Grassroots cultural heritage institutions must confirm that the item in front of a camera is the same item as the one in their register, yet they work with a few images per object, a single CPU-grade terminal, and no practical capacity to retrain a network whenever a new item is accessioned. This paper reports a digital identity solution designed around that constraint and evaluates it using a purpose-collected archival corpus rather than a public benchmark. Each object is described by a 1280-dimensional descriptor extracted using a frozen ImageNet-pretrained MobileNetV2 network with GeM pooling; the registration views of each object are summarised into an identity template, a query is matched using cosine similarity, and an acceptance threshold, together with mandatory manual review, determines whether an identity match is confirmed. Identity is stored in a structured record containing the identifier, the template, collection metadata, the model and template version, and an integrity hash. The corpus contains 42 objects and 435 images in four object groups, of which 35 objects are registered and 7 are used as unknowns, with registration and query images separated according to their acquisition conditions; a perceptual-hash audit found no near-duplicate pair across that split. Closed-set top-1 accuracy is 67.2% and top-5 accuracy is 92.7%. Two principal findings structure the discussion. First, the same objects matched under an identical acquisition condition achieve 77.7%, whereas the identical pipeline achieves only 67.2% across different acquisition conditions; this 10.5-percentage-point gap is not eliminated by backbones containing up to 85.8 million parameters, by metric-learning heads trained with triplet and relational-distillation losses, or by colour normalisation and re-ranking, so the principal limitation is the acquisition condition rather than the capacity of the model. Second, the quantised model occupies 2.49 MB and runs at a median latency of 4.14 ms with a 1.1-point accuracy loss, meeting the deployment target without any training stage. We therefore argue that, for small archival corpora, engineering effort should be directed towards acquisition protocols and record design, and that an open-set equal error rate of 33.1% requires the system to be specified as a candidate-ranking aid for human review rather than as an autonomous identity decision-making system.