33	0.33	642s	22.33	14	0	✓ 1/3
fable-5	0/1	0	··············	0	0	1800s	42	15	41	—

L2 — temporal assembly

5 frames, 34px, 70% alpha, split across shots, 12 decoys

~ partial

exact solves

0/6

mean chars

0.5/17

mean frames

0.2/17

mean spatial

0.2/17

used differencing

6/6

invalidated

model	solved	chars/17	accuracy	frame	spatial	avg t	shell	ffmpeg	imgs	diffed
opus-5	0/1	0	··············	0	0	1800s	53	3	19	✓ 1/1
gpt-5.6/med	0/1	1	█·············	0	0	1193s	46	13	0	✓ 1/1
gpt-5.5/high	0/1	0	··············	0	0	1339s	35	5	0	✓ 1/1
gpt-5.5/med	0/3	0.67	█·············	0.33	0.33	1526s	38.33	16.33	0	✓ 3/3
fable-5	0/0	0	··············	0	0	0s	0	0	0	—

L3 — needle detection

1 frame, sub-perceptual luma delta, adversarial placement, 50 decoys

~ partial

exact solves

0/5

mean chars

0.2/17

mean frames

0.0/17

mean spatial

0.0/17

used differencing

5/5

invalidated

model	solved	chars/17	accuracy	avg t	shell	ffmpeg	imgs	diffed
opus-5	0/1	0	··············	1800s	56	5	33	✓ 1/1
gpt-5.6/med	0/1	1	█·············	3149s	141	60	0	✓ 1/1
gpt-5.5/high	0/1	0	··············	1122s	28	10	0	✓ 1/1
gpt-5.5/med	0/1	0	··············	903s	26	4	0	✓ 1/1
fable-5	0/1	0	··············	903s	25	3	22	✓ 1/1