Evidence, not a scorecard
Measure the route, then the result.
yantraOS distinguishes deterministic fast paths from supervised visual loops. A useful benchmark must keep those routes separate and verify what changed on the machine.
Publication status
No Hosted Runtime benchmark is published yet.
Hosted Runtime remains in development. This page documents the reporting standard and the verified public alpha routing model; it does not claim throughput, latency, or success-rate results that have not been captured.
Benchmark contract
Every published run should include these fields.
Completion
Verified final state, not whether a model proposed a plausible action.
Route split
Fast-path, visual, and sandboxed actions reported separately.
Latency
Elapsed time from approved action to verified outcome.
Safety
Confirmation, policy denial, and audit outcomes retained with the run.
Route A
Fast path
Known, supported operations take a typed, allowlisted route after confirmation. They should be benchmarked against verified file or application state.
Route B
Visual loop
Desktop work uses a supervised visual bridge with explicit verification. It is a separate class of evidence, not a slower version of the fast path.
See the interactive route replay for a non-executing explanation, read the action-routing ADR, or follow the release gates.