Deep Searcher includes a frozen-query evaluation of Jev as a search-stopping judge; this experiment is separate from the full research agent.
Jev judges enough_information from the main query and accumulated intermediate context; code compares stopping decisions under a bounded round protocol.
Makes stopping behavior and call traces inspectable; the protocol explicitly does not score final-answer quality.
Fixed-version documentation and implementation material were inspected on 2026-10-24. No 100-query comparison or live API was run. The evaluation directory does not prove the main application enables this path by default.
Evaluate when a research search loop has enough information to stop.
Material analysed：
The manifest protocol explicitly excludes final-answer quality scoring. Matching another model’s stopping time cannot establish that the resulting answer is correct or complete.
Inspect the supporting source ↗full100_core.py hashes URL/body, stores call observations and failed responses, and separates request-rate gates. Cached responses must not be mistaken for newly measured latency.
Inspect the supporting source ↗Choose this as a stopping-policy experiment, not a drop-in replacement for Deep Searcher’s entire research agent. Measure downstream answer quality before deploying a new stopping threshold.
Inspect make_manifest in run_full100.py and DurableAPI.request in full100_core.py. Confirm the frozen question set, seven-round cap and cache markers, then distinguish a decision trace from a scored final answer.