BenchHandoff · Experiment recovery
Verify declared inputs and completed outputs before resuming an interrupted experiment batch. Records hashes and logs, rechecks completed outputs, and rejects changed evidence before an approval-bound retry.
Python · Go · Kubernetes
BenchHandoff examples · BenchHandoff checks
DCPInvariant · Checkpoint verification
Check PyTorch checkpoint restoration with fixed single-host CPU scenarios. Covers training-state restoration, process-count changes, asynchronous snapshots, stale publishers, and damaged checkpoint files.
Python · PyTorch · Distributed Checkpoint
DCPInvariant scenarios · DCPInvariant source · DCPInvariant checks
- EvalFence — a Rust CLI for checking evaluation metrics, duplicate records, and declared agent inputs.
- EffectWitness — compares client observations with durable tool effects when a response is lost.
- CacheInvariant — checks inference-cache behavior against a pinned CPU fixture.
The project READMEs document the supported scenarios, runtime, and runnable checks.