# Quickstart: your first caught bug (60 seconds)

The fastest way to understand Neurarch is to watch a check fire. This walkthrough plants a classic bug and shows you the catch. No account, no install; everything runs in the browser.

## The bug

Multi-head attention with `embedDim: 768` and `numHeads: 10`. It reads fine in code review. But 768 / 10 is not an integer, so PyTorch raises `AssertionError: embed_dim must be divisible by num_heads` when the module is constructed, typically on the training node, after the job was queued and the dataset downloaded.

## The path

1. Open <https://neurarch.com/?example=attention-heads>. A small transformer with the planted bug loads onto the canvas.
2. The pre-flight report opens on the finding: a blocking shape check, `embedDim % numHeads != 0 (768 % 10)`.
3. Close the report. The attention layer carries a red issue dot: hover it for the explanation and the suggested fix; click it to jump to the exact finding in the Architecture Advisor.
4. Change the head count to 12 (or any divisor of 768). The health badge in the stats bar turns green.

A second planted example, a Linear layer whose `inFeatures` does not match its upstream width, is at `https://neurarch.com/?example=linear-width`. That one is the bug class simple linters miss; it fails on the first batch with `RuntimeError: mat1 and mat2 shapes cannot be multiplied`.

## After the catch

- **Build something real**: describe your task in the prompt box (agent features need sign-in), pick a task in the 3-question wizard, or load one of 48 templates. See [overview](/docs/overview.md).
- **Check your own model**: paste PyTorch code, an arXiv link, or a HuggingFace id and the graph is reconstructed with shapes inferred. See [import and export](/docs/import-export.md).
- **See every rule**: all 50 checks with trigger conditions are public at <https://neurarch.com/rules.html>, and the measured evidence they track reality is at <https://neurarch.com/caught-bugs.html>.