System One: Hill-Climbing Multimodal Embeddings from 77% Zero-Shot to 99% Calibrated Classification in 30 μs
How a frozen multimodal encoder (google/embeddinggemma-2) paired with calibrated linear probes,
Wilson statistical gates, and leave-one-group-out novelty floors climbs from 77.4% zero-shot to 99.0% Top-1
(100% answered accuracy, 100% OOD vegetable rejection) in a 241 KB binary file running in 29.8 μs on CPU.
Includes the modality-gap collapse trap, few-shot curves, and deploying natively on Cloudflare Workers AI.