CrazyJev Lab
Reaction methodology: what each millisecond actually measures
The Reaction Lab starts a human response test and a Jev request from one application signal. It shows the original measured times alongside an adjustable hypothetical delay. The hypothetical line does not call a chat model and is not a competitor benchmark.
What each clock measures
| Value | Start and end | Included |
|---|---|---|
| Your response | App schedules GREEN → first accepted input | Display scheduling, perception, input and event handling |
| Jev request | Browser starts the request → response received and parsed | Browser/server transport, server handling, provider call and retries |
| Server-to-provider call | Server starts calling the provider → returns a result | Transport, provider processing and server retries; not pure inference |
| Hypothetical delay | User-selected value | No measurement and no actual model call |
The app schedules the GREEN state and dispatches the request in the same callback. It does not have hardware timestamps for when pixels reach your eyes. Network and display paths are different. This is an interactive demonstration, not a controlled human-versus-machine benchmark.
The game handicap is a different number
The original Reaction game offers ÷2, ÷4 and ÷8 difficulty targets. Those are disclosed game handicaps applied to request time. They are not estimates of one-way network time or inference latency. The Lab's comparison chart uses the original request time without a divisor.
Reproduce a small test
- Use the same browser, device and network; close background workloads and keep the tab visible.
- Start a round and wait for GREEN. Click or press Space once. An early input is a false start.
- Wait for the live Jev result. Do not discard a slow or failed round.
- Repeat and record sample count, false starts, missing player inputs and failed requests.
- Report the distribution, not just the fastest attempt. The interface shows the most recent ten finished rounds for this page session; false starts are counted separately for the session.
The median and p95 use the nearest-rank method on successful request durations in that visible window. With ten samples, p95 is the largest observed value. These small-window summaries are descriptive, not a population estimate. Errors remain in the finished-round denominator but do not get invented latency values. A valid model response that declines to press is not a winning reaction.
Comparing another model honestly
A real comparison needs a named provider/model, equivalent input and output rules, region, timeout, retry policy, concurrency, request count and collection date. Prefer interleaved trials and include failures. Publish raw sanitized samples and explain which transport costs are included.
Until that experiment exists, the slider only answers: what would this timeline look like if a response took the selected number of milliseconds? It does not establish how fast a particular model is, or how much more useful Jev would be in every application.