Work in progress · how this page was made
This page is shared to explain the method. The test is small and will be repeated with more photos and paid models, so treat the numbers as indicative, not final. Written by Claude (Opus 4.8) from two sources: the statistics captured during the runs, and the working session between Bharat and the model in which the experiment was designed and read. Not yet reviewed by a human.

How five AI models read shelf pictures

What we did: we ran five open-weight vision models over photographs of one bookshelf, reading every spine into a title and author with the same prompt and the same scoring, and changed only the model.
Why we did it: to see whether a small, free, local model can read a real, messy bookshelf — and, more usefully, what decides the result: the model, or the photograph. It is one bookshelf and eight photographs, not a verdict on any model. ShelfBench · Antilibrary · 2026-07-13.

Sibling study: How seven AI models cluster one library runs the same kind of head-to-head one step later — organizing the books instead of reading them.

The control

Change one thing at a time
Held constant — identical for all five
🔒8 photographs · 189 books 🔒One extraction prompt 🔒Split-and-merge tiling 🔒Same fuzzy scorer 🔒Human answer key (Opus 4.8 + person)
The one variable
THE MODEL
× 5 engines
Nothing was tuned per model. If the numbers move, the model moved them — not the setup.
The photographs and the scoring are fixed. So when a score changes, the model changed it — which is what lets us compare the five directly, and what lets one flipped photo isolate orientation.

The five models we tried

All open-weight · small enough to run on a laptop, no cloud
Open-weight · local, free
Alibaba · 7B · open-weight
Qwen 7B
The anchor — best reader here: most books, fewest inventions.
Alibaba · 3B · open-weight
Qwen 3B
The small sibling — isolates scale inside one family.
Meta · 11B · open-weight
Llama 11B
Cross-family. Reads well, but over-eager — it invents more.
Google · 12B · open-weight
Gemma 12B
Cross-family. Reads titles; muddles the authors.
Hugging Face · 2.2B · open-weight
Smol 2B
The floor — effectively blind to the task.

The photographs

One real, messy bookshelf · eight shots, chosen to fail in different ways

We did not tidy the shelf or stage the shots. Real bookshelves are crowded, crooked and badly lit, and a tool that reads only a clean one is no use. The eight photos vary on purpose — one tidy, one crammed three shelves deep, one of repeated titles, one at an angle, one a jumble of orientations and languages — so that no two fail for the same reason.

The busiest photo caught three shelves at once. The models read the shelf below too — real books, counted as mistakes.

What came back

The scoreboard
Every model, every axis. The last column — books simply invented — is the one to watch.
ModelSizeRecallPrecisionAuthorHallucinationsFabricationsTime
Qwen 7B7B0.730.880.692176m
Llama 11B11B0.720.760.70621822m
Gemma 12B12B0.580.670.4166177m
Qwen 3B3B0.460.680.532954m
Smol 2B2.2B0.010.010.001481296m
Qwen 7B leads — and it beat both larger rivals, an 11B from Meta and a 12B from Google. More parameters did not mean better sight. The 2B model could not do the task at all.
Read-rate, shelf by shelf — and what happens when two photos are flipped
modelT001
13
T001 ↻
flip
T002
17
T003
17
T004
26
T005
15
T006
69
T006 ↻
flip
T007
19
T008
13
Qwen 7B1.001.001.000.880.461.000.520.591.000.92
Llama 11B1.000.850.820.820.690.930.610.320.580.85
Gemma 12B0.850.610.590.710.610.800.410.300.680.54
Qwen 3B1.000.690.940.000.540.600.200.220.530.85
Smol 2B0.000.000.060.000.000.000.010.030.000.00
Deeper green = more books read correctly. = the same photo flipped 90° (same books; a paired control, excluded from the totals).
Two shelves defeat everyone: T004 (three shelves crammed in one frame) and T006 (69 books, mixed orientation). The two columns are the same books turned 90°: on the tidy shelf (T001) the weaker models drop and even the best model starts inventing; on the already-hard mixed shelf (T006) rotating just reshuffles which books read.

What we found

1. The picture matters more than the books. We built the eight shelves to be hard in different ways. One shelf held the same journal nine times over. Another held six books by a single author. A third packed in look-alike covers. These were the traps, and we expected them to break the models. They did not. When the type was large and the shelf was tidy, the models read the repeats and the look-alikes without trouble. The thing that decided success was not the books at all. It was the photograph: how it was framed, which way the books faced, the angle of the camera, and how clearly the spines could be read.

