Interestingly, a tuned length-threshold baseline performs very competitively and even beats Wayfinder on one of the benchmark sets. My intuition is that structural signals become more useful as prompt formats get more heterogeneous (code reviews, logs, tables, long instructions, etc.), but I’d love to see more real-world data.

Most routing systems I found either used another model as a classifier or relied on provider-specific routing. I was interested in the narrower question: how far can deterministic heuristics get before you actually need another model in the loop?