What this is worth testing
Your agent keeps the larger task and any action that follows. System One asks Jev for a typed answer on this check.
- Research value
- HighEvidence confidence is moderate. This rating describes the research record, not a live Jev probability or a measured saving.
- Sample size
- About 170 input tokensThe recipe sample length divided by four. This is a planning estimate, not a measured tokenizer count.
$0.07 at Jev's $0.042 per million list rate
$5.10 at a $3 per million example rate
$5.03 input-cost difference for 10,000 uses of this sample
Excludes output charges, host tool calls, retries, hosting, and paid plans. Equal input volume does not establish equal answer quality or savings on a fixed-price subscription. $3 per million is an example rate, not a quoted model price.
When an agent uses this
System One asks Jev for a typed answer. Your agent keeps the larger task, the original evidence, and any action that follows.
Example evidence
This is the recipe sample. Replace it with the evidence from your task. It is not a measured result.
Original query: how does engine handle errors? Result: 0 matches in candidate files; corpus has technical TypeScript definitions.What your agent keeps
Advice only. The agent generates the concrete query string using its reasoning model and domain knowledge.
Finding
DR-DCI (arXiv:2606.14885) and Bergum (2026) demonstrate the unreasonable effectiveness of tuned BM25 and grep for agentic search trajectories. Jev provides bounded micro-evaluations without expensive context ingestion.
Research value: high. Evidence confidence: moderate. That rating describes the research record, not a live Jev probability.
Try it
Connect System One, then ask your agent for recipe corpus-query-reformulate. Inspect the input on the recipe page before you send your own evidence.