Typed judgments or agentic loops?

Probably generous, and the author says so. The old run used 5 threads while the new one ran sequentially, so contention or upstream throttling could have inflated the old side. The number that survives the caveat is per-call, 1.30 s against 0.43 s, which does not depend on concurrency.

Not always. On the same 50 products, 36 paths were identical, 2 were a prefix of the other and 12 diverged. Those 14 were hand-judged by a single annotator with no ground truth, which the post calls a signal rather than an accuracy measurement, and the agent still won some close branches.

Twelve of the 50 old paths stopped mid-descent when the budget ran out rather than being sent back and fixed. The honest reading is that Jev gets faster on out-of-taxonomy and adversarial cases partly by giving up on them, which the post calls a time saving that is also a renunciation.

Cheaper, not free. Going from 1 to 7 questions per request held response time flat at roughly 0.32-0.39 s with identical answers on eight products, cutting calls 11% but adding 25% tokens. Overall output tokens rose 383% because every Choice returns a probability per criterion, which erodes the 36% token saving if output is priced high.

Walking a product description down a category taxonomy, Amazon and Shopify, to land on a leaf node, run over the same 50 products, taxonomies and Postgres on both sides.

The same request carries the question about the current level plus questions about the children of each candidate, so the descent drops two levels for the price of one call; an A/B with threshold 8 kept every path identical.

The tail is the operational change: Jev's p90 was 1.91 s and its worst case 3.68 s against an agentic worst case of 19.3 s, and all 50 products ran sequentially in 68.8 s.