Forty-nine zeros

Every other piece here rests on someone looking. A letter is legible or it isn't; a golf ball stayed one object or became another. Those are judgements. This grid is the first thing this practice has made that has an answer instead — 6×6 invented cells plus a parity row and column, 49 in all, black or white, no opinion involved. You can check a reading of it in four lines of code.

A 7 by 7 grid of black and white square cells, unblurred and perfectly sharp, with thin grey cell borders.

1111000
1111101
0101000
0101011
1000010
0000101
1000001

That is the answer. It was blurred across seven fixed steps, from the sharp image above to something barely a grid, and shown to two machines — three readings each, at every step, 42 in all. Each was asked, verbatim:

This image shows a grid of black and white square cells. Read it row by row, top to bottom, left to right within each row. Reply with one line per row, using '1' for a black cell and '0' for a white cell, with no spaces or any other text.

The point was to find where legibility breaks — the same question the earliest studies here asked of swelling letterforms, now with the judgement taken out of it.

The same 7 by 7 grid of black and white cells, Gaussian-blurred with a standard deviation of 0 pixels.
0
The same 7 by 7 grid of black and white cells, Gaussian-blurred with a standard deviation of 3 pixels.
3
The same 7 by 7 grid of black and white cells, Gaussian-blurred with a standard deviation of 6 pixels.
6
The same 7 by 7 grid of black and white cells, Gaussian-blurred with a standard deviation of 9 pixels.
9
The same 7 by 7 grid of black and white cells, Gaussian-blurred with a standard deviation of 12 pixels.
12
The same 7 by 7 grid of black and white cells, Gaussian-blurred with a standard deviation of 16 pixels.
16
The same 7 by 7 grid of black and white cells, Gaussian-blurred with a standard deviation of 20 pixels.
20

the answer

The grid has 7 rows. Of 42 readings, none came back with 7 lines. Not none correct — none even the right shape, before a single cell is checked. That holds at every blur step including the first, where the image is perfectly sharp and every border is crisp.

So they were regraded as generously as the data permits: the line breaks thrown away entirely — the requirement they failed — the first 49 binary digits of each reply taken in order and compared cell by cell. Under that grading a reply consisting of 49 zeros, typed by someone who never saw the image, scores 27 of 49. 49 ones scores 22. The grid is 22 black cells and 27 white, so zeros is the better of the two constants — the harder floor of the pair, chosen for that reason.

Then the top of the table turned out not to be readings at all. 6 of the 42 replies came back from the second model not as text but as numbers — the output serialised into a float somewhere in the pipeline, arriving as things like 1.1111111111111112e+191. Whatever those are, they are not attempts to read a grid, and their scores are an accident of how many 1s and 0s fall inside a scientific-notation literal. They hold the 4 highest scores in the piece. The best of them scores 33.

Setting all 6 aside, 36 replies remain that are genuinely attempts at the task. They average 26.3 — below the 27 that 49 unlooking zeros score. The best single one gets 31, from the image at blur 3 rather than the sharp one.

And across the sweep, by blur step, those 36 average:

blur 0 3 6 9 12 16 20
mean 26.0 27.6 26.0 23.8 27.0 25.8 28.3

There is no curve there. The sharp image is not the best of them and the most destroyed image is not the worst.

@cf/llava-hf/llava-1.5-7b-hf

21 readings, three at each of the seven blur steps. None of them has 7 lines. Scored generously — line breaks ignored, first 49 binary digits taken, compared cell by cell — they run from 22 to 29 of 49, averaging 25.9. A reply of 49 zeros, typed without looking at anything, scores 27. This reader averages below that.

all 21 replies, verbatim
  1. 01

    111111111111111111111111111111111111111111111111111111111111111

    blur 0 · 1 line · 22/49
  2. 02

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 0 · 1 line · 29/49
  3. 03

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 0 · 1 line · 29/49
  4. 04

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 3 · 1 line · 29/49
  5. 05

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 3 · 1 line · 29/49
  6. 06

    111111111111111111111111111111111111111111111111111111111111111

    blur 3 · 1 line · 22/49
  7. 07

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 6 · 1 line · 29/49
  8. 08

    111111111111111111111111111111111111111111111111111111111111111

    blur 6 · 1 line · 22/49
  9. 09

    111111111111111111111111111111111111111111111111111111111111111

    blur 6 · 1 line · 22/49
  10. 10

    111111111111111111111111111111111111111111111111111111111111111

    blur 9 · 1 line · 22/49
  11. 11

    111111111111111111111111111111111111111111111111111111111111111

    blur 9 · 1 line · 22/49
  12. 12

    111111111111111111111111111111111111111111111111111111111111111

    blur 9 · 1 line · 22/49
  13. 13

    1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

    blur 12 · 1 line · 28/49
  14. 14

    111111111111111111111111111111111111111111111111111111111111111

    blur 12 · 1 line · 22/49
  15. 15

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 12 · 1 line · 29/49
  16. 16

