Studio guide / Pro + Team
Build and run without hiding what the SDK does.
Reliability Studio turns the SDK workflow into a guided interface. Configure controlled conditions, inspect the generated code, prepare immutable resources, and choose where the job executes.
01 / Workflow
Four deliberate stages
- 01Configure
Choose an environment and declare the controlled changes, seeds, purpose, and limits.
- 02Inspect code
Review the equivalent SDK code before Studio persists or launches anything.
- 03Prepare
Create an immutable scenario version and evaluation suite.
- 04Execute
Choose your CPU, Iso-managed CPU, or a customer-hosted GPU runner and authorize the exact job.
Sign in to Studio and open Build & run from the main navigation. It is available to Pro and Team workspaces. Community can use the SDK and customer-hosted execution, but the guided Studio builder is not included.
02 / Inputs
What Studio expects
Studio validates these values before creating a scenario, preparing a suite, or authorizing billable compute.
| Input | Expected shape |
|---|---|
| Environment | A registered simulator, benchmark, rig, or declared dynamics source. |
| Version label | A stable, immutable label such as latency-85ms or modal-l4-v1. |
| Scenario | A controlled condition named without predicting the outcome. |
| Seeds | One or more integers. Each seed becomes a separate run. |
| Purpose | The intended evaluation and the limits of the claim. |
| Perturbations | A family, name, JSON parameters, and optional active step window. |
| Advanced config | A JSON object passed to the runner as pinned scenario context. |
03 / Plans
Pro guides; Team opens the full workspace
PRO / GUIDED + TRANSPARENT
Configure with controls and inspect the code.
Pro supplies guided inputs and generates equivalent Python and JSON. You can audit or copy the workflow before preparing it, but the versioned manifest is edited through the guided form.
TEAM / FULL WORKSPACE
Edit the complete versioned workflow manifest.
Team adds a bidirectional manifest editor. Applying a valid manifest updates the guided controls; it does not start compute. The prepared scenario and suite remain immutable resources.
{
"schema_version": "iso-obs.studio-scenario-workflow.v1",
"projectId": "prj_...",
"environmentId": "env_...",
"scenarioName": "payload-and-friction-shift",
"version": "modal-l4-v1",
"purpose": "Evaluate stopping behavior under a heavier payload.",
"perturbations": [{
"family": "dynamics_shift",
"name": "heavy-payload",
"parameters": { "payload_mass": 1.2 },
"start_step": 0,
"end_step": 25
}],
"config": { "max_steps": 64 },
"seeds": [2026071902]
}04 / Execution
Choose where the work runs
YOUR CPU RUNNER
Run locally or in your infrastructure.
Studio queues one exact job. Start iso-runner with the displayed job ID, API base URL, and a workspace API key.
ISO-MANAGED CPU
Size and authorize before billing.
Studio estimates worker-hours, shows the available balance, and requires explicit authorization before the dedicated CPU job is queued.
CUSTOMER GPU
Keep GPU credentials in your provider.
A GPU-capable runner claims the exact queued job and uploads traces, metrics, artifacts, and completion state. Studio does not receive your provider secret or pretend that queueing ran the job.
Customer CPU and GPU code executes outside Studio today. The runner performs the assigned adapter or training work and returns the authorized evidence. Managed Studio execution is CPU-only.
05 / Results
Follow the run through the reliability loop
- 01Watch execution
Track queue state, progress, runtime, and the runner handoff.
- 02Inspect evidence
Review observations, actions, metrics, artifacts, and failures.
- 03Compare versions
Select a baseline and candidate and locate sustained divergence.
- 04Promote failures
Turn a reproducible failure into a pinned regression case.
Environment rendering requires simulator-declared geometry. Recorded playback requires positions or exact frame artifacts in the trace. Studio reconstructs only declared or recorded evidence and does not invent missing boundaries.