Forty-nine zeros

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.

Whatever you actually did a few paragraphs up: if your one reading of the sharp image beat 31 of 49, you already outread the best single attempt either machine made in 42 tries. If it beat 27, you outread all 42 of them, since none climbed past what a constant reply of unlooking zeros scores for free.

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.

@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.

the owed control

Four more plates, same prompt, same sampling as above: a blank field with no grid at all, this same 7×7 grid with every cell white, the same grid with every cell black, and the swollen-letterform plates from Four and B — unrelated content, same canvas size. The white/black pair is the decisive one: identical structure, opposite content, a computable answer either way. A reader whose reply tracks the image has to separate them; a reader emitting the same string regardless does not.

llava-1.5-7b-hf: ones-rate 1.00 on the all-white plate, 1.00 on the all-black one. Its replies did not change with the image — the same wall this piece already reported.

llama-3.2-11b-vision-instruct: ones-rate 0.50 on the all-white plate, 1.00 on the all-black one (on only 2 genuine replies across both plates — the rest arrived as numbers again, same failure as the main table). Its replies changed with the image — a clean falsification of ‘no signal to degrade’.

So the answer is not one thing. llava reproduces the wall of ones this piece already found — on the blank field, on an all-white grid, on an all-black grid, all indistinguishable — which is the strongest version of “no signal to degrade” this piece could have earned: not merely failing to transcribe the grid, blind to whether there was a grid at all. llama does not: its one genuine reading of each decisive plate is different, and different in the direction the image actually points. That does not rescue llama's transcription — neither reply comes close to the required format or the right cell count — but it means llama's failure in the main table above is not indifference to the picture. Something in it responded to the picture and still could not say what it saw correctly. That is a real complication to “there was no signal in it to degrade” two sections up: true of one model's replies, not of the other's.

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 other pieces on this site. Three of the four read machine output as evidence about machine vision; the fourth does the same with a language model and a passage of text.

Their material has no computable answer underneath it, so a reply that ignored the image and a reply that saw it and got it wrong look identical there, and get described the same way. This grid is the first object here with an answer a machine can check, 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. (A separate, later question — not whether degradation mattered, but whether the image mattered at all — is the owed control above, and its answer is not this clean for both readers.)

Two corrections to that, both owed to the older pieces, and both found by re-reading them rather than trusting a summary of them. B got to the floor effect first: a hundred readings that never changed, at every destruction level including the perfectly legible one. It also ran a control before this piece had one — two unrelated images, correctly reported as containing no text. That is a cruder instrument than a parity grid, but it is the same instrument, and it was here first; this piece's own control, above, is the parity-grid version of the same check, run months later. And Golf ball cuts the other way: sixteen rounds of descriptions that visibly track what they were shown, an abstract field becoming metal becoming a golf ball. Whatever is happening here, it is not that these models don't look.

So the claim has to be narrower than it first wanted to be. Not that machine reading is unchecked in general — that these four pieces cannot check the readings they publish, and that on this one task, strict transcription of a small structured thing, the readings carry nothing. 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.