OOD research, evaluating best function to identify OOD

I Tested 4 OOD Detectors on a Real Model Failure — the Baseline Won Can a CNN know when an input is foreign? In my previous research, a model that scored 93% on its European benchmark (EuroSAT) detected only 19% of real Amazon deforestation — it labeled dense jungle as "SeaLake" with confidence 0.83, and nothing ever warned that the inputs were out of distribution (OOD). This study attaches four known OOD warning functions to that same model and gives them a real exam: separate 4,050 home images (the EuroSAT test split) from 289 foreign images (real Amazon patches), with every grading rule fixed before the results were seen. Result alarm separation caught% confidence 0.911 54.0% mumble (energy) 0.854 49.5% center distance 0.864 31.8% knn distance 0.775 19.4% The ranking inverted my written predictions almost perfectly. I predicted knn best (0.89) and confidence weakest (0.55); reality placed confidence first and knn last. The predictions were committed in ink before grading, so the reversal is a measured result, not hindsight. The verdict readers beat the map readers. The alarms that read the model's scores won; the alarms that measure distances on its internal 2,048-dimensional feature space lost — consistent with the idea that the network's own measurements hide the foreignness of foreign inputs. The honest asterisk: at a threshold that leaves 95% of home images alone, even the winning alarm catches only 54% of the foreign pile. No alarm here is deployment-ready; the exam measured how far from ready. The exam Trained ResNet50 (93% on EuroSAT) → HOME pile = EuroSAT test split (4,050 images, same split seed as the original training — never trained on) vs FOREIGN pile = 289 Sentinel-2 Amazon patches (foreign by construction) → each alarm assigns every image one danger score → graded by separation (probability a random foreign image out-scores a random home image) and caught% (threshold set from the home pile only — no tuning on foreign data). Files File What it is OOD_research.pdf Full paper (two-column) with figures notebook.ipynb All code: piles, signal collection, 4 alarms, grading, histograms eurosat_resnet50.pt The trained model under test Honesty note Predictions were written down before any grading ran. The threshold rule and both grading metrics were fixed in advance. The center function is the simple (non-shape-aware) version by design — this benchmarks what a practitioner can cheaply attach to an existing model. AI assistance was used for debugging, verification, and language editing; all experiments, predictions, decisions, and findings are my own. Related Part 3 of a research arc on benchmark-vs-reality collapse: Part 1 — tabular robustness study: [link to repo 1] Part 2 — the 93%→19% deforestation deployment study: [link to repo 2]