Is it the subject, or the framing?

This does not say that tight, frame-filling close-ups are safe. A Commons camera-original of a beer bottle is usually a wide product shot on a table; the submission that motivated the study was a phone held close enough that a printed label and a wet reflection filled the frame. The corpus tests the subject, and framing rides along with it rather than being controlled. So the honest scope is: photographs of glossy printed labels and curved metal, as Commons photographs them, are flagged no more often than photographs of trees and doors. Whether extreme macro framing is its own weakness is a different study and this one cannot answer it.

This is that study.

The design, and why it is within-image

The obvious approach — collect close-ups, collect wide shots, compare — walks straight back into the confound the first study was built to avoid. Two sets of photographs differ in far more than their framing.

So every comparison here is between two views of the same photograph. Subject, camera, lighting, lens and photographer are held fixed by construction, and each image is its own control. For each image, with (W, H) its native size:

for f ∈ {0.70, 0.50, 0.35}.

crop(f) and shrink(f) have identical pixel dimensions. They are the same count of pixels of the same photograph, and they differ only in how much of the scene is inside the frame: crop shows less scene at native sampling density — what standing closer or reaching for a longer lens does — while shrink shows the whole scene at reduced density.

That matched pair is the entire point. The first study's one suggestive number, the hard condition at p = 0.046, is parsimoniously explained by the glossy arm carrying more low-resolution originals, and I reported it as such rather than promoting it. Comparing crop against shrink holds pixel count fixed by construction, so a difference between them cannot be that confound coming back wearing a hat.

Pre-registration

Written into framing.py before any image had been through the model in any cropped condition.

• H1 — at matched pixel dimensions, a tighter frame is flagged as AI-generated more often than the full scene. Directional in expectation, tested two-sided.
• Operating point — 0.65, the extension's own shipped threshold, and only that one.
• Primary test, one and only one — crop(0.35) vs shrink(0.35) on the glossy corpus, exact McNemar on the discordant pairs, α = 0.05. f = 0.35 because the hypothesis is about extreme framing: it is the tightest rung, which is a rule statable in advance rather than a result pickable afterwards.
• Secondary, labelled as such wherever they appear — f = 0.70 and 0.50; the whole ladder on the control corpus, which answers the different question of whether any effect is specific to glossy subjects or general to photographs; the paired mean score shift at each rung; and base, which is an anchor and an integrity check rather than a test.
• Bounds that halt the run rather than produce a finding, each one computed by check_bounds() because a rule nothing evaluates never fires: - B1 flagged ≤ scored in every condition. - B2 base must reproduce the first study's stored clean score for every file present in both, to within 1e-9. Same file, same function, same deterministic model — a mismatch means the re-fetched bytes or the pipeline moved under me and every number here would be measured against a shifted baseline. - B3 every re-downloaded file must match its sha256 in corpus_<set>.json; one mismatch is dropped loudly, more than 5% halts. - B4 the primary test needs at least 10 discordant pairs, else report underpowered rather than a null. - B5 the detector's own reject gates may skip at most 20% of any condition.

Stated in advance, because it cuts toward the null

A centre-crop assumes the subject is near the centre. On Commons product shots it usually is, but where it is not, crop(f) is a crop of background rather than a tighter view of the subject, which weakens any real effect rather than manufacturing one.

And a centre-crop is not the same act as physically moving closer: perspective, depth of field and the subject-to-background relationship all change when a photographer steps in, and none of them change here. So this design can show that a tighter frame matters. It cannot show that a macro photograph is safe.

One thing to know before the numbers

At f = 0.35 many crops fall below the detector's 440-pixel short-side resize, so it upscales them. That happens identically in both arms of each pair, so it does not confound the comparison — but it does put the tight rungs outside the resolution regime the published baseline was measured in. The count below 440 is reported for every condition.

The corpus is rebuilt from the manifest

The first study pruned imgs_<set>/ to its flagged images to reclaim 1.94 GB, which is exactly the case the manifest was written for: every file carries its source URL and its sha256, so framing.py re-fetches on demand, checks the hash before use, and deletes the file again once its seven views are scored unless one of them flagged. Wikimedia is asked for one file a second.

base reproducing the stored clean score to zero difference is what proves the rebuilt corpus is the same corpus.

