ghost trials.Give Ghost Trials an executable target and a bounded task. It searches for failures, preserves the execution evidence, and checks the next version against the same case. Start with a CPU-simulated robot controller or a controlled AI-agent reference target.
Interactive illustration · not an executed simulation
Under what tested conditions does this executable policy fail? Ghost Trials observes actual simulation or agent execution, applies domain-specific assertions, and retains the case for replay. A bounded finding is not a universal safety or security guarantee.
Run a gantry-arm controller in headless CPU MuJoCo. Vary supported mass, friction, position, and delay; inspect measured trajectories and reproducible task failures.
Execute bundled reference agents against adversarial input and mock-tool responses. Inspect tool calls, permission violations, deterministic assertions, and corrected-version regression.
Vary physical controls or agent inputs and tool responses. Candidates stay within declared bounds, and observed outcomes inform the next tests.
Configure a Hunt ↗Inspect measured robot state or agent tool-call traces and assertions. Observations stay separate from optional model interpretation.
Review trial runs ↗Verify retained evidence, reproduce the finding, and test a compatible next version against that exact case.
Open Ghost Cases ↗Search, execution, deterministic verification, and interpretation are distinct stages.
Choose a controller or supported reference agent, its task, and bounded test budget.
target → search spaceExecute physics or the controlled agent harness and retain state or tool-call events.
candidate → traceApply measured task rules or versioned agent assertions. Provider errors are not target failures.
trace → outcomeReproduce a saved Ghost Case and compare a compatible candidate on that case.
case → regressionA pass or fail cannot tell the whole story. Physical trials retain simulator measurements; agent trials retain tool events and assertion references. Optional Nemotron analysis is interpretation, not an observed outcome.
Inside an evidence record ↗GHOST CASEILLUSTRATIVE SCHEMAdomainphysical_ai | ai_agenttarget_versionversion + digesttest_configuration{ bounded recipe }observations{ trace, assertions }evidence_statusverified | reviewregressiontested case only