A Gradia Universe is the governed world: its state, tools, actors, time, interruptions, visibility rules, and evidence. Benchmarks are the frozen measurement editions run inside it. One world can support many task families and benchmark editions without pretending they are the world itself.
See the measurements produced inside them →A workflow with possibilities
Keep the starting conditions fixed. Compare what happens when an agent waits, asks for approval or takes another permitted action.
See how a world changes →
Inspect the history and the information each actor could see.
Check the stop receipt and settled actions before continuing.
Restore supported state into separate continuations and compare outcomes.
Concept artwork. Interruption and independent restoration are demonstrated in a bounded synthetic browser workflow. Each additional application, external effect and state component needs its own qualification.
No Universe has passed the public release gate
Candidate evidence stays private until an authorized reviewer signs its exact bundle digest. Unreleased samples never appear here by accident.
Every benchmark needs an environment. Only an exact, reviewed, signed bundle becomes a released Gradia Universe.
Research candidate · exact public release not authorized
Measured by: Conditionally Approved
A branchable synthetic world where evidence can arrive, change, expire, or be retracted while an agent coordinates documents, authority, deadlines, and handoffs.
Boundary: Scripted controls and preliminary machine evidence exist. The exact public Universe bundle still requires its own rights, review, signature, and release receipt.
Inspect research record →Independent public research environment · Gradia conversion pending
Measured by: The Value Engine Benchmark
A synthetic multi-touch sales world with changing stakeholders, commitments, evidence, and negotiation state.
Boundary: Its public benchmark is real, but this environment has not been ingested and released as a signed Gradia Universe.
Inspect research record →Published research environment · not a released Universe
Measured by: Published evaluator-metrology study
An adversarial evaluation environment that mutates frozen scorers, preserves controls, localizes exploitable shortcuts, and re-attacks repairs.
Boundary: The study has evaluator evidence, not optimizer-backed training lift or a certificate-backed public Universe release.
Inspect research record →