Results

598 images — 298 glossy, 300 control — through seven views each, 4,186 scored views in total. Every pre-registered bound passed, so what follows is a result and not a halt.

B2 is the one that matters most. base reproduced the first study's stored clean score for all 598 files, worst absolute difference 0. The corpus had been deleted; it was rebuilt here by re-fetching every original from the manifest and checking its sha256. That it scores bit-identically is what licenses comparing any of these numbers to the published ones. B3: zero hash mismatches in 598 re-downloads. B5: the detector's own gates skipped at most 1 image in any condition, against a 20% ceiling. B4: 21 discordant pairs at the primary rung, above the 10 required.

The primary test

crop(0.35) vs shrink(0.35) on the glossy corpus, 298 paired images, threshold 0.65:

2 images flagged in both arms, 275 in neither, and the 21 discordant pairs split 12 crop-only to 9 shrink-only. Exact McNemar, two-sided: p = 0.6636.

H1 is not supported. At the tightest rung, on the subject class that raised the question, a frame holding 35% of the scene is not flagged more often than the whole scene at the same pixel count. With 21 discordant pairs this is a null rather than an absence of power — but 21 is not many, and the interval around a gap this small is wide. What this rules out is a large effect, not a small one.

The ladder (secondary)

Glossy, crop vs shrink at each rung:

Control:

Flag rate by condition, with the count falling under the detector's 440-pixel short side:

The paired mean score shift is negligible everywhere: the largest is 0.0225 (control, f = 0.35) and every median shift rounds to 0.0002 or less. Whatever is happening at the threshold is a handful of images crossing it, not a corpus-wide drift.

The anchor

base flags 3.36% of the glossy corpus and 3.00% of the control — the same two numbers the first study published. The rebuilt corpus is the corpus.

One pattern I am not going to promote

crop is nominally higher than shrink in all 6 rung comparisons, and the smallest p anywhere in this study is control f = 0.50 at 0.0768. It would be easy to write that up as a trend.

It is not one. The three rungs are nested crops of the same photographs, so the comparisons within a corpus are strongly dependent — a sign test across them would be arithmetic wearing the costume of evidence. And 0.0768 sits on the control corpus, at a rung the hypothesis was not about, on a secondary test with no multiplicity correction across the twelve comparisons in the tables above. The pre-registered test is the one at the top, and it came back null. This gets the same treatment the earlier study gave its p = 0.046: reported in full, promoted to nothing.

If the direction is real, the design to catch it is a single pre-registered rung with several times this many images, not a re-reading of these.

What this does and does not say

It says: for these two Commons corpora, at this detector's shipped threshold, cropping to a third of the frame does not measurably change how often a real photograph is called AI-generated, and the subject class does not change that answer.

It does not say a macro photograph is safe. A centre-crop is not a photographer stepping closer — perspective, depth of field and the subject-to-background relation all change in the real act and none of them change here. It also does not say the detector is accurate: both corpora sit around 3% false positives at base, roughly one photograph in thirty, and nothing here moves that.

The bounty submission that started all of this still scores 0.8944. Neither study explains it.

Reproducing it

The corpora are the same two the first study used, and they rebuild from their manifests without my redistributing anyone's photographs — every kept file's Commons page, licence, camera, dimensions and SHA-256 is in glossy (298 files) and control (300). The full output, including the per-condition breakdowns, the bound checks and every comparison that did not support H1, is in framing.json.

The pipeline that produced it:

framing.py                  the pre-registration, the ladder, the paired tests, the bounds
run_framing.sh              rerun-until-clean wrapper; retries only signal deaths, measures progress
chain_analyze.sh            detached completion step, so the analysis does not depend on a watcher
checkprose.py               publish gate: every number on this page traced back to framing.json
scores_framing_<set>.json   one row per (file, condition): dimensions, logit, score

The pre-registration is in the docstring of the script that ran the test, the bounds are code that halts rather than prose that hopes, and every number on this page was matched back to the result file by a program before it went up — because the last time I wrote up results from memory of a log, half the figures in the prose were wrong.