Tools

The arithmetic, on your numbers.

Each of these settles in about a minute something teams argue about for a quarter. No sign-up, no wall, and nothing you type leaves your browser.

Agent reliability
What your per-step number actually delivers

Reliability multiplies across steps, so a chain of individually good steps is a bad system. Enter your steps and per-step reliability to get the end-to-end figure, the per-step number your target really requires, and what a tree of individually capped runs can cost.

Answers: At 95 percent per step, what does a 20-step task deliver?
Edge and latency
Is your latency bytes, or arithmetic?

On fixed hardware the cost of running a model is moving weights and activations across a bus, not doing math. Enter your model size, precision and bandwidth to get the memory-bound floor, how much of your measured latency it explains, and how much silicon is sitting idle.

Answers: Does 7B at 8-bit fit a 40 ms budget at 200 GB/s?
Evaluation
Can your test set detect a shortcut?

A random split measures how much your test data resembles your training data. Enter your row count, group count and the largest source's share to see what that split does with it, and what your effective sample size really is once you split by group.

Answers: If one speaker is 70% of the data, what does a random split prove?
Biosignal front end
Can your reference amplifier survive its own actuator?

The current you inject is a loop, and it comes home through the electrode you added to protect your front end. One multiplication and one comparison tell you whether the amplifier can absorb it or rails and takes your signal with it.

Answers: Does 0.8 microamps through 1.2 megohms clear a single-supply rail?