Learned Eyes, Deterministic Judgment

A perception model is very good at answering “what is in this image, where is it, and how big is it.” It should never be asked “and is that acceptable?” Those are different questions with different failure modes, and our position is that they belong in different layers of the system, with the boundary between them drawn explicitly rather than left to emerge.

The number that moves

The public benchmark for this problem is ELPV: 2,624 solar cell images at 300 × 300 pixels, extracted from 44 modules — 18 monocrystalline and 26 polycrystalline — each annotated with a defect probability^[1]. The paper that introduced the classification task reported 88.42% accuracy for a CNN and 82.44% for an SVM pipeline on the binary functional-versus-defective task, training on 1,968 of those cells and testing on 656^[1].

Now look at what happens to that number in other hands. A 2023 study benchmarking several architectures on the same binary ELPV task reports 78.38% for a plain CNN, 84.01% for VGG16, 87.61% for DenseNet121, and 96.17% for its own proposed model^[2]. A 2022 study reports 98.15% on the binary task^[3], and a 2025 study reports a peak of 90.13%^[4].

On nominally the same task and the same dataset, published accuracy spans roughly twenty points. The ordering is not even monotonic in publication year.

The 2022 result is the most instructive, because the authors are candid about where it comes from. The same paper reports accuracies of “48.39% and 87.67% for Inception-V3, 52.68% and 91.60% for ResNet50, finally 98.15% and 95.35% for combined features… before and after applying data augmentation”^[3]. A preprocessing decision moves the headline by fifty points.

We want to be careful about what this does and does not show. These are not all the same task — some of those multi-class figures are severity bins, others are defect-type classifications, and comparing across them would be sloppy. The binary functional-versus-defective figures are comparable, and those alone span 78.38% to 98.15%. That is the honest version, and it is enough.

Why that is an argument about architecture

A model whose reported accuracy moves twenty points with the choice of backbone, and fifty with the choice of augmentation, is not a thing you want holding a quality criterion. Not because it is bad — 96% on a hard visual task is genuinely impressive — but because the number is a property of an experimental setup as much as of the world, and an acceptance criterion has to be a property of the product.

The field is candid about the boundary of what has been established. A 2025 review of EL diagnostics concludes that while transfer learning and attention mechanisms have proven effective, “there is a need for greater generalization across defect types and datasets”^[5]. We could not find a study that trains an EL defect model on one dataset, evaluates it on another, and reports the degradation — the experiment you would most want before trusting a published accuracy figure to describe your own line. Absent that, we assume our distribution is not theirs.

The split

So we separate the system into three layers, and we are strict about what crosses each boundary.

Perception is learned. It detects, classifies, segments and measures, and it emits those measurements with confidence attached. It does not know what a good module is and is never asked. Judgment is deterministic: it takes measurements and applies the quality criteria, and it emits a disposition together with the criterion that produced it and the evidence that satisfied it. Attribution is statistical and comes last — trends, correlations, where to look next — and it deliberately has no authority over the unit in front of the camera.

There is precedent for this shape in the standards. The IEC technical specification for module electroluminescence sets out methods to capture images, to “process images to obtain metrics about the images taken in quantitative terms”, and to “provide guidance to qualitatively interpret the images”^[6]. Capture, measurement, interpretation — three separable activities, and the specification itself sets no pass/fail threshold. The thresholds are somebody’s decision, made deliberately, and the standard declines to make it for you.

What this buys: a disposition you can replay

The practical payoff is auditability. Because judgment is deterministic, every disposition decomposes into three artefacts: the image, the measurement extracted from it, and the criterion applied. Any decision can be replayed. If a criterion changes, we can re-run history and see exactly which units would have been dispositioned differently, without retraining anything. If a measurement looks wrong, the argument is about the measurement, not about the disposition.

With end-to-end learned grading, none of that exists. You have a score and a saliency map, and the saliency map is not the reason — it is a post-hoc story about the reason. The argument against relying on such explanations is uncomfortably tight: if an explanation “was completely faithful to what the original model computes, the explanation would equal the original model, and one would not need the original model in the first place, only the explanation”^[7]. Explanations of a black box are approximations by construction.

We should be precise about how far that argument reaches. It is made primarily about structured data with meaningful features, and its author is more cautious about raw computer vision — which is exactly our perception layer. We are not claiming the perception model should be interpretable. We are claiming the decision should be, and that the way to get there is not to explain a black box but to keep the criterion outside it. On the same view, the claimed trade-off is overstated in the first place: “It is a myth that there is necessarily a trade-off between accuracy and interpretability”^[7].

The evidence that transparency changes outcomes

There is at least one controlled result in our own setting. In a preregistered experiment with experienced factory workers doing visual defect inspection, workers supported by black-box AI achieved balanced accuracy of 88.6% while workers given an explainable version reached 96.3%, which the authors describe as “a five-fold decrease in the median error rate”^[8].

The detail that makes this relevant rather than merely encouraging: the AI’s prediction was identical in both arms. Only the explanation was added. So this is not evidence that interpretable models are more accurate — it is evidence that a person given the basis for a machine judgment catches machine errors that they otherwise accept. Which is the same argument as the layer split, arriving from the human side.

What we give up

This architecture is more work. An end-to-end model can be trained on dispositions alone; ours needs measurements to be meaningful in their own right, which means defining what is measured before the model exists. It also forecloses a category of gain: a learned grader can exploit correlations in the image that no explicit measurement captures, and we will not have those.

We think that is the right trade for a decision that has to be defended rather than merely made. A number that swings twenty points across published papers on the same benchmark is a fine detector and a poor judge. We would rather build the judge ourselves, in code we can read, and let the model do the thing it is genuinely extraordinary at.