My blur test came back inconclusive. The naive version of it came back confirmed.

The inconclusive result is the boring half, and I will show my working on why it is inconclusive rather than negative. What is worth your time is that the naive version of the same analysis — same 140 images, same two variables, one design decision changed — returns -0.254, comfortably past the bound I had set in advance. I would have published it.

The rival hypothesis, named before the test

Bellingcat's 2023 evaluation of a different detector found six false positives in twenty competition photographs and attributed them to sharpness, fine detail and large depth of field. That is the opposite sign from mine. Naming it in advance is what made the study worth running at all: two hypotheses that disagree about the sign of one number can both be wrong, but they cannot both be confirmed by whichever way it happens to lean.

The design

Variance of the Laplacian — the standard sharpness statistic — against the detector's score, over the 140 real photographs of my own corpus under the clean condition, scored by ft44s-2026-08-19 at threshold 0.65.

The whole design turns on one nuisance variable. Those 140 photographs come from 14 source datasets, and the datasets differ enormously in both blur and score: tight face crops are sharp and score low, flat textures are sharp and score high. A correlation computed across the pool would mostly be measuring which datasets I happened to include. So the estimate is within source — rank both variables inside each dataset, then correlate — and the interval is bootstrapped over the 14 datasets, not the 140 images, because the images nest inside them.

All of that, plus the decision table below, was fixed in PREREG-blur.md before a single pixel was measured.

The result

It leans the way I guessed, it is smaller than the bound I set, and the interval crosses zero. Verdict: inconclusive.

The number I would have published

-0.254 against -0.127 is the entire argument for pre-registering the estimator and not just the hypothesis. Both are computed from the same 140 images and the same two variables. One of them clears my bound. The difference between them is a decision — whether to hold the source dataset fixed — that I would have been making after seeing both numbers if I had not written it down first.

I have made the version of this mistake that ships. This is the one that did not.

Does a null from this design mean anything?

At 14 clusters the honest worry is that this design could not detect anything, in which case a null says nothing about the world. So, after the result and therefore not in the pre-registration: inject a known effect into the real blur measure and see how often the interval notices.

The top row is the calibration check: with no effect present the interval wrongly excludes zero 7% of the time, against the 5% it advertises. The row at -0.20 is the power at the size I pre-registered: 57%.

So the instrument does not clear the bar, and the correct reading of the primary result is that this design is too weak to answer the question — not that the answer is no. That is a worse outcome than a clean null and it is the one I have.

The board itself

Ranking all 17 submissions by the same blur measure: the winning photograph, at 0.9588, is the 9th blurriest of 17. Mid-pack. The blurriest photograph anyone sent me was not flagged at all.

That is descriptive and nothing more: 17 entries with 3 positives, two of which are one person photographing one dog, is not a sample anything can be inferred from, and no p-value is computed on it anywhere — PREREG-blur.md declared this study descriptive-only in advance, precisely so a suggestive-looking rank could not be promoted to a result after the fact. It is here because it is the population my hunch was originally about, and because the picture that started the hunch turns out to be unremarkable on the very axis I formed the hunch from.

What this does and does not say

It does not say blur is safe, and — this is the part I would rather not be writing — it does not say blur is irrelevant either. What it says is that across 140 photographs from 14 datasets, with the dataset held fixed, any relationship between blur and this detector's score is small enough that a design with 14 clusters cannot resolve it. The honest summary of the primary result is not measured, not measured and absent.

So blur does not join glossy surfaces, tight framing and web recompression on the pile of explanations this board has actually ruled out. It goes on a different pile: things I tried to measure and could not, at this corpus size. Getting it onto the first pile needs more source datasets — 14 is the binding constraint, not the 140 images — and that is a different piece of work than the one I did here.

Files

Everything is in the bounty page's bundle: PREREG-blur.md (the bound, written first), blur.py (the analysis), blur.json (its output, including the verdict field the bounty page's prose is generated from), blur_power.py and blur_power.json (the instrument check, written after the result), and scores_clean.json (the scored corpus).

The 140 images themselves are not in the bundle. The corpus includes face datasets and a mugshot set, and republishing those is not mine to do. What is shipped instead is the derived per-image measures — blur_measures.json for Study A and blur_measures_subs.json for the board — which are the complete input to every statistic above. blur.py prefers the images when they are present and falls back to these, so the analysis reproduces from the bundle alone; anyone holding the source datasets can recompute the measures from scratch, because every row names its file.