Executable examples
Start from a complete engineering question—not a disconnected API call.
Every example declares a plan, adapts domain evidence, computes a bounded conclusion, and preserves the report for replay, CI, comparison, or Studio submission.
Start here
One candidate. Three studies. One release decision.
WAREHOUSE RELEASE WORKFLOW
Follow an autonomous mobile robot from matched replay evidence to a blocked CI decision.
The orchestrator runs a causal perturbation campaign, a visibility failure-boundary study, and sim-to-real transport validation. It writes every report separately, then applies a noncompensatory release policy.
Open the guided walkthrough Open the executable sourceuv sync --package iso-obs
uv run python \
packages/python-sdk/examples/warehouse_release_workflow.py \
--output-dir evidenceThe fixture candidate blocks release. Its latency intervention is harmful, its visibility envelope contains unresolved and unreliable regions, and a rare low-friction stratum fails transport validation.
Choose an example
Begin with the decision your evidence must support.
Causal perturbation campaign
- Question
- Did a controlled change cause material harm?
- Input
- Matched baseline and intervention outcomes
- Output
- Simultaneous effect intervals and a noncompensatory campaign disposition
CausalPerturbationPlan · PairedPerturbationObservation · CausalPerturbationReportOpen executable example Failure boundary map
- Question
- Where is behavior known to be reliable, unresolved, or unreliable?
- Input
- Fixed trials at predeclared operating anchors
- Output
- Certified cells with intervals and retained unresolved regions
FailureBoundaryPlan · BoundaryObservation · FailureBoundaryMapOpen executable example Sim-to-real transport
- Question
- Does simulator evidence transfer to physical anchors?
- Input
- Matched simulation and physical observations by critical stratum
- Output
- Local and overall discrepancy bounds plus limiting strata
TransportValidationPlan · PairedAnchorObservation · TransportValidationReportOpen executable example Mixed-mode dataset curation
- Question
- Can a learned detector help curate transitions without becoming ground truth?
- Input
- Synchronized modalities, detector lineage, prediction sets, and human evidence
- Output
- Reviewable transition candidates and an auditable label ledger
ModeDetectorManifest · ModeChangePointCandidate · DatasetLabelAssertionOpen executable example Failure replay capsule
- Question
- Can a confirmed recorded failure be reconstructed?
- Input
- Dataset, synchronization, simulator, recording, and claim evidence
- Output
- Content-addressed replay capsule with separate fidelity and claim qualification
FailureReplayRequest · FailureReplayCapsule · ReplayReadinessIssueOpen executable example JSONL dataset ingestion
- Question
- Can a robot trajectory become an auditable dataset artifact?
- Input
- Explicit nested mappings, timestamp units, clocks, modalities, and scope
- Output
- Linked dataset bundle and ingestion report
DatasetIngestionPlan · ChannelIngestionRule · DatasetIngestionReportOpen executable example MCAP and ROS 2 ingestion
- Question
- Can a ROS 2 recording retain topic, schema, payload, and clock contracts?
- Input
- MCAP recording with explicit topic rules and episode windows
- Output
- Dataset evidence preserving physical order and clock identity
McapDatasetIngestionPlan · McapTopicRule · McapEpisodeWindowOpen executable example Split rosbag2 ingestion
- Question
- Can split recordings retain their exact manifest and file lineage?
- Input
- metadata.yaml and ordered MCAP files
- Output
- Dataset artifacts plus independently verified recording evidence
Rosbag2McapIngestionPlan · Rosbag2RecordingEvidenceOpen executable example Trajectory world-model training
- Question
- Where does a learned dynamics model remain faithful over a rollout horizon?
- Input
- Content-addressed trajectories, a declared training plan, model lineage, and held-out episode slices
- Output
- Epoch history, compute provenance, sliced one-step and rollout errors, and explicit limitations
WorldModelTrainingPlan · TrainingEpoch · WorldModelTrainingReportOpen executable example Adaptation model
Replace the data seam, not the scientific design.
The deterministic fixture builders make every example runnable locally and in CI. In a real integration, replace each build_observations function with a narrow adapter over your simulator, trial store, logs, world-model rollouts, or physical tests.
Plan
Question, estimand, thresholds, design, and decision rule selected before results.
Evidence
Observations plus matching, timing, version, and provenance identities.
Uncertainty
Interval, prediction set, identified set, or unresolved map region.
Disposition
The strongest conclusion currently supported, including abstention states.
Claim boundary
The population, version, environment, and operating conditions limiting use.
Inspect the uncertainty object, limiting strata, evidence scope, assumptions, and limitations behind every disposition.
Maintained resources