Experiments you can replay.
Each lab is run once against the live APIs and recorded — every request, every response, every millisecond. What you open here is that recording, replayed. No key on this site, no live calls, and every visitor sees the same evidence.
- 1 · Record — Run live against the real APIs, once.
- 2 · Commit — The recording is a file in the repo. Never edited.
- 3 · Replay — This page plays it back at recorded speed.
Jev vs. an LLM on a decision: which one gives an answer you can act on, how fast, and how honest is its confidence?
- Jev 1.13
- 185 ms
- median, 30 calls
- GPT-4o-mini
- 1128 ms
- median, 30 calls
The same request and the same toolkit prompt on workbooks shaped like real finance exports: subtotals that include intercompany, negatives in parentheses, a renamed region, units in a note, blank cells for a new region.
- GPT-6 Luna
- 40086 ms
- median, 40 calls
- Gemini 3.8 Flash
- 63029 ms
- median, 40 calls
On messy enterprise data, what buys more accuracy per dollar: describing the data, or buying a bigger model?
- Jev 1.13
- 266 ms
- median, 135 calls
- GPT-6 Luna
- 4016 ms
- median, 135 calls
A refund agent meets an injected instruction. Does a permission gate stop it, or labels on what the agent may trust? Which plane stops which failure?
- Jev 1.13
- 189 ms
- median, 320 calls
- GPT-6 Luna
- 3616 ms
- median, 320 calls