2. One model was clearly best. Qwen 7B led on every measure. It read the most books and invented the fewest. It also beat two larger rivals, an 11-billion model from Meta and a 12-billion model from Google. In this task, more parameters did not mean better sight. The gap ran the same way inside one family: the 7-billion model was far stronger than its 3-billion sibling. The smallest model, at 2 billion, could not do the task at all.

3. Recall falls as a shelf grows denser. Recall is the share of the real books that a model reads. It fell as the number of books in the frame rose. On the 13-book shelf, the best model read every title. On the 69-book shelf, it read about half. Every model showed the same slope. A crowded photograph gives the model more to read, less room per spine, and more chances to lose its place.

4. Framing creates false errors. Our answer key covers one shelf per photograph. But several photographs caught two or three shelves at once, and the models read all of them. The extra reads were correct: real books, read right, from the shelf above or below. They simply were not the books we asked about, so the scorer marked them wrong. On the busiest photograph, the models “found” titles like Data Science for Business and Trigonometry. Those books are real. They sit one shelf down. The fault was in the framing, not the model, which means the scoreboard is partly measuring the camera.

5. Orientation is a large effect. We took the easiest shelf and turned it upside-down. The books were identical; only the angle changed. The weaker models lost up to a third of their recall. The best model held its recall, but failed in a worse way. It still found all 13 real books. Then it added 15 that were not there. Upside-down text did not stop the model from reading the words. It stopped it from telling where one book ended and the next began. Turning the picture did not stop it from reading. It stopped it from stopping.

6. Flat books may read more easily than upright ones. One shelf held both kinds: some books standing up, some lying flat in a stack. The models read the flat books better. Across the models, the flat books scored about half right; the upright books, about a third. Four of the five models showed this. The reason is likely plain. A flat book’s title runs left-to-right, the way we read. An upright spine’s title runs sideways, so the model must read it turned on its side. We cannot call this settled, though, and here is the caveat. The upright books on that shelf were also small and faint, so part of the gap is legibility, not orientation. And the best model read the upright books slightly better, not worse. So we treat this as a hint, not a result. The clean test is still to come: photograph the same books once standing and once flat, and change nothing else.

7. Honesty matters more than accuracy. When a model cannot read a spine, it fails in one of two ways. It garbles a real title, or it invents a fake one. We call the invented kind a fabrication, and the difference between the two is everything. Garbling is recoverable: you can see the model reaching for a real book, and All the Light We Cannot See comes back as “We Cannot See.” Fabrication gives you no such warning. It is a clean, confident, wrong answer. The best model rarely fabricated. The smallest did little else. Shown a shelf of science books it could not make out, it reported The Great Gatsby, 1984, and The Lord of the Rings. None were there. It had stopped reading the shelf and started guessing what a shelf usually holds. The number that separates a safe model from a risky one is the fabrication rate, not the accuracy score.

8. You need a good-enough model first, then a good photograph. It is tempting to say the photograph matters more than the model. It does not, and one table shows why. These are the read-rates on the clean, upright, legible shelves — the ones where the picture is not the problem:

Read-rate on the clean, upright, legible shelves — where the picture is not the problem. Same easy photos; the spread is the model.
ModelRead-rate on clean shelves
Qwen 7B1.00
Llama 11B0.83
Qwen 3B0.77
Gemma 12B0.73
Smol 2B0.01

Same easy photographs. The results run from perfect down to nothing, and that whole spread is the model, not the picture. Only the best model read the clean shelves in full. The others missed a fifth to a quarter of the books even when nothing was wrong with the shot, and the smallest read almost none of them. So there is an order to it. First the model has to be good enough to read a clean shelf. Below that bar, the model is the ceiling, and no photograph can lift it. Above that bar, the model reads any clean shelf, and from there the photograph decides the rest.

The practical rule follows from both halves. Pick a model good enough to read a clean shelf — here that was Qwen 7B, not the larger models. Then give it a good photograph: one shelf, so there are no neighbours to misread; straight on, so the small author text stays legible; in good light, for contrast; and right-side up. Get both right, and a free model on a laptop will read your shelf well.

T001, as shot — every capable model reads it in full.
The same shelf, turned upside-down. Same books.

Turning the easiest shelf upside-down, with the same books, is the cleanest test of orientation. The weaker models lost up to a third of their recall; the best model kept its recall but began inventing books — 0 became 15.

What this does and does not show