The environment for autoresearch loops that optimize LLM inference. The engine lives in inference-server.
- Lab: the agent loop, its tools, ledger, sandbox and cloud GPUs
- Workloads: real-trace corpus and the knowledge base
- Integrity verdict: did the agent earn its result
- Session feedback: facts for the agent, diagnostics for the researcher
- Load regimes: drives a server the eight ways it gets used and finds its goodput
- Correctness gate: optimized outputs still match the original model
- Canaries: planted cheats that must get caught
- Validity checks: held-out budget, cache audit, short-vs-full test agreement
- Kernel checker: rewritten kernels compute the same thing
Diagnostics are developed on the public InferenceBench trajectories.