Field note
Raindrop Signals 2 creates narrow binary classifiers from traces produced by live AI applications. A signal can ask whether an agent broke a rule, lost the user's intent, or misused memory, and the input may include multi-turn dialogue, tool calls, and subagent traces.
The system combines deterministic filters, a small semantic classifier, and selective frontier-model calls when uncertainty rises. Raindrop reports median latency around 100 milliseconds and costs up to 1,600 times lower than GPT-5.6 Sol at xhigh effort on its evaluated workloads. Those figures are vendor benchmarks.
The architectural idea is stronger than the headline number: expensive judgment should scale with ambiguity rather than raw traffic. A production monitor can handle common cases cheaply, escalate uncertain examples, and feed reviewed mistakes back into the classifier. That creates an observable drift loop instead of an unlimited LLM-as-judge bill.