I had 14 recorded team calls. They contained 30,143 things people said. Some of my instructions for work in progress were buried in there, and the first agent reading the calls had missed 10 of them.
I gave Jev a smaller job: look at a few lines at a time and tell me whether I was giving an instruction, making a decision, expressing a preference or rejecting something. Jev marked 7% of the short passages for a full read. The scan took 108 seconds and cost $0.32. Reading those marked passages surfaced the 10 missed instructions.
If your startup records calls, this is the useful part: you don't need another summary of everything everyone said. You need a reliable way back to the moment someone said the thing your agent must now obey. Jev helped me build that reading queue. It did not decide what the instructions meant.
Why search missed them
Keyword search is good when you know the words to type. It is less helpful when a decision was spoken in different words, between many mentions of the same topic.
In my calls, words such as “video,” “photo” and “keyword” appeared again and again. A keyword hit sent my first reader to plenty of relevant talk, but it did not reliably pick out the sentence where I actually made a decision.
Jev asked a different question of each short passage: is the person I care about making a decision here? It could mark a rejection even if I never said the project's usual keyword.
| Search method | Share of passages sent for full reading | Known instructions it touched |
|---|---|---|
| Keyword list | 15% | 49 of 54 |
| Jev | 7% | 51 of 54 |
| Both together | The union of both reading lists | 53 of 54 |
The methods did not find exactly the same things. Keywords alone found 2 Jev missed; Jev alone found 4 keywords missed; 1 needed a full-call read. So I would keep both reading lists. I published the measured pattern and its limits with the Jev skill.
How to give Jev a meeting archive
A whole meeting is too much for 1 narrow question. I cut the transcripts into 7,562 overlapping passages. Each held about 6 consecutive utterances, roughly 40 seconds, and shared 2 utterances with the next passage. The overlap kept a decision from falling between 2 cuts.
Before sending them to an external model, I relabelled speakers as Founder and Other and scrubbed personal names. The saved input looked like this:
{"id":"call-07#103","passage":"Other: ...\nFounder: I want the next page to show the video library.\nOther: ..."}That line is a format example, not a verbatim transcript. The real inputs were scrubbed and remain private.
Jev answered a focused yes/no question about each passage:
{
"decision_here": {
"type": "noul",
"instructions": "Does Founder give an instruction, make a decision, express a preference, or reject a proposal about <the project>?"
}
}Replace <the project> with a short description of your own topic. The public transcript recipe also asks Jev which part of my directory the passage concerns. Rewrite those options for your work. Do not copy my hotel categories into your product.
What I read, and what Jev read
Jev evaluated all 7,562 passages. I selected passages scoring at least 0.3, merged neighboring hits and read the original transcript around them. That gave 544 marked passages, about 7% of the total. Among passages my first reader had not yet read, Jev's flags led to 10 missed instructions.
The 0.3 line was chosen to catch more possible decisions, even if it meant reading extra passages. It is not a universal “safe” score. On a different archive, I would first mark a few decisions I know are present and check that the chosen line finds them.
There is another detail hidden in the cost claim. Jev still read all the short passages. The saving was in human or agent reading time: a reader examined the marked sections in full instead of trying to reread every call. The $0.32 was the measured Jev input cost for this run, not a promise that your archive will cost the same.
Hand this to your agent
The Jev skill and transcript recipe give your agent the question shape and CLI. This is the job I would give it:
I have an archive of calls, tickets or chats and need to recover decisions on one topic. First identify the speaker or role whose decisions matter. Build overlapping passages of about 6 utterances, with 2 utterances shared between neighbors. Scrub names and secrets before sending anything to Jev. Ask 1 yes/no question per passage about a decision, instruction, preference or rejection on the topic. Keep a low review threshold, then read every flagged passage in its original context. Run keyword search alongside Jev and read the union. Before reporting coverage, check both methods against at least 5 decisions I already know are in the archive. Give me the exact source passage for every recovered instruction. Do not turn a model label into an instruction without reading the passage.
The output I want is a ledger: decision, exact passage, date, speaker role, current status. A list of “themes discussed” would not tell my agent what to do next.
Where this result stops
This was 1 archive about 1 project. The same first reader who built the reference list of 54 instructions had already used keyword search, so the benchmark is not a clean independent contest. The source also does not establish whether the 10 newly surfaced instructions are included in those 54. I report them as 2 observations, not as numbers to add together.
Automatic transcripts contain recognition errors. A model can miss a decision because the words were transcribed badly. Jev also missed 2 instructions the keywords found. The combined list found 53 of 54, and the last one required reading the whole call.
Why I would save this pattern
The next time an agent tells me “I read the meetings,” I want proof that it found the decisions that matter, with links back to the exact moment they were spoken. On this archive, Jev turned 30,143 utterances into a much smaller reading queue in 108 seconds for $0.32, and that queue surfaced 10 instructions an earlier pass missed.
Start with a question you can recognize in a short passage, keep the original words attached, and test recall against decisions you already know. Jev can help you find where to look. Your agent or a person still has to understand what was decided and whether it is still current.