A Vision Model That Measures Catches and Catches Cheaters

Competitive fishing runs on measurement and honesty, and both are hard to enforce from a phone photo submitted in the field. Two questions have to be answered at once: how long is this fish, really, and has the image been manipulated to make it look bigger or to reuse an old catch? Human judges cannot scale to thousands of live submissions, and simple rules are easy to game. This is a measurement-and-verification problem, and it is exactly the kind of narrow, high-value task a small, purpose-built model is good at.

We are building a compact vision model — small enough to run on-device and in real time — that estimates the true measurement of a catch from an image and, in the same pass, looks for the signatures of tampering and fraud: edited pixels, inconsistent scale references, and reused or staged submissions. We are deliberately not publishing the model architecture, the training data, or the detection heuristics, because the value of a cheat detector is precisely that cheaters cannot read the manual. What we will say is that it is the same discipline as the rest of our work: a measured base model, a held-out verification gate, and a bias toward small and local so results are fast and private.

Measure-a-thing-from-an-image and flag-when-the-image-cannot-be-trusted is a general pipeline, and tournament integrity is a forgiving place to harden it before the stakes get higher. The same approach — precise measurement plus tamper and anomaly detection, running small and on-device for privacy — maps directly onto medical imaging, where measurement accuracy and image integrity are matters of patient safety and where data cannot leave the building. Fishing is where we prove the pipeline in the wild. Medicine is where it goes next.