JAX PathwaysCheck your starting point
Two simulated examples receive the same random key and run the same sampling function. Should you expect independently allocated draws?
Compare your reasoning
No. The same key and operation replay the same draw in the tested environment. Split or derive distinct keys at the ownership boundary; unequal observed samples alone do not prove independence.
If this is unfamiliar, work through Random keys without surprises before attempting the project.
Build in stages
1. Name and structure the state
Draw the key ownership tree and identify parameter, optimizer and metadata leaves. Check both tree structure and leaf values.
2. Audit repeated and conditional transitions
Implement a recurrence with fixed carry shapes. Save per-step observations as outputs, compare with a Python loop, then test both conditional branches and their boundary.
Try a changed condition
Make a prediction before opening the reasoning. Keep your answer and the evidence you would collect.
Resume a simulation halfway through with the saved position but a newly created key. What must you compare to determine whether it is a true continuation?
Compare an approach
Compare the next draw and next state, then the remaining trajectory, against an uninterrupted reference. Restore the original key and counter as well as position. A similar final average is not a continuation check.
Diagnose a likely failure
Symptom: A compiled loop fails when its history grows each step.
Check next: Keep carry structure, shapes and dtypes fixed. Return the per-step history through scan outputs.
Decide what is ready
Use foundation-toolkit stage 4. Save and reload state in a fresh process and verify the next transition and full remaining trajectory, not only the final mean.
Where to put your project code
Write your code in: projects/foundation-toolkit/my_toolkit.py (copy projects/foundation-toolkit/starter/toolkit.py first: cp projects/foundation-toolkit/starter/toolkit.py projects/foundation-toolkit/my_toolkit.py).
Verify your implementation: python3 projects/foundation-toolkit/tests/check.py --stage all --impl projects/foundation-toolkit/my_toolkit.py
Run the staged project checks on your implementation in projects/foundation-toolkit/my_toolkit.py and record your transfer experiment results in your portfolio.
What still needs further work
Local deterministic state is separate from concurrent-writer or multi-host recovery; the recovery and operations phases extend these boundaries.