Save the experiment, not just the waveform.

Updated 18 October 2026

Put a small manifest next to every run.

A waveform shows what happened in one simulation. Reproduction also needs what produced it. Capture this information automatically:

• Source: DUT and testbench revisions, plus a hash or saved patch for local changes.
• Configuration: widths, parameters, register expectations, compile flags and exclusions.
• Environment: simulator, language runtime and dependency versions; a container digest or lock file when used.
• Invocation: exact test selection, seed, command and project watchdog limits.
• Finding: failed check, sampled cycle, accepted-transaction context, logs and waveform location.

Keep the original failure.

First replay it unchanged. Then make the fix and rerun the same case before broadening the regression. If you minimize the test, preserve the original bundle too: a smaller reproduction is useful, but it is a new experiment.

Random seeds are only one input. Simulator scheduling, library changes and an updated driver can alter a run. Pin those inputs before using repeatability as evidence.

Test that a failing test fails well.

Introduce a small known fault, such as read data changing during a stall. Check that the intended monitor identifies that fault. A compile error, unrelated timeout or runner crash is a different result.

Tool reference: Verilator simulation and coverage documentation. Pin the version you actually use; a moving documentation page is not a compatibility guarantee.

See the planned Header7F report standard →