Background. Middle and external ear disease is common and largely treatable, but accurate otoscopic interpretation requires training that is scarce outside specialist ENT services. Published deep learning models for otoscopy report high accuracy, yet most are optimised for closed set classification rather than the referral decision a non specialist actually makes, report uncalibrated probabilities, and do not report the near duplicate controls that video derived otoscopic datasets demand. We developed, validated and deployed a screening aid framed around referral rather than diagnosis. Methods. We merged 4,763 otoscopic images from 16 folders across three public datasets into 12 clinically meaningful classes. Every image was passed through a single field of view normalisation function (crop to the illuminated field, square, resize to 224 px, mask to one fixed circle) shared identically by training and inference, to remove the framing cue that otherwise identifies the source dataset. Because one source is video derived, images were grouped by cosine similarity of generic ImageNet features (threshold 0.93, joined transitively) and whole groups assigned to a single split, giving 3,326 training, 719 validation and 718 test images. A MobileNetV2 backbone initialised from ImageNet was fine tuned from block9expand. Probabilities were averaged over the eight dihedral transforms of each image and temperature scaled. The referral score was one minus the probability of a normal ear, and its threshold was set on the validation abnormal ears as the highest value still meeting a target sensitivity of 97%, applied with an inclusive comparison. Model selection, calibration and threshold were frozen before the test split was scored once. Results. On the 718 held out test images, referral sensitivity was 97.0% (95% CI 95.1 to 98.2) and specificity 75.6% (95% CI 69.4 to 80.8), with AUROC 0.974 for abnormal versus normal; a cluster bootstrap over the 363 similarity groups in the test split widens those intervals to 93.8 99.6% and 66.0 83.4% respectively. Of 501 abnormal ears, 15 were not referred; 11 of these were myringosclerosis and none was cholesteatoma. Twelve class macro F1 was 0.686, balanced accuracy 68.3%, and accuracy 85.0% (86.2% with chronic subtypes merged). Temperature scaling (T = 1.2) reduced expected calibration error from 0.0498 to 0.0430. A low confidence flag fired on 4.0% of test images, separating 87.2% accuracy on unflagged from 31.0% on flagged cases. Recall was consistently highest on the largest source dataset and fell on the other two, for example acute otitis media 95.5% on source A versus 50.0% on source B. Conclusions. A compact, calibrated model run on ordinary CPU hardware achieved the pre specified 97% referral sensitivity with clinically interpretable failure modes. The source stratified gap, and an earlier perceptual hash split that inflated scores until it was corrected, indicate that headline accuracy in this field is strongly dependent on leakage control and acquisition style. The system is deployed with mandatory specialist review, a daily backtest and approval gated retraining, so that prospective performance on local equipment is measured rather than assumed. It is research software and not a medical device.