Interleave: make concurrency bugs visible, then test the fix

[pralav-25]

Two workers increment a counter, but the final value is 1 instead of 2. Interleave lets you control the execution and see exactly where the update disappears.

Open a failing replay · Source and setup

Try it

1. Open the replay and inspect each worker's reads and writes.
2. Rewind a step and choose a different worker to explore another execution.
3. Select With the fix, then explore the terminal schedules in the model.
4. Share the replay or export a Markdown trace so someone else can reproduce it.

What is included

• Six experiments: lost updates, oversold inventory, double payments, stale search responses, deadlocks, and write skew.
• Deterministic state transitions, explicit invariants, and exploration of the finite built-in models.
• Optional private saved investigations and notes after sign-in.
• Optional local AI assistance for discussing the selected trace. It requires WebGPU and roughly a 1 GB initial model download; the core lab works without it.

The lab models two workers with finite steps and sequentially consistent memory. It is an educational concurrency tool. The optional model's real GPU inference quality and speed have not yet been validated; see the AI validation notes.

Feedback welcome: which experiment helped you understand the bug, and which small concurrency scenario should come next? A replay link makes unexpected behavior much easier to investigate.

Architecture and model boundaries · Contribution guide