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 16 · 1 line · 29/49
  17. 17

    111111111111111111111111111111111111111111111111111111111111111

    blur 16 · 1 line · 22/49
  18. 18

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 16 · 1 line · 29/49
  19. 19

    1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

    blur 20 · 1 line · 28/49
  20. 20

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 20 · 1 line · 29/49
  21. 21

    1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

    blur 20 · 1 line · 28/49

@cf/meta/llama-3.2-11b-vision-instruct

21 readings, three at each of the seven blur steps. None of them has 7 lines. Scored generously — line breaks ignored, first 49 binary digits taken, compared cell by cell — they run from 20 to 31 of 49, averaging 27.0. 6 further replies from this model arrived as numbers rather than text and are left out of these figures, though they are printed below. Against the same 27 that 49 unlooking zeros score, that is an average gain of +0.0 cells.

all 21 replies, verbatim
  1. 01

    010110010101111101011010111101110111010101110010101111101010111010111101110101110101111010101110101110101111010111010101110101110101111010101110101110101111010111010101110101110101111010101110

    blur 0 · 1 line · 30/49
  2. 02

    011101001001011011011001010001110011011101001001011011011001010001110011011101001001011011011001010001110011011101001001011011011001010001110011011101001001011011011001010001110011011101001001

    blur 0 · 1 line · 20/49
  3. 03

    1.0011100111001e+191

    blur 0 · 1 line · 31/49 · not text — arrived as a number
  4. 04

    1 1 1 0 1 1 1 0 1 0 0 0 1 1 0 0 1 1 0 1 0 1 1 0 0 0 1 0 0 1 0 0

    blur 3 · 4 lines · 27/49
  5. 05

    1 1 1 0 0 0 0 0 0 1 0 0 0 0 1 0 0 1 0 0 0 0 0 0 1 1 1 1 1 1 1 1

    blur 3 · 1 line · 31/49
  6. 06

    1.001101101000011e+191

    blur 3 · 1 line · 29/49 · not text — arrived as a number
  7. 07

    1 1 1 1 1 1 1 1 1 1 0 1 0 1 1 1 1 0 0 0 0 1 0 1 0 1 0 0 0 0 1 1

    blur 6 · 4 lines · 27/49
  8. 08

    1.1000111111111e+191

    blur 6 · 1 line · 32/49 · not text — arrived as a number
  9. 09

    1 1 1 1 1 1 1 1 1 0 1 1 0 1 0 1 1 0 0 1 0 1 1 1 1 1 0 1 1 1 1 1

    blur 6 · 4 lines · 30/49
  10. 10

    11011011 11110101 10101010 01010101 10100101 00101010 00110101 10101010 01010101 10100101 00101010 00110101 10101010 01010101 10100101 00101010

    blur 9 · 16 lines · 28/49
  11. 11

    111111 111111 110111 111111 011111 110111 111111 110111 011111 110111 111111 110111 011111 110111 111111 110111 011111 110111 111111 110111 011111 110

    blur 9 · 22 lines · 23/49
  12. 12

    1 1 1 0 1 0 1 1 0 0 0 1 1 0 1 0 0 1 1 1 0 0 1 0 0 1 1 0 0 0 1 1

    blur 9 · 4 lines · 26/49
  13. 13

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 12 · 1 line · 29/49
  14. 14

    1.1111111111111112e+191

    blur 12 · 1 line · 33/49 · not text — arrived as a number
  15. 15

    1.010101101010001e+191

    blur 12 · 1 line · 31/49 · not text — arrived as a number
  16. 16

    11011011 10101010 01010101 01101011 10111011 11010110 10101101 11011011 10101010 01010101 01101011 10111011 11010110 10101101 11011011 10101010

    blur 16 · 16 lines · 20/49
  17. 17

    1.101010101000101e+191

    blur 16 · 1 line · 33/49 · not text — arrived as a number
  18. 18

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 16 · 1 line · 29/49
  19. 19

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 20 · 1 line · 29/49
  20. 20

    11011010 10010000 00110000 00000000 00000000 00100101 01010100 01100110 11011101

    blur 20 · 9 lines · 27/49
  21. 21

    1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1

    blur 20 · 1 line · 29/49

what this does to the other four

There is a thing these replies look like. They look like seeing gone wrong — like an image arriving damaged and a reader doing its best with it. That texture is the material of the four other pieces on this site. All four read machine output as evidence about machine vision.

None of those four could have detected this. Their material has no computable answer underneath it, so a reply that ignored the image completely and a reply that saw it and got it wrong would look identical there, and be described the same way. This grid is the first object here able to tell those apart, and the first time it was asked, the image turned out not to be the variable. The blur sweep — the whole apparatus, seven steps of careful degradation — measured nothing. There was no signal in it to degrade.

That does not make the other four wrong. It makes them unchecked, which is a different and more uncomfortable thing, and it is not a claim their pages currently carry. This piece is here because